Datasets:

jamescalam

/

ai-arxiv-chunked

like 30

2108.10934

101

Then Chernozhukov and Hong (2003); Lyddon et al. (2018) show that under regularity conditions the following asymptotic result holds p N ^(N) IW D!N 0;JIW( IW)1 asN!1 whenis distributed according to the general Bayesian posterior almost surely w.r.t. x1:1. Similarly, if we define J0() :=Z r2 `0(x;)dPD(x); then we have that under the standard Bayesian posterior (Chernozhukov and Hong, 2003; Kleijn et al., 2012; Lyddon et al., 2018) p N ^(N) 0 D!N 0;J0( 0)1 almost surely w.r.t. x1:1. Now it follows from the importance sampling identity that  IW= arg min 2Z `IW(x;)dPG(x) = arg min 2Z `0(x;)dPD(x) = 0; JIW() =Z r2

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

102

2Z `0(x;)dPD(x) = 0; JIW() =Z r2 `IW(x;)dPG(x) =Z bw(x)r2 `0(x;)dPG(x) =Z r2 `0(x;)dPD(x) =J0() Moreover ^(N) 0and^(N) IWare also consistent estimates of  0under the same regularity conditions. This establishes the result. B.6.1 Finite Sample Importance-Weighted Bayesian posterior To complement the asymptotic results connecting the importance weighted general Bayesian posterior given data from pGand the standard BayesianpDwe can consider the difference between these two for finite n=m. This is formulated in the following proposition. Proposition 4. The expected KLD beween standard Bayesian posterior (jx1:n)and its importance weighted approximation IW(jz1:m)in expectation over the generating distributions for x1:nPDandz1:mPG, forn=mis

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

103

IW(jz1:m)in expectation over the generating distributions for x1:nPDandz1:mPG, forn=mis ExpD[EzpG[KLD ((jx1:n)jjIW(jz1:m)]] =nExpD E(
jx1:n) logf(x;)Ex0pD logf(x0;)
 1IW(jx1:N)and(jx1:N)are here interpreted as random probability measures, and functions of the random observations x1:N. 24 Proof. We have ExpD[EzpG[KLD ((jx1:n)jjIW(jz1:m)]] =ExpD EzpGZ (jx1:n) log(jx1:n) IW(jz1:m)d =ExpD" EzpG" E(jx1:n)"nX

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

104

=ExpD" EzpG" E(jx1:n)"nX i=1logf(xi;)mX j=1bw(zi) logf(zi;)### : Now by Fubini we can reorder these integrals assuming that they all exist =ExpD" E(
jx1:n)" nX i=1logf(xi;)mX j=1EzpG[bw(zi) logf(zi;)]!## =ExpD" E(
jx1:n)" nX i=1logf(xi;)mEx0pD logf(x0;)!## : Now assuming n=m, we have =ExpD" E(
jx1:n)"nX i=1 logf(xi;)Ex0pD logf(x0;)
## =nExpD E(
jx1:n) logf(x;)Ex0pD

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

105

=nExpD E(
jx1:n) logf(x;)Ex0pD logf(x0;)
 : C EXPERIMENTS C.1 EXPERIMENTAL DETAILS Please refer to Table 4 for an overview of the data sets used. We considered a random 80=20train test split for all data sets except for MNIST for which the default split was used. Data # training observations # features prediction problem Iris 150 4 3-class classification tgfb 262 7 regression Boston 506 10 regression Breast 569 30 binary classification Banknote 1372 4 binary classification MNIST 60000 784 10-class classification Table 4: Characteristics of the analysed data sets We obtained the code for PrivBayes from https://github.com/DataResponsibly/DataSynthesizer , and the code for DPCGAN from https://github.com/ricardocarvalhods/dpcgan . This code was used and changed to write the code for DPGAN. For the logistic regression alternatives we use an adaption of the sklearn implementation. DPGAN was trained on labelled

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

106

DPGAN. For the logistic regression alternatives we use an adaption of the sklearn implementation. DPGAN was trained on labelled data by concatenating the features with the one hot encoding of the labels. Our implementation will be made available online. We train different downstream tasks on the synthetic data and test them on test data to ensure their utility for the setting of supervised learning. The downstream algorithms were trained using sklearn with default parameters. Hyperparameter tuning is a non-private operation as it queries private data to evaluate the model at validation time. To ensure that we do not undermine the performance of the baselines we tuned them for = 1:, and chose default parameters for our method. PrivBayes is trained in correlated attribute mode, and with optimal bandwidth computation. For the GAN alternatives, we tuned the norm clip (1.0, 0.5), the batch size (32, 64), and number of epochs (50, 100) with grid search on a validation set (10% split of training). The noise multiplier was chosen such that the desired privacy budget was reached. The models were then retrained on the full training data set. Note that these hyperparameters are chosen smaller than in a non-private setting as the noise to be added would otherwise explode. The optimal

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

107

these hyperparameters are chosen smaller than in a non-private setting as the noise to be added would otherwise explode. The optimal hyperparameters can be found in the GitHub repository. Further we chose learning rate of the discriminator and generator as 0.15, and the 25 number of hidden dimensions as dfollowing Jordon et al. (2019). For the MNIST experiment, we chose to use the hyperparameters found by Torkzadehmahani et al. (2019). The regularisation parameter of the logistic regression for weight estimation was chosen from 0:1;1;2. The MLP for likelihood ratio estimation was computed based on the tensorflow andtensorflow_privacy package. To ensure the privacy of the MLP, we started with a configuration of one epoch, a batch size of 1, an L2 norm clip of 1, a noise multiplier of 5.2, 20 microbatches and a learning rate of 0.1. We computed the using built-in functions and increased/decreased the noise multiplier and the number of epochs until the desired privacy level was reached. We chose NS=NDunless otherwise mentioned. To compute the output-noised weights we computed the largest NSsuch that the scale restriction was satisfied and conducted the downstream analysis on this smaller dataset.

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

108

output-noised weights we computed the largest NSsuch that the scale restriction was satisfied and conducted the downstream analysis on this smaller dataset. C.2 COMPUTATIONAL TIME OF IMPORTANCE WEIGHT ESTIMATION Please refer to Table 5 for an overview of the additional time needed to compute the importance weights. All experimental results were computed by training on a single Tesla V100 GPU. We observe that the estimation of the importance weights comes with negligible computational overhead. weighting Iris Banknote Housing Breast MNIST BetaNoised 0:00640:0002 0:00840:0002 0:01330:0011 0:08240:0206 51:56059:0042 BetaDebiased 0:02370:0125 0:01120:0003 0:07420:0083 0:18560:0858 59:072310:5120 DP-MLP 0:83380:0964 5:46490:0654 1:73030:1104 2:93630:1208 87:26934:7303

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

109

Discriminator 0:00000:0000 0:00000:0000 0:00000:0000 0:00000:0000 0:00000:0001 LogReg 0:00710:0004 0:00990:0003 0:01430:0012 0:09100:0210 52:03319:1285 MLP 0:77410:1436 1:58950:0261 1:74910:1414 1:44800:1441 30:19686:3155 Table 5: Additional computational time in seconds needed for the computation of importance weights averaged over 10 seeds and SDGP for = 1. C.3 CHOICE OF PRIVACY SPLIT In Figure 3, we plot the change in evaluation metrics for different values of privacy budget splits. We notice that the impact of the split parameter decreases the larger is. Similarly, the variability in the metrics for different splits decreases, the larger IWis, whereIW denotes the privacy budget dedicated to the importance weight estimation. While a larger split of 30-50% seems beneficial for DP-MLP,

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

110

denotes the privacy budget dedicated to the importance weight estimation. While a larger split of 30-50% seems beneficial for DP-MLP, the fraction of dedicated to the importance weighting model should be chosen relatively small, i.e. 10%. Note that we chose these default values based on their performance on the Adult, Credit and Spam data set. Tuning them to the underlying data and task characteristics will be able to improve their results. As hyperparameter tuning is an unsolved problem in DP, we leave the procedure for choosing the optimal privacy split per data set for future work. We note that an additional intricacy appears in DP because of the noise injection which increases the variability of the model’s performances. C.4 MSE OF IMPORTANCE WEIGHT ESTIMATION For each of our experiments, we compute the mean squared error between the privatised parameters of the logistic regression for importance weight estimation and the parameters of an unperturbed logistic regression trained on the private data. Please refer to Table 6 for the results. We observe that debiasing almost always decreases the MSE in the low-privacy regimes. For large privacy budgets, the scale of the perturbations can be negligible for low-dimensional data sets which is why both approaches perform similarly on Iris and

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

111

scale of the perturbations can be negligible for low-dimensional data sets which is why both approaches perform similarly on Iris and Banknote, but debiasing still helps with larger data sets such as Breast. C.5 BAYESIAN UPDATING EXPERIMENTAL DETAILS In addition to the logistic regression ROC-AUC score distributions presented in the main body of the paper, we applied importance weighted posteriors to updating and learning the parameters of linear regression and multinomial logistic regression models applied to the TGFB and Iris datasets respectively, see Figures 4a and 4b. It can be seen that in the case of linear regression, the DP-MLP and MLP IW methods are again very effective, with the performance improving across all SDGPs. Other methods again tend to reduce variance in the results whilst not damaging performance and so can be seen to be effective in at least ensuring greater robustness and consistency when learning under synthetic data. In the case of the Iris data, we calculated 1 vs all ROC-AUC scores for each class separately, then averaged these per-class ROC-AUCs to get a single multi-class average ROC-AUC. Again, MLP and DP-MLP are stand-out in their 26 0.10 0.20 IW %

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

112

26 0.10 0.20 IW % 0.60.70.80.9MLP-AUC 0.10 0.20 IW % 12345 MSE 0.10 0.20 IW % 0.51.01.52.0WST 0.10 0.20 IW % 0.10.20.30.40.5MMD 1.0 6.0 IW DP-MLP BetaNoised BetaDebiasedDPCGAN on Breast 0.10 0.20 IW % 0.50.60.70.80.91.0MLP-AUC 0.10 0.20 IW % 2468 MSE 0.10 0.20 IW % 0.51.01.52.0WST 0.10 0.20 IW % 0.10.20.30.40.5MMD 1.0 6.0 IW DP-MLP BetaNoised BetaDebiasedDPGAN on Breast 0.10 0.20 IW % 123MLP-MSE 0.10 0.20 IW % 0.00.51.01.52.0 MSE 0.10 0.20 IW %

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

113

0.10 0.20 IW % 0.00.51.01.52.0 MSE 0.10 0.20 IW % 0.000.050.100.150.20WST 0.10 0.20 IW % 0.20.30.40.50.60.7MMD 1.0 6.0 IW DP-MLP BetaNoised BetaDebiasedDPGAN on Boston 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.750.800.850.900.95MLP-AUC 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 1.001.251.501.752.002.25 MSE 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.40.60.81.01.21.4WST 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.150.200.25MMD IW % 0.04 0.08 0.1 0.16DPCGAN on Breast

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

114

IW % 0.150.200.25MMD IW % 0.04 0.08 0.1 0.16DPCGAN on Breast 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.60.70.80.9MLP-AUC 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 1234 MSE 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.40.60.81.0WST 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.150.200.250.30MMD IW % 0.04 0.08 0.1 0.16DPGAN on Breast 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.51.01.52.0MLP-MSE 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.00.10.20.3 MSE

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

115

0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.00.10.20.3 MSE 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.0000.0250.0500.0750.100WST 0.1 0.2 0.3 0.4 0.5 0.6 0.7 IW % 0.30.40.5MMD IW % 0.04 0.08 0.1 0.16 0.2DPGAN on BostonFigure 3: Multiple metrics measured across a range of privacy splits on Breast and Boston averaged over 10 seeds, and displayed with standard errors. The maximum mean discrepancy (MMD) was included as a measure of divergence between the weighted SDGP and the test distribution. 27 = 1 = 6SDGP dataBetaNoised BetaDebiased BetaNoised BetaDebiased CGAN Breast 1:48330:9603 0:07750:0197 0:00240:0006 0:00200:0004

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

116

CGAN Breast 1:48330:9603 0:07750:0197 0:00240:0006 0:00200:0004 Banknote 0:04200:0211 0:04130:0196 0:00140:0007 0:00140:0007 Iris 8:75224:9893 3:46871:3044 0:11600:0240 0:12900:0311 GAN Housing 8:20817:7702 1:44060:8314 3:79163:3246 1:54791:0430 DPCGAN Breast 0:05820:0165 0:04450:0162 0:00150:0003 0:00140:0003 Banknote 0:04200:0211 0:04130:0196 0:00220:0013 0:00210:0012 Iris 0:78340:2341 1:23000:7050 0:25020:1627 0:28060:1760

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

117

Iris 0:78340:2341 1:23000:7050 0:25020:1627 0:28060:1760 DPGAN Breast 6:04873:7927 3:76292:2881 0:02510:0245 0:02380:0234 Banknote 0:05820:0353 0:06100:0397 0:00620:0057 0:00610:0056 Iris 2:64861:3518 1:36981:1554 0:07410:0228 0:08640:0274 Housing 5:91752:8546 0:83980:6328 1:90441:1426 2:11111:3450 Table 6: Mean squared error averaged over 10 runs with standard errors reported in brackets for (= 1;= 105)and (= 6;= 105)whereIW= 0:1. performance, significantly improving the performance measured by this metric, especially under synthetic data from the CGAN, DPCGAN

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

118

performance, significantly improving the performance measured by this metric, especially under synthetic data from the CGAN, DPCGAN and PrivBayes generators. Similar gains can be seen across the majority of the methods for the DPCGAN, especially at the higher = 6. All of these models were implemented in the Turing.jl PPL Ge et al. (2018). We then ran an experiment for each model and dataset on a defined grid across all seeds, synthetic generators and values. For each combination, we generated 10,000 samples across 4 chains (not counting 1,000 discarded warm-up samples per chain) for each of the importance weighting methods, as well as once for a model fit on the synthetic data with its standard non-weighted posterior, and once for the real data. We used Turing’s implementation of the NUTS sampling algorithm with a target acceptance ratio of 0:65for sampling the linear regression models’ parameters, and for the logistic and multinomial logistic regression models we used HMC with a leapfrog step size of 0:05and10leapfrog steps per iteration. The logistic and multinomial logistic regression models’ coefficients (including intercepts) were given centred Normal priors with = 1. The linear

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

119

and multinomial logistic regression models’ coefficients (including intercepts) were given centred Normal priors with = 1. The linear regression models’ coefficient priors were given the same centred Normal priors with = 1; its variance was given a non-informative prior via a truncated Normal distribution ensuring positivity with = 10 . We then took all 10,000 samples and calculated our evaluation metrics on the test set for each sample, storing all of these. We then present the distributions of metric scores that arise in the included box-plot figures. C.6 ILLUSTRATIVE EXAMPLE OF THE IMPLICATIONS OF BIAS MITIGATION Figure 5: Illustrative example of debiasing with IW on PrivBayes synthesised Banknote data.In Figure 5, we visualise the benefit of debiasing: We fitted a logistic regression as a downstream classifier on the private data to get the truecoefficients . The predictedcoefficients are estimated by training the logistic classifier on the importance weighted synthetic data. Each dot in the figure plots one dimension

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

120

importance weighted synthetic data. Each dot in the figure plots one dimension of the predicted coefficients against its true counterpart for one training run (out of ten). An optimal classifier would reconstruct the true coefficients. In this case all lines would be on the diagonal. An unbiased estimator would on average reconstruct the true coefficients: For each true coefficient, the predicted coefficients would be centred around the true value. We observe that coefficients learned without importance weighting exhibit the largest distance to the diagonal line, while the importance weighting alternatives push the dots closer to the diagonal line. Our method, DP-MLP, is particularly successful in decreasing the bias in thecoefficients. C.7 COMPLETE UCI RESULTS The complete experimental results on the UCI data sets can be found in Tables 7 to 10. Each table displays the performance of the different weight estimators for private and non-private synthetic data generative models for 2f1;6g,IW= 0:1andIW= 0:3. We observe

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

121

that importance weighting brings significant gains especially in low privacy regimes. For high privacy regimes this effect is reduced as the SDGP gets closer to the DGP. 28 ϵ = 1 ϵ = 6 None LogReg MLP BetaNoised BetaDebiased DP-MLP Discriminator2-2202224 2-2202224 IW MethodT est Predictions MSESynth Model DPGAN GAN PRIVBAYES(a) Test set prediction MSE distributions calculated via chains of parameters sampled from a Bayesian linear regression model fit on synthesised TGFB data across 10 seeds. ϵ = 1 ϵ = 6 None LogReg MLP BetaNoised BetaDebiased DP-MLP Discriminator0.40.60.81.0 0.40.60.81.0 IW MethodMulti-Class Averaged ROC-AUCSynth Model CGAN DPCGAN DPGAN PRIVBAYES (b) Multi-class averaged ROC-AUC distributions calculated via chains of parameters sampled from a Bayesian multinomial logistic regression model fit on synthesised Iris data across 10 seeds. 29 SDGP CGAN DPCGAN DPGAN PrivBayes= 1

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

122

multinomial logistic regression model fit on synthesised Iris data across 10 seeds. 29 SDGP CGAN DPCGAN DPGAN PrivBayes= 1 MLP-ROC-AUC"None 0:46190:1010 0:47170:1103 0:53570:0752 0:52430:1299 BetaNoised 0:58240:0931 0:58410:0831 0:54870:0803 0:66510:0884 BetaDebiased 0:56690:1237 0:59130:1136 0:59980:1141 0:50050:0793 DP-MLP 0:62990:0984 0:57250:0859 0:54480:0912 0:61430:0374 Discriminator 0:58090:0840 0:59950:0982 0:64750:0701 LogReg 0:49800:0780 0:49080:0950 0:48060:0806 0:62450:1235

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

123

MLP 0:72300:0791 0:62730:0988 0:57700:1199 0:67780:0923MSE#None 1:35940:3789 1:04600:2457 3:89550:9764 0:35110:0753 BetaNoised 1:49440:2321 1:11330:1911 4:15651:0469 0:47390:0469 BetaDebiased 1:36820:3080 1:33470:2830 4:16940:9246 0:81470:1690 DP-MLP 0:61090:0481 1:06630:1411 4:49861:2881 0:19620:0413 Discriminator 1:04540:3012 0:94040:1024 3:90490:6010 LogReg 1:33450:2725 0:95570:1356 4:19711:1035 0:36590:0660

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

124

MLP 0:60910:0546 0:83160:1630 4:51091:3057 0:15510:0162WST#None 0:72260:0543 0:74480:0423 0:79190:0458 0:50550:0111 BetaNoised 0:27710:0490 0:10140:0519 0:18930:0266 0:14120:0493 BetaDebiased 0:23400:0210 0:09890:0062 0:14570:0143 0:10590:0032 DP-MLP 0:39600:0561 0:23760:0196 0:26130:0627 0:34510:0253 Discriminator 0:26980:0383 0:16960:0371 0:10030:0003 LogReg 0:23410:0687 0:14440:0406 0:16110:0178 0:35310:0357

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

125

MLP 0:26770:0693 0:09670:0287 0:07520:0261 0:13960:0139= 6 MLP-ROC-AUC"None 0:46620:1039 0:52020:0928 0:52520:0844 0:48750:1139 BetaNoised 0:58420:0900 0:55310:1093 0:56030:0980 0:62180:1304 BetaDebiased 0:60290:1100 0:69920:0801 0:64450:0906 0:53880:1258 DP-MLP 0:60070:1060 0:60540:0951 0:51810:0957 0:56390:0483 Discriminator 0:58940:0829 0:58060:1014 0:59090:0903 LogReg 0:50730:0852 0:53530:0793 0:49340:1051 0:70880:0843

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

126

MLP 0:72060:0774 0:71180:0774 0:59230:1130 0:67340:0881MSE#None 1:41110:3882 1:02620:1866 2:07100:3284 0:26500:0610 BetaNoised 1:28940:2726 0:95070:3017 2:82841:0195 0:33380:0701 BetaDebiased 1:26790:2854 0:95110:3113 2:82561:0359 0:34920:0719 DP-MLP 0:59280:0682 0:77730:2286 4:11121:1372 0:25590:0527 Discriminator 1:04340:3014 0:94490:2838 2:12030:5427 LogReg 1:26060:2771 0:96040:3155 2:84091:0311 0:36030:0806

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

127

MLP 0:61740:0523 0:51020:1630 3:94031:1462 0:12830:0252WST#None 0:73990:0445 0:65980:1077 0:67700:0379 0:42550:0208 BetaNoised 0:27030:0492 0:30320:0697 0:26220:0229 0:44670:0200 BetaDebiased 0:30350:0601 0:31710:0746 0:27700:0332 0:33830:0070 DP-MLP 0:45070:0722 0:53740:0654 0:44450:0635 0:48500:0160 Discriminator 0:21340:0419 0:21680:0032 0:21780:0037 LogReg 0:30900:0612 0:28360:0742 0:26010:0262 0:45910:0121

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

128

MLP 0:20640:0819 0:13430:0299 0:27110:0235 0:19810:0192 Table 7: Results on Iris averaged over 10 seeds. 30 SDGP CGAN DPCGAN DPGAN PrivBayes= 1 MLP-ROC-AUC"None 0:74080:0522 0:85460:0213 0:68630:0436 0:76300:0495 BetaNoised 0:74690:0522 0:84950:0274 0:60630:0510 0:89430:0173 BetaDebiased 0:78640:0888 0:87290:0310 0:58680:1005 0:76320:0517 DP-MLP 0:73130:0613 0:76970:0419 0:56570:0570 0:89530:0299

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

129

DP-MLP 0:73130:0613 0:76970:0419 0:56570:0570 0:89530:0299 Discriminator 0:75110:0523 0:86950:0167 0:71140:0424 LogReg 0:79860:0391 0:81720:0327 0:60340:0534 0:91020:0129 MLP 0:72530:0521 0:82910:0333 0:59740:0627 0:85940:0231MSE#None 15:32782:5238 11:02151:8377 39:32433:7708 8:17240:3987 BetaNoised 11:76362:1960 8:42981:0383 35:28624:0365 5:70010:1885 BetaDebiased 8:49461:7858 8:35082:3127 32:99095:9024 6:68620:1458

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

130

BetaDebiased 8:49461:7858 8:35082:3127 32:99095:9024 6:68620:1458 DP-MLP 14:66442:9599 17:15972:5448 36:46184:1011 3:55190:2895 Discriminator 14:95372:5553 12:54712:3124 30:92825:4283 LogReg 11:77772:2000 8:47601:0406 35:29644:0396 5:67510:1785 MLP 15:45843:0826 17:93902:4926 35:52114:2147 2:62860:3761WST#None 0:67020:0282 0:47460:0214 0:74420:0333 0:32370:0162 BetaNoised 0:31060:0475 0:25090:0436 0:43550:0456 0:23180:0035

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

131

BetaNoised 0:31060:0475 0:25090:0436 0:43550:0456 0:23180:0035 BetaDebiased 0:38370:0990 0:40150:0766 0:46180:0832 0:23690:0061 DP-MLP 0:14180:0283 0:20350:0427 0:42980:0433 0:04560:0061 Discriminator 0:63660:0273 0:33820:0399 0:10870:0415 LogReg 0:30920:0470 0:25080:0432 0:43480:0460 0:23480:0034 MLP 0:04940:0141 0:09130:0259 0:38600:0452 0:00210:0004= 6 MLP-ROC-AUC"None 0:72120:0491 0:89580:0179 0:83230:0301 0:83570:0354

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

132

BetaNoised 0:78110:0423 0:87710:0227 0:82160:0320 0:85880:0295 BetaDebiased 0:69510:0958 0:89920:0334 0:70610:1083 0:81360:0648 DP-MLP 0:68790:0547 0:85820:0330 0:74450:0511 0:88990:0148 Discriminator 0:73320:0529 0:89760:0148 0:80710:0362 LogReg 0:79530:0421 0:88670:0207 0:78710:0351 0:86680:0336 MLP 0:69600:0456 0:85990:0291 0:80250:0212 0:84040:0400MSE#None 19:29594:0480 8:30741:6718 18:08352:5051 7:90520:3837

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

133

BetaNoised 14:43502:3116 6:46830:9572 23:05903:2307 5:47360:1792 BetaDebiased 13:15782:9727 5:68901:0695 19:16276:1430 6:47760:1134 DP-MLP 18:70593:0658 8:88201:4421 24:04333:4451 3:08830:2703 Discriminator 18:91944:0483 8:06821:5928 13:62671:9313 LogReg 14:44642:3126 6:47010:9581 23:06963:2327 5:47060:1781 MLP 18:24003:1143 9:71111:4901 23:02683:2550 2:45890:3184WST#None 0:66420:0270 0:47230:0294 0:56450:0219 0:29280:0118

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

134

BetaNoised 0:25070:0384 0:30780:0231 0:26080:0370 0:22690:0036 BetaDebiased 0:23160:0670 0:28920:0442 0:30290:0883 0:21760:0076 DP-MLP 0:13950:0262 0:09570:0183 0:17300:0413 0:11420:0017 Discriminator 0:63030:0278 0:35960:0470 0:04360:0100 LogReg 0:25040:0384 0:30830:0231 0:26070:0370 0:22720:0035 MLP 0:06580:0208 0:04090:0104 0:07870:0325 0:20250:0004 Table 8: Results on Banknote averaged over 10 seeds. 31 SDGP GAN DPGAN PrivBayes= 1 MLP MSE#None 1:44640:1591 1:88510:5262 0:19730:0108

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

135

MLP MSE#None 1:44640:1591 1:88510:5262 0:19730:0108 BetaNoised 0:64550:0942 1:00570:1973 0:22000:0154 BetaDebiased 0:64210:1290 0:90240:1244 0:21390:0122 DP-MLP 0:82790:0974 0:94620:1702 0:18770:0174 Discriminator 1:51260:1639 1:62560:2394 LogReg 0:62920:0909 1:06060:2648 0:25150:0305 MLP 0:62660:1273 1:09790:2225 0:16970:0079MSE#None 0:10170:0118 0:18670:0434 0:00110:0002 BetaNoised 0:06010:0172 0:17610:0948 0:00880:0028 BetaDebiased 0:06080:0190 0:06670:0188 0:00770:0022

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

136

BetaDebiased 0:06080:0190 0:06670:0188 0:00770:0022 DP-MLP 0:03630:0192 0:15300:0812 0:00480:0024 Discriminator 0:09400:0100 0:15670:1825 LogReg 0:07070:0194 0:07490:0279 0:00370:0016 MLP 0:00580:0007 0:14760:0804 0:00080:0002WST#None 1:30600:0319 2:20130:0945 1:39380:0231 BetaNoised 1:00600:0023 2:09220:0419 1:30090:0338 BetaDebiased 1:00230:0009 2:09300:0393 1:27050:0290 DP-MLP 1:00360:0015 2:05420:0184 1:02650:0035

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

137

DP-MLP 1:00360:0015 2:05420:0184 1:02650:0035 Discriminator 0:94720:0764 2:01450:0141 LogReg 1:00700:0042 2:20510:0819 1:40780:0492 MLP 1:00010:0001 2:03500:0158 1:00720:0009= 6 MLP MSE#None 1:82180:1514 1:80160:1771 0:16330:0074 BetaNoised 0:53180:0806 0:65290:0814 0:19400:0156 BetaDebiased 0:56470:1065 0:90250:1462 0:18100:0131 DP-MLP 0:97370:1178 1:09020:1486 0:14280:0068 Discriminator 1:83980:1446 1:86310:1986 LogReg 0:55010:0540 0:90500:1553 0:19340:0224

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

138

MLP 0:47250:0736 0:74640:1185 0:15810:0076MSE#None 0:12300:0110 0:14500:0174 0:00090:0002 BetaNoised 0:06950:0203 0:06080:0231 0:00220:0006 BetaDebiased 0:06930:0207 0:06130:0240 0:00180:0004 DP-MLP 0:00300:0006 0:03540:0112 0:00080:0002 Discriminator 0:11350:0098 0:22740:0375 LogReg 0:06970:0207 0:06060:0237 0:00180:0004 MLP 0:00630:0011 0:02120:0060 0:00080:0001WST#None 1:37270:0249 1:56810:0368 1:33060:0271 BetaNoised 1:00310:0012 1:06150:0304 1:39060:0410

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

139

BetaNoised 1:00310:0012 1:06150:0304 1:39060:0410 BetaDebiased 1:00310:0012 1:05980:0286 1:41060:0432 DP-MLP 1:01400:0032 1:03380:0126 1:24050:0133 Discriminator 1:04810:0752 1:38440:0654 LogReg 1:00310:0012 1:06230:0298 1:40330:0406 MLP 1:00010:0000 1:00810:0045 1:00970:0010 Table 9: Results on Boston averaged over 10 seeds. 32 SDGP CGAN DPCGAN DPGAN PrivBayes= 1 MLP-ROC-AUC"None 0:68010:0655 0:63740:0421 0:67910:0966 0:83660:0579 BetaNoised 0:77320:0589 0:61100:0477 0:65460:0727 0:70760:0983

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

140

BetaNoised 0:77320:0589 0:61100:0477 0:65460:0727 0:70760:0983 BetaDebiased 0:71510:1146 0:68200:0510 0:71730:0842 0:85570:0765 DP-MLP 0:71660:1038 0:79420:0404 0:56860:0823 0:73530:0887 Discriminator 0:86070:0485 0:69920:0839 0:72900:0720 LogReg 0:71410:0755 0:66310:0469 0:64840:1081 0:76180:1019 MLP 0:69420:1262 0:77300:0412 0:73580:1017 0:75730:0738MSE#None 2:36460:2983 2:06430:2012 4:98281:5701 2:39040:1050

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

141

BetaNoised 1:49000:1807 2:75320:2650 2:50250:3763 2:11440:2400 BetaDebiased 1:54130:2378 2:83370:3842 2:23241:0446 1:82660:2392 DP-MLP 0:99770:1617 2:39650:2083 3:88650:6043 2:31300:2195 Discriminator 1:85540:3263 1:45910:1837 4:06120:9523 LogReg 1:19400:1610 2:69340:2667 2:21560:3366 1:53330:2138 MLP 1:01200:1383 2:39990:2040 3:83430:7032 1:65810:2020WST#None 1:84260:1329 2:36650:0982 1:58530:1333 2:11170:1740

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

142

BetaNoised 1:31090:0507 1:43370:1114 2:22320:2325 1:23220:0823 BetaDebiased 1:06490:0120 1:89220:1237 1:99130:3507 1:18250:0933 DP-MLP 1:47370:1027 1:45700:1492 1:03150:1415 1:21900:0795 Discriminator 1:88140:1682 1:00070:0004 1:00010:0001 LogReg 1:43740:0467 1:64510:1168 2:29530:2121 1:46630:1152 MLP 1:30560:0524 1:61290:1404 1:07090:1579 1:41410:1216= 6 MLP-ROC-AUC"None 0:61770:0737 0:97900:0058 0:97560:0042 0:94350:0152

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

143

BetaNoised 0:71850:0898 0:97150:0031 0:97100:0065 0:96990:0121 BetaDebiased 0:90700:0434 0:97230:0033 0:97240:0066 0:98200:0064 DP-MLP 0:72030:1028 0:97030:0040 0:97280:0059 0:97540:0063 Discriminator 0:87120:0471 0:97630:0071 0:97370:0065 LogReg 0:68690:0760 0:97060:0033 0:97190:0049 0:98250:0061 MLP 0:68990:1290 0:95840:0080 0:97670:0043 0:95060:0250MSE#None 2:36020:4035 0:98860:2287 1:06530:1229 0:91420:1575

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

144

BetaNoised 1:24000:1637 1:03290:0732 1:15860:1312 1:04650:1358 BetaDebiased 0:93880:0802 1:01500:0783 1:16170:1936 0:98430:1766 DP-MLP 0:99490:1486 1:01190:0698 0:89690:0837 1:34420:0900 Discriminator 1:75880:3421 0:85390:2323 0:54230:0457 LogReg 1:22210:1598 1:03100:0719 1:14840:1276 1:02340:1274 MLP 1:08450:1210 1:09530:0844 0:92750:0938 1:53540:1343WST#None 1:84360:1257 1:33780:0282 1:64490:0849 2:04370:2188

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

145

BetaNoised 1:41640:0483 0:65260:0463 1:54850:0635 1:48080:0943 BetaDebiased 1:33140:0459 0:66410:0482 1:51560:0935 1:41330:1346 DP-MLP 1:71760:1206 0:79310:0380 1:55510:0826 1:49230:0685 Discriminator 1:85230:1553 0:23630:0425 1:10200:0158 LogReg 1:41400:0493 0:65970:0470 1:52810:0622 1:48240:0952 MLP 1:34870:0591 0:37620:0383 1:23090:0387 1:34060:0792 Table 10: Results on Breast averaged over 10 seeds. 33 C.8 COMPARISON TO EXPERIMENTAL RESULTS REPORTED BY RELATED WORK

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

146

Table 10: Results on Breast averaged over 10 seeds. 33 C.8 COMPARISON TO EXPERIMENTAL RESULTS REPORTED BY RELATED WORK We compare our results to PATE-GAN and DPGAN as DP synthetic data generators (Jordon et al., 2019; Xie et al., 2018). The PATEGAN implementation is taken from https://github.com/vanderschaarlab/mlforhealthlabpub . For DPGAN we chose the code from the DataSynthesizer package. In the implementation of the PATE-GAN method, Jordon et al. (2019) generate 50 independent synthetic data sets for each function call, returning the best synthetic data set as defined by a comparison with non-private validation data. The relative level of privacy violation in these situations is unknown, making interpretation of results and comparison between methods in tables and figures challenging. On re-implementing the methods to generate DP synthetic data, we find a substantial and significant drop in performance, which nonetheless is improved through bias mitigation. weightPATE-GAN DPGAN WST# MD# SVM" RF" MLP" WST# MD# SVM" RF" MLP"

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

147

weightPATE-GAN DPGAN WST# MD# SVM" RF" MLP" WST# MD# SVM" RF" MLP" None 1.5472 0.0670 0.4876 0.1686 0.0938 1.4997 0.0592 0.5263 0.2848 0.1548 BetaNoised 0.0023 0.0462 0.5482 0.5172 0.5020 0.0050 0.0375 0.4450 0.4973 0.2062 OutputLaplace 5.7380 300.24 0.6777 0.2225 0.4234 5.3239 300.59 0.4807 0.3760 0.5217 OutputNorm 6.3058 311.23 0.5590 0.2637 0.4221 5.2081 317.79 0.6503 0.3271 0.6153 DP-MLP 0.1769 0.0495 0.6196 0.4683 0.5517 0.0744 0.0466 0.3994 0.4054 0.3476Breast Discriminator 1.5194 0.0670 0.4867 0.1923 0.0898 1.4975 0.0592 0.5260 0.2592 0.1021

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

148

None 3.0221 0.1962 0.4966 0.4508 0.4269 0.6436 0.0050 0.5293 0.3957 0.4483 BetaNoised 0.1863 0.1163 0.4751 0.4237 0.4783 0.0498 0.0427 0.6178 0.3756 0.5853 OutputLaplace 11.0003 547.71 0.5267 0.4338 0.4075 10.0815 532.70 0.5944 0.4114 0.4152 OutputNorm 12.0701 580.33 0.4096 0.3422 0.4775 11.7703 588.92 0.5555 0.4460 0.4463 DP-MLP 0.0117 0.1249 0.4564 0.4230 0.4959 0.0003 0.0472 0.6048 0.3577 0.5929Spam Discriminator 2.9582 0.1963 0.4945 0.4150 0.4485 0.6185 0.0043 0.4938 0.4781 0.4148

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

149

None 0.9406 0.0548 0.4594 0.5196 0.4910 1.0668 0.0499 0.5515 0.5015 0.4222 BetaNoised 0.0001 0.0155 0.4919 0.5519 0.4878 0.2868 0.0182 0.5089 0.4363 0.4350 OutputLaplace 2.4455 219.44 0.4888 0.4925 0.4609 2.3973 219.83 0.4780 0.4741 0.5212 OutputNorm 2.4401 225.63 0.4851 0.4837 0.4620 2.5196 224.12 0.4502 0.5035 0.4509 DP-MLP 0.0001 0.0102 0.5078 0.5661 0.4788 0.0895 0.0200 0.5267 0.4360 0.4252Credit Discriminator 0.9247 0.0549 0.4597 0.5208 0.4935 1.0555 0.0474 0.5006 0.5030 0.4388 Table 11: Wasserstein-1 distance (WST), maximum mean discrepancy (MD), support vector classifier AUC (SVM), random

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

150

Table 11: Wasserstein-1 distance (WST), maximum mean discrepancy (MD), support vector classifier AUC (SVM), random forest classifier AUC (RF), multi-layer perceptron classifier AUC (MLP) for (= 6;= 105). weightPATEGAN DPGAN WST# MD# SVM" RF" MLP" WST# MD# SVM" RF" MLP" Discriminator 1.5194 0.0670 0.4867 0.1923 0.0898 1.4975 0.0592 0.5260 0.2592 0.1021 PSIS 1.5890 0.0754 0.5978 0.2992 0.1307 1.5209 0.0613 0.4416 0.2365 0.1159Breastcalibrated 1.6098 0.0754 0.5985 0.3156 0.0718 1.5223 0.0613 0.4417 0.2349 0.1306 Discriminator 2.9582 0.1963 0.4945 0.4150 0.4485 0.6185 0.0043 0.4938 0.4781 0.4148

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

2108.10934

151

PSIS 2.9598 0.1960 0.4760 0.3611 0.5284 2.3378 0.0988 0.5997 0.3953 0.5784Spamcalibrated 3.0072 0.1960 0.4771 0.3566 0.5095 2.3060 0.0982 0.5998 0.3972 0.5589 Discriminator 0.9247 0.0549 0.4597 0.5208 0.4935 1.0555 0.0474 0.5006 0.5030 0.4388 PSIS 0:8803 0:0505 0:4507 0:5395 0:5284 0.8723 0.0473 0.5060 0.6121 0.4444Creditcalibrated 0:8890 0:0505 0:4508 0:5365 0:4872 0.8123 0.0003 0.5059 0.6121 0.5101 Table 12: Results for the parameters (= 6:0;= 1e5)(Wasserstein distance, maximum mean discrepancy, support vector classifier ROC-AUC, random forest classifier ROC-AUC, multi-layer perceptron classifier ROC-AUC) 34

2108.10934

Mitigating Statistical Bias within Differentially Private Synthetic Data

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.

http://arxiv.org/pdf/2108.10934

stat.ML

20210824

20220519

1701.08118

0

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis Bj¨orn Ross Michael Rist Guillermo Carbonell Benjamin Cabrera Nils Kurowsky Michael Wojatzki Research Training Group ”User-Centred Social Media” Department of Computer Science and Applied Cognitive Science University of Duisburg-Essen firstname.lastname@uni-due.de Abstract Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability,

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

1

showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation. 1 Introduction Social media are sometimes used to disseminate hateful messages. In Europe, the current surge in hate speech has been linked to the ongoing refugee crisis. Lawmakers and social media sites are increasingly aware of the problem and are developing approaches to deal with it, for example promising to remove illegal messages within 24 hours after they are reported (Titcomb, 2016). This raises the question of how hate speech can be detected automatically. Such an automatic detection method could be used to scan the large amount of text generated on the internet for hateful contentand report it to the relevant authorities. It would also make it easier for researchers to examine the diffusion of hateful content through social media on a large scale. From a natural language processing perspective, hate speech detection can be considered a classification task: given an utterance, determine whether or not it contains hate speech. Training a classifier requires a large amount of data that is unambiguously

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

2

not it contains hate speech. Training a classifier requires a large amount of data that is unambiguously hate speech. This data is typically obtained by manually annotating a set of texts based on whether a certain element contains hate speech. The reliability of the human annotations is essential, both to ensure that the algorithm can accurately learn the characteristics of hate speech, and as an upper bound on the expected performance (Warner and Hirschberg, 2012; Waseem and Hovy, 2016). As a preliminary step, six annotators rated 469 tweets. We found that agreement was very low (see Section 3). We then carried out group discussions to find possible reasons. They revealed that there is considerable ambiguity in existing definitions. A given statement may be considered hate speech or not depending on someone’s cultural background and personal sensibilities. The wording of the question may also play a role. We decided to investigate the issue of reliability further by conducting a more comprehensive study across a large number of annotators, which we present in this paper. Our contribution in this paper is threefold: To the best of our knowledge, this paper presents the first attempt at compiling a German hate speech corpus for the refugee crisis.1 We provide an estimate of the reliability of

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

3

presents the first attempt at compiling a German hate speech corpus for the refugee crisis.1 We provide an estimate of the reliability of hate speech annotations. We investigate how the reliability of the annotations is affected by the exact question asked. 1Available at https://github.com/UCSM-DUE/ IWG_hatespeech_publicarXiv:1701.08118v1 [cs.CL] 27 Jan 2017 2 Hate Speech For the purpose of building a classifier, Warner and Hirschberg (2012) define hate speech as “abusive speech targeting specific group characteristics, such as ethnic origin, religion, gender, or sexual orientation”. More recent approaches rely on lists of guidelines such as a tweet being hate speech if it “uses a sexist or racial slur” (Waseem and Hovy, 2016). These approaches are similar in that they leave plenty of room for personal interpretation, since there may be differences in what is considered offensive. For instance, while the utterance “the refugees will live off our money” is clearly generalising and maybe unfair, it is unclear if this is already hate speech. More precise definitions from

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

4

“the refugees will live off our money” is clearly generalising and maybe unfair, it is unclear if this is already hate speech. More precise definitions from law are specific to certain jurisdictions and therefore do not capture all forms of offensive, hateful speech, see e.g. Matsuda (1993). In practice, social media services are using their own definitions which have been subject to adjustments over the years (Jeong, 2016). As of June 2016, Twitter bans hateful conduct2. With the rise in popularity of social media, the presence of hate speech has grown on the internet. Posting a tweet takes little more than a working internet connection but may be seen by users all over the world. Along with the presence of hate speech, its reallife consequences are also growing. It can be a precursor and incentive for hate crimes, and it can be so severe that it can even be a health issue (Burnap and Williams, 2014). It is also known that hate speech does not only mirror existing opinions in the reader but can also induce new negative feelings towards its targets (Martin et al., 2013). Hate speech has recently gained some interest as a research topic on the one hand – e.g. (Djuric

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

5

feelings towards its targets (Martin et al., 2013). Hate speech has recently gained some interest as a research topic on the one hand – e.g. (Djuric et al., 2014; Burnap and Williams, 2014; Silva et al., 2016) – but also as a problem to deal with in politics such as the No Hate Speech Movement by the Council of Europe. The current refugee crisis has made it evident that governments, organisations and the public share an interest in controlling hate speech in social media. However, there seems to be little consensus on what hate speech actually is. 2“You may not promote violence against or directly attack or threaten other people on the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or disease. We also do not allow accounts whose primary purpose is inciting harm towards others on the basis of these categories.”, The Twitter Rules3 Compiling A Hate Speech Corpus As previously mentioned, there is no German hate speech corpus available for our needs, especially not for the very recent topic of the refugee crisis in Europe. We therefore had to compile our own corpus. We used Twitter as a source as it offers recent comments on current events. In our study we only considered the textual content of tweets

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

6

corpus. We used Twitter as a source as it offers recent comments on current events. In our study we only considered the textual content of tweets that contain certain keywords, ignoring those that contain pictures or links. This section provides a detailed description of the approach we used to select the tweets and subsequently annotate them. To find a large amount of hate speech on the refugee crisis, we used 10 hashtags3that can be used in an insulting or offensive way. Using these hashtags we gathered 13 766 tweets in total, roughly dating from February to March 2016. However, these tweets contained a lot of non-textual content which we filtered out automatically by removing tweets consisting solely of links or images. We also only considered original tweets, as retweets or replies to other tweets might only be clearly understandable when reading both tweets together. In addition, we removed duplicates and near-duplicates by discarding tweets that had a normalised Levenshtein edit distance smaller than .85 to an aforementioned tweet. A first inspection of the remaining tweets indicated that not all search terms were equally suited for our needs. The search term #Pack (vermin or lowlife) found a potentially large amount of hate speech not directly linked to the refugee crisis. It was therefore discarded. As

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

7

term #Pack (vermin or lowlife) found a potentially large amount of hate speech not directly linked to the refugee crisis. It was therefore discarded. As a last step, the remaining tweets were manually read to eliminate those which were difficult to understand or incomprehensible. After these filtering steps, our corpus consists of 541 tweets, none of which are duplicates, contain links or pictures, or are retweets or replies. As a first measurement of the frequency of hate speech in our corpus, we personally annotated them based on our previous expertise. The 541 tweets were split into six parts and each part was annotated by two out of six annotators in order to determine if hate speech was present or not. The annotators were rotated so that each pair of annotators only evaluated one part. Additionally the offensiveness of a tweet was rated on a 6-point Likert scale, the same scale used later in the study. 3#Pack ,#Aslyanten ,#WehrDich ,#Krimmigranten , #Rapefugees ,#Islamfaschisten ,#RefugeesNotWelcome , #Islamisierung ,#AsylantenInvasion ,#Scharia

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

8

#Rapefugees ,#Islamfaschisten ,#RefugeesNotWelcome , #Islamisierung ,#AsylantenInvasion ,#Scharia Even among researchers familiar with the definitions outlined above, there was still a low level of agreement (Krippendorff’s a=:38). This supports our claim that a clearer definition is necessary in order to be able to train a reliable classifier. The low reliability could of course be explained by varying personal attitudes or backgrounds, but clearly needs more consideration. 4 Methods In order to assess the reliability of the hate speech definitions on social media more comprehensively, we developed two online surveys in a betweensubjects design. They were completed by 56 participants in total (see Table 1). The main goal was to examine the extent to which non-experts agree upon their understanding of hate speech given a diversity of social media content. We used the Twitter definition of hateful conduct in the first survey. This definition was presented at the beginning, and again above every tweet. The second survey did not contain any definition. Participants were randomly assigned one of the two surveys.

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

9

and again above every tweet. The second survey did not contain any definition. Participants were randomly assigned one of the two surveys. The surveys consisted of 20 tweets presented in a random order. For each tweet, each participant was asked three questions. Depending on the survey, participants were asked (1)to answer (yes/no) if they considered the tweet hate speech, either based on the definition or based on their personal opinion. Afterwards they were asked (2)to answer (yes/no) if the tweet should be banned from Twitter. Participants were finally asked (3)to answer how offensive they thought the tweet was on a 6-point Likert scale from 1 (Not offensive at all) to 6 (Very offensive). If they answered 4 or higher, the participants had the option to state which particular words they found offensive. After the annotation of the 20 tweets, participants were asked to voluntarily answer an open question regarding the definition of hate speech. In the survey with the definition, they were asked if the definition of Twitter was sufficient. In the survey without the definition, the participants were asked to suggest a definition themselves. Finally,

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

10

survey without the definition, the participants were asked to suggest a definition themselves. Finally, sociodemographic data were collected, including age, gender and more specific information regarding the participant’s political orientation, migration background, and personal position regarding the refugee situation in Europe. The surveys were approved by the ethical committee of the Department of Computer Science andApplied Cognitive Science of the Faculty of Engineering at the University of Duisburg-Essen. 5 Preliminary Results and Discussion Since the surveys were completed by 56 participants, they resulted in 1120 annotations. Table 1 shows some summary statistics. Def. No def. p r Participants 25 31 Age (mean) 33.3 30.5 Gender (% female) 43.5 58.6 Hate Speech (% yes) 32.6 40.3 .26 .15 Ban (% yes) 32.6 17.6 .01 -.32 Offensive (mean) 3.49 3.42 .55 -.08 Table 1: Summary statistics with p values and effect size estimates from WMW tests. Not all participants chose to report their age or gender. To assess whether the definition had any effect, we calculated, for each participant, the percentage of tweets they considered hate speech or suggested

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

11

To assess whether the definition had any effect, we calculated, for each participant, the percentage of tweets they considered hate speech or suggested to ban and their mean offensiveness rating. This allowed us to compare the two samples for each of the three questions. Preliminary Shapiro-Wilk tests indicated that some of the data were not normally distributed. We therefore used the Wilcoxon-MannWhitney (WMW) test to compare the three pairs of series. The results are reported in Table 1. Participants who were shown the definition were more likely to suggest to ban the tweet. In fact, participants in group one very rarely gave different answers to questions one and two (18 of 500 instances or 3.6%). This suggests that participants in that group aligned their own opinion with the definition. We chose Krippendorff’s ato assess reliability, a measure from content analysis, where human coders are required to be interchangeable. Therefore, it measures agreement instead of association, which leaves no room for the individual predilections of coders. It can be applied to any number of coders and to interval as well as nominal data. (Krippendorff, 2004) This allowed us to compare agreement between

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

12

of coders and to interval as well as nominal data. (Krippendorff, 2004) This allowed us to compare agreement between both groups for all three questions. Figure 1 visualises the results. Overall, agreement was very low, ranging from a=:18to:29. In contrast, for the purpose of content analysis, Krippendorff recommends a minimum of a=:80, or a minimum of :66for applications where some uncertainty is un0.00.10.20.3 1 2 3 QuestionAlphaGroup No definition DefinitionFigure 1: Reliability (Krippendorff’s a) for the different groups and questions problematic (Krippendorff, 2004). Reliability did not consistently increase when participants were shown a definition. To measure the extent to which the annotations using the Twitter definition (question one in group one) were in accordance with participants’ opinions (question one in group two), we calculated, for each tweet, the percentage of participants in each group who considered it hate speech, and then calculated Pearson’s correlation coefficient. The two series correlate strongly ( r=:895;p< :0001 ), indicating that they measure the same underlying construct. 6 Conclusion and Future Work

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

13

correlate strongly ( r=:895;p< :0001 ), indicating that they measure the same underlying construct. 6 Conclusion and Future Work This paper describes the creation of our hate speech corpus and offers first insights into the low agreement among users when it comes to identifying hateful messages. Our results imply that hate speech is a vague concept that requires significantly better definitions and guidelines in order to be annotated reliably. Based on the present findings, we are planning to develop a new coding scheme which includes clear-cut criteria that let people distinguish hate speech from other content. Researchers who are building a hate speech detection system might want to collect multiple labels for each tweet and average the results. Of course this approach does not make the original data any more reliable (Krippendorff, 2004). Yet, collecting the opinions of more users gives a more detailed picture of objective (or intersubjective) hatefulness. For the same reason, researchers might want to consider hate speech detection a regression problem, predicting, for example, the degree of hatefulness of a message, instead of a binary yes-or-no classification task.In the future, finding the characteristics that

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

14

of a message, instead of a binary yes-or-no classification task.In the future, finding the characteristics that make users consider content hateful will be useful for building a model that automatically detects hate speech and users who spread hateful content, and for determining what makes users disseminate hateful content. Acknowledgments This work was supported by the Deutsche Forschungsgemeinschaft (DFG) under grant No. GRK 2167, Research Training Group ”UserCentred Social Media”. References Peter Burnap and Matthew Leighton Williams. 2014. Hate Speech, Machine Classification and Statistical Modelling of Information Flows on Twitter: Interpretation and Communication for Policy Decision Making. In Proceedings of IPP 2014 , pages 1–18. Nemanja Djuric, Robin Morris Jing Zhou, Mihajlo Grbovic, Vladan Radosavljevic, and Narayan Bhamidipati. 2014. Hate Speech Detection with Comment Embeddings. In ICML 2014 , volume 32, pages 1188–1196. Sarah Jeong. 2016. The History of Twitter’s Rules. VICE Motherboard .

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

15

1188–1196. Sarah Jeong. 2016. The History of Twitter’s Rules. VICE Motherboard . Klaus Krippendorff. 2004. Reliability in Content Analysis: Some Common Misconceptions and Recommendations. HCR , 30(3):411–433. Ryan C Martin, Kelsey Ryan Coyier, Leah M VanSistine, and Kelly L Schroeder. 2013. Anger on the Internet: the Perceived Value of Rant-Sites. Cyberpsychology, behavior and social networking , 16(2):119– 22. Mari J Matsuda. 1993. Words that Wound - Critical Race Theory, Assaultive Speech, and the First Amendment . Westview Press, New York. Leandro Silva, Mainack Mondal, Denzil Correa, Fabr´ıcio Benevenuto, and Ingmar Weber. 2016. Analyzing the Targets of Hate in Online Social Media. InProceedings of ICWSM 2016 , pages 687–90. James Titcomb. 2016. Facebook and Twitter promise to crack down on internet hate speech. The Telegraph . William Warner and Julia Hirschberg. 2012. Detecting Hate Speech on the World Wide Web. In Proceedings of LSM 2012 , pages 19–26. ACL.

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

1701.08118

16

William Warner and Julia Hirschberg. 2012. Detecting Hate Speech on the World Wide Web. In Proceedings of LSM 2012 , pages 19–26. ACL. Zeerak Waseem and Dirk Hovy. 2016. Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter. In Proceedings of NAACL-HLT , pages 88–93.

1701.08118

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.

http://arxiv.org/pdf/1701.08118

Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Processing for Computer-Mediated Communication (Bochum), Bochumer Linguistische Arbeitsberichte, vol. 17, sep 2016, pp. 6-9

cs.CL

20170127

20170127

2301.11305

0

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature Eric Mitchell1Yoonho Lee1Alexander Khazatsky1Christopher D. Manning1Chelsea Finn1 Abstract The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM’s probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model’s log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

1

(e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See ericmitchell.ai/detectgpt for code, data, and other project information. 1. Introduction Large language models (LLMs) have proven able to generate remarkably fluent responses to a wide variety of user queries. Models such as GPT-3 (Brown et al., 2020), PaLM (Chowdhery et al., 2022), and ChatGPT (OpenAI, 2022) can convincingly answer complex questions about science, mathematics, historical and current events, and social trends. 1Stanford University. Correspondence to: Eric Mitchell <eric.mitchell@cs.stanford.edu >. Proceedings of the 40thInternational Conference on Machine Learning , Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright 2023 by the author(s). Candidate passage : “Joe Biden recently made a move to the White House that included bringing along his pet German Shepherd…” DetectGPT x

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

2

Candidate passage : “Joe Biden recently made a move to the White House that included bringing along his pet German Shepherd…” DetectGPT x ...GPT-3(1)Perturb(2) Score(3) Compare 🤖 from GPT-3 Yes(reword with T5) “made a move” “moved”→“pet” “dog”→Delete “bringing along” ... 🤔 from other source No Figure 1. We aim to determine whether a piece of text was generated by a particular LLM p, such as GPT-3. To classify a candidate passagex, DetectGPT first generates minor perturbations of the passage ~xiusing a generic pre-trained model such as T5. Then DetectGPT compares the log probability under pof the original samplexwith each perturbed sample ~xi. If the average log ratio is high, the sample is likely from the source model. While recent work has found that cogent-sounding LLMgenerated responses are often simply wrong (Lin et al., 2022), the articulate nature of such generated text may still make LLMs attractive for replacing human labor in some contexts, notably student essay writing and journalism. At

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

3

2022), the articulate nature of such generated text may still make LLMs attractive for replacing human labor in some contexts, notably student essay writing and journalism. At least one major news source has released AI-written content with limited human review, leading to substantial factual errors in some articles (Christian, 2023). Such applications of LLMs are problematic for a variety of reasons, making fair student assessment difficult, impairing student learning, and proliferating convincing-but-inaccurate news articles. Unfortunately, humans perform only slightly better than chance when classifying machine-generated vs human-written text (Gehrmann et al., 2019), leading researchers to consider automated detection methods that may identify signals difficult for humans to recognize. Such methods might give teachers and news-readers more confidence in the human origin of the text that they consume. As in prior work (Jawahar et al., 2020), we study the 1arXiv:2301.11305v2 [cs.CL] 23 Jul 2023 Zero-Shot Machine-Generated Text Detection using Probability Curvature machine-generated text detection problem as a binary classification problem. Specifically, we aim to classify whether

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

4

machine-generated text detection problem as a binary classification problem. Specifically, we aim to classify whether acandidate passage was generated by a particular source model . While several works have investigated methods for training a second deep network to detect machine-generated text, such an approach has several shortcomings, including a tendency to overfit to the topics it was trained on as well as the need to train a new model for each new source model that is released. We therefore consider the zero-shot version of machine-generated text detection, where we use the source model itself, without fine-tuning or adaptation of any kind, to detect its own samples. The most common method for zero-shot machine-generated text detection is evaluating the average per-token log probability of the generated text and thresholding (Solaiman et al., 2019; Gehrmann et al., 2019; Ippolito et al., 2020). However, this zeroth-order approach to detection ignores the local structure of the learned probability function around a candidate passage, which we find contains useful information about the source of a passage. This paper poses a simple hypothesis: minor rewrites of model-generated text tend to have lower log probability under the model than the original sample, while minor rewrites

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

5

This paper poses a simple hypothesis: minor rewrites of model-generated text tend to have lower log probability under the model than the original sample, while minor rewrites ofhuman-written text may have higher or lower log probability than the original sample. In other words, unlike human-written text, model-generated text tends to lie in areas where the log probability function has negative curvature (for example, near local maxima of the log probability). We empirically verify this hypothesis, and find that it holds true across a diverse body of LLMs, even when the minor rewrites, or perturbations , come from alternative language models. We leverage this observation to build DetectGPT, a zero-shot method for automated machine-generated text detection. To test if a passage came from a source model p, DetectGPT compares the log probability of the candidate passage under pwith the average log probability of several perturbations of the passage under p(generated with, e.g., T5; Raffel et al. (2020)). If the perturbed passages tend to have lower average log probability than the original by some margin, the candidate passage is likely to have come fromp. See Figure 1 for an overview of the problem and

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

6

to have lower average log probability than the original by some margin, the candidate passage is likely to have come fromp. See Figure 1 for an overview of the problem and DetectGPT. See Figure 2 for an illustration of the underlying hypothesis and Figure 3 for empirical evaluation of the hypothesis. Our experiments find that DetectGPT is more accurate than existing zero-shot methods for detecting machine-generated text, improving over the strongest zero-shot baseline by over 0.1 AUROC for multiple source models when detecting machine-generated news articles. Contributions. Our main contributions are: (a) the identification and empirical validation of the hypothesis that the curvature of a model’s log probability function tends to be significantly more negative at model samples than for human text, and (b) DetectGPT, a practical algorithm inspired by this hypothesis that approximates the trace of the log logp/uni03B8(x) xfake/uni223Cp/uni03B8(x)˜xfake1˜xfake2˜xfake3˜xfake4xreal/uni223Cphuman(x)˜xreal1˜xreal2˜xreal3˜xreal4

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

7

Fake/real samplePerturbed fake/real sampleLog likelihood…logp/uni03B8(x)Figure 2. We identify and exploit the tendency of machinegenerated passages xp(
)(left) to lie in negative curvature regions of logp(x), where nearby samples have lower model log probability on average. In contrast, human-written text xpreal(
)(right) tends not to occupy regions with clear negative log probability curvature; nearby samples may have higher or lower log probability. probability function’s Hessian to detect a model’s samples. 2. Related Work Increasingly large LLMs (Radford et al., 2019; Brown et al., 2020; Chowdhery et al., 2022; OpenAI, 2022; Zhang et al., 2022) have led to dramatically improved performance on many language-related benchmarks and the ability to generate convincing and on-topic text. GROVER (Zellers et al., 2019) was the first LLM trained specifically for generating plausible news articles. Human evaluators found GROVER-generated propaganda at least as trustworthy as human-written propaganda, motivating the authors to study GROVER’s ability to detect its own generations by finetuning a detector on top of its features; they found GROVER

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

8

human-written propaganda, motivating the authors to study GROVER’s ability to detect its own generations by finetuning a detector on top of its features; they found GROVER better able to detect GROVER-generated text than other pretrained models. However, Bakhtin et al. (2019); Uchendu et al. (2020) note that models trained explicitly to detect machine-generated text tend to overfit to their training distribution of data or source models. Other works have trained supervised models for machinegenerated text detection on top of neural representations (Bakhtin et al., 2019; Solaiman et al., 2019; Uchendu et al., 2020; Ippolito et al., 2020; Fagni et al., 2021), bag-of-words features (Solaiman et al., 2019; Fagni et al., 2021), and handcrafted statistical features (Gehrmann et al., 2019). Alternatively, Solaiman et al. (2019) notes the surprising efficacy of a simple zero-shot method for machine-generated text detection, which thresholds a candidate passage based on its average log probability under the generative model, serving as a strong baseline for zero-shot machine-generated text detection in our work. In our work, we similarly use the

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

9

average log probability under the generative model, serving as a strong baseline for zero-shot machine-generated text detection in our work. In our work, we similarly use the generating model to detect its own generations in a zero shot manner, but through a different approach based on estimating local curvature of the log probability around the sample rather than the raw log probability of the sample itself. See Jawahar et al. (2020) for a complete survey on machine2 Zero-Shot Machine-Generated Text Detection using Probability Curvature generated text detection. Other work explores watermarks for generated text (Kirchenbauer et al., 2023), which modify a model’s generations to make them easier to detect. Our work does not assume text is generated with the goal of easy detection; DetectGPT detects text generated from publicly available LLMs using standard LLM sampling strategies. The widespread use of LLMs has led to much other contemporaneous work on detecting LLM output. Sadasivan et al. (2023) show that the detection AUROC of the an detector is upper bounded by a function of the TV distance between the model and human text. However, we find that AUROC of DetectGPT is high even for the largest publiclyavailable models (Table 2), suggesting that TV distance may

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

10

AUROC of DetectGPT is high even for the largest publiclyavailable models (Table 2), suggesting that TV distance may not correlate strongly with model scale and capability. This disconnect may be exacerbated by new training objectives other than maximum likelihood, e.g., reinforcement learning with human feedback (Christiano et al., 2017; Ziegler et al., 2020). Both Sadasivan et al. (2023) and Krishna et al. (2023) show the effectiveness of paraphrasing as a tool for evading detection, suggesting an important area of study for future work. Liang et al. (2023) show that multi-lingual detection is difficult, with non-DetectGPT detectors showing bias against non-native speakers; this result highlights the advantage of zero-shot detectors like DetectGPT, which generalize well to any data generated by the original generating model. Mireshghallah et al. (2023) study which proxy scoring models produce the most useful log probabilities for detection when the generating model is not known (a large-scale version of our Figure 6). Surprisingly (but consistent with our findings), they find that smaller models are in fact better proxy models for performing detection with perturbation-based methods like DetectGPT.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

11

in fact better proxy models for performing detection with perturbation-based methods like DetectGPT. The problem of machine-generated text detection echoes earlier work on detecting deepfakes, artificial images or videos generated by deep nets, which has spawned substantial efforts in detection of fake visual content (Dolhansky et al., 2020; Zi et al., 2020). While early works in deepfake detection used relatively general-purpose model architectures (G¨uera & Delp, 2018), many deepfake detection methods rely on the continuous nature of image data to achieve stateof-the-art performance (Zhao et al., 2021; Guarnera et al., 2020), making direct application to text difficult. 3. The Zero-Shot Machine-Generated Text Detection Problem We study zero-shot machine-generated text detection, the problem of detecting whether a piece of text, or candidate passage x, is a sample from a source model p. The problem is zero-shot in the sense that we do not assume access to human-written or generated samples to perform detection. As in prior work, we study a ‘white box’ setting (Gehrmann et al., 2019) in which the detector may evaluate the log prob-Algorithm 1 DetectGPT model-generated text detection

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

12

et al., 2019) in which the detector may evaluate the log prob-Algorithm 1 DetectGPT model-generated text detection 1:Input: passagex, source model p, perturbation function q, number of perturbations k, decision threshold  2:~xiq(
jx); i2[1::k]// mask spans, sample replacements 3:~ 1 kP ilogp(~xi)// approximate expectation in Eq. 1 4:^dx logp(x)~ // estimate d(x;p;q) 5:~2 x 1 k1P i(logp(~xi)~)2// variance for normalization 6:if^dxp~x>then 7: return true // probably model sample 8:else 9: return false // probably not model sample ability of a sample logp(x). The white box setting does notassume access to the model architecture or parameters. Most public APIs for LLMs (such as GPT-3) enable scoring text, though some exceptions exist, notably ChatGPT. While most of our experiments consider the white box setting, see Section 5.2 for experiments in which we score text using

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

13

text, though some exceptions exist, notably ChatGPT. While most of our experiments consider the white box setting, see Section 5.2 for experiments in which we score text using models other than the source model. See Mireshghallah et al. (2023) for a comprehensive evaluation in this setting. The detection criterion we propose, DetectGPT, also makes use of generic pre-trained mask-filling models in order to generate passages that are ‘nearby’ the candidate passage. However, these mask-filling models are used off-the-shelf, without any fine-tuning or adaptation to the target domain. 4. DetectGPT: Zero-shot Machine-Generated Text Detection with Random Perturbations DetectGPT is based on the hypothesis that samples from a source model ptypically lie in areas of negative curvature of the log probability function of p, unlike human text. In other words, if we apply small perturbations to a passage xp, producing ~x, the quantity logp(x)logp(~x) should be relatively large on average for machine-generated samples compared to human-written text. To leverage this

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

14

should be relatively large on average for machine-generated samples compared to human-written text. To leverage this hypothesis, first consider a perturbation function q(
jx) that gives a distribution over ~x, slightly modified versions of xwith similar meaning (we will generally consider roughly paragraph-length texts x). As an example, q(
jx)might be the result of simply asking a human to rewrite one of the sentences of x, while preserving the meaning of x. Using the notion of a perturbation function, we can define the perturbation discrepancy d(x; p; q): d(x; p; q),logp(x)E~xq(
jx)logp(~x)(1) We state our hypothesis more formally as the Local Perturbation Discrepancy Gap Hypothesis, which describes a gap in the perturbation discrepancy for model-generated text and human-generated text. Perturbation Discrepancy Gap Hypothesis. Ifqproduces samples on the data manifold, d(x; p; q)is positive and

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

15

Perturbation Discrepancy Gap Hypothesis. Ifqproduces samples on the data manifold, d(x; p; q)is positive and large with high probability for samples xp. For humanwritten text, d(x; p; q)tends toward zero for all x. 3 Zero-Shot Machine-Generated Text Detection using Probability Curvature 0.0 0.1 0.2 0.30204060gpt2-xl Human Model 0.0 0.1 0.2 0.3EleutherAI/gpt-neo-2.7B 0.0 0.1 0.2 0.30204060EleutherAI/gpt-j-6B 0.0 0.1 0.2 0.3EleutherAI/gpt-neox-20b 0.0 0.2 0.4 0.6 0.8 1.0 Log Probability Change (Perturbation Discrepancy)0.00.20.40.60.81.0Frequency Figure 3. The average drop in log probability (perturbation discrepancy) after rephrasing a passage is consistently higher for modelgenerated passages than for human-written passages. Each plot shows the distribution of the perturbation discrepancy d(x;p;q)

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

16

shows the distribution of the perturbation discrepancy d(x;p;q) forhuman-written news articles andmachine-generated articlesof equal word length. Human-written articles are a sample of 500 XSum articles; machine-generated text, generated from models GPT-2 (1.5B), GPT-Neo-2.7B (Black et al., 2021), GPT-J (6B; Wang & Komatsuzaki (2021)) and GPT-NeoX (20B; Black et al. (2022)), is generated by prompting each model with the first 30 tokens of each XSum article, sampling from the raw conditional distribution. Discrepancies are estimated with 100 T5-3B samples. If we define q(
jx)to be samples from a mask-filling model such as T5 (Raffel et al., 2020), rather than human rewrites, we can empirically test the Perturbation Discrepancy Gap Hypothesis in an automated, scalable manner. For real data, we use 500 news articles from the XSum dataset (Narayan et al., 2018); for model samples, we use the output of four different LLMs when prompted with the first 30 tokens of

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

17

et al., 2018); for model samples, we use the output of four different LLMs when prompted with the first 30 tokens of each article in XSum. We use T5-3B to apply perturbations, masking out randomly-sampled 2-word spans until 15% of the words in the article are masked. We approximate the expectation in Eq. 1 with 100 samples from T5.1Figure 3 shows the result of this experiment. We find the distribution of perturbation discrepancies is significantly different for human-written articles and model samples; model samples tend to have a larger perturbation discrepancy. Section 5.3 explores a relaxation of the assumption that qonly produces samples on the data manifold, finding that a gap, although reduced, still exists in this case. Given these results, we can detect if a piece of text was generated by a model pby simply thresholding the perturbation discrepancy. In practice, we find that normalizing the perturbation discrepancy by the standard deviation of the observed values used to estimate E~xq(
jx)logp(~x)provides a slightly better signal for detection, typically increasing 1We later show in Figure 8 that varying the number of samples

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

18

a slightly better signal for detection, typically increasing 1We later show in Figure 8 that varying the number of samples used to estimate the expectation effectively allows for trading off between accuracy and speed.AUROC by around 0.020, so we use this normalized version of the perturbation discrepancy in our experiments. The resulting method, DetectGPT, is summarized in Alg. 1. Having described an application of the perturbation discrepancy to machine-generated text detection, we next provide an interpretation of this quantity. Interpretation of perturbation discrepancy as curvature While Figure 3 suggests that the perturbation discrepancy may be useful, it is not immediately obvious what it measures. In this section, we show that the perturbation discrepancy approximates a measure of the local curvature of the log probability function near the candidate passage, more specifically, that it is proportional to the negative trace of the Hessian of the log probability function.2To handle the non-differentiability of discrete data, we consider candidate passages in a latent semantic space, where small displacements correspond to valid edits that retain similar meaning to the original. Because our perturbation function (T5) models natural text, we expect our perturbations to roughly capture such meaningful variations of the original passage, rather than arbitrary edits.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

19

(T5) models natural text, we expect our perturbations to roughly capture such meaningful variations of the original passage, rather than arbitrary edits. We first invoke Hutchinson’s trace estimator (Hutchinson, 1990), giving an unbiased estimate of the trace of matrix A: tr(A) =Ezz>Az (2) provided that the elements of zqzare IID with E[zi] = 0 andVar(zi) = 1 . To use Equation 2 to estimate the trace of the Hessian of fatx, we must therefore compute the expectation of the directional second derivative z>Hf(x)z. We approximate this expression with finite differences: z>Hf(x)zf(x+hz) +f(xhz)2f(x) h2(3) Combining Equations 2 and 3 and simplifying with h= 1, we have an estimate of the negative Hessian trace tr(Hf(x))2f(x)Ez[f(x+z) +f(xz)]:(4) If our noise distribution is symmetric , that is, p(z) =p(z) for all z, then we can simplify Equation 4 to

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

20

If our noise distribution is symmetric , that is, p(z) =p(z) for all z, then we can simplify Equation 4 to tr(Hf(x)) 2f(x)Ezf(x+z): (5) We note that the RHS of Equation 5 corresponds to the perturbation discrepancy (1) where the perturbation functionq(~xjx)is replaced by the distribution qz(z)used in Hutchinson’s trace estimator (2). Here, ~xis a highdimensional sequence of tokens while qzis a vector in a 2Rather than the Hessian of the log likelihood with respect to model parameters (the Fisher Information Matrix), here we refer to the Hessian of the log probability with respect to the sample x. 4 Zero-Shot Machine-Generated Text Detection using Probability Curvature XSum SQuAD WritingPrompts Method GPT-2 OPT-2.7 Neo-2.7 GPT-J NeoX Avg. GPT-2 OPT-2.7 Neo-2.7 GPT-J NeoX Avg. GPT-2 OPT-2.7 Neo-2.7 GPT-J NeoX Avg.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

21

logp(x) 0.86 0.86 0.86 0.82 0.77 0.83 0.91 0.88 0.84 0.78 0.71 0.82 0.97 0.95 0.95 0.94 0.93* 0.95 Rank 0.79 0.76 0.77 0.75 0.73 0.76 0.83 0.82 0.80 0.79 0.74 0.80 0.87 0.83 0.82 0.83 0.81 0.83 LogRank 0.89* 0.88* 0.90* 0.86* 0.81* 0.87* 0.94* 0.92* 0.90* 0.83* 0.76* 0.87* 0.98* 0.96* 0.97* 0.96* 0.95 0.96* Entropy 0.60 0.50 0.58 0.58 0.61 0.57 0.58 0.53 0.58 0.58 0.59 0.57 0.37 0.42 0.34 0.36 0.39 0.38

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

22

DetectGPT 0.99 0.97 0.99 0.97 0.95 0.97 0.99 0.97 0.97 0.90 0.79 0.92 0.99 0.99 0.99 0.97 0.93* 0.97 Diff 0.10 0.09 0.09 0.11 0.14 0.10 0.05 0.05 0.07 0.07 0.03 0.05 0.01 0.03 0.02 0.01 -0.02 0.01 Table 1. AUROC for detecting samples from the given model on the given dataset for DetectGPT and four previously proposed criteria (500 samples used for evaluation). From 1.5B parameter GPT-2 to 20B parameter GPT-NeoX, DetectGPT consistently provides the most accurate detections. Bold shows the best AUROC within each column (model-dataset combination); asterisk (*) denotes the second-best AUROC. Values in the final row show DetectGPT’s AUROC over the strongest baseline method in that column. compact semantic space. Since the mask-filling model samples sentences similar to xwith minimal changes to semantic meaning, we can think of the mask-filling model as first

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

23

sampling a similar semantic embedding ( ~zqz) and then mapping this to a token sequence ( ~z7!~x). Sampling in semantic space ensures that all samples stay near the data manifold, which is useful because we would expect the log probability to always drop if we randomly perturb tokens. We can therefore interpret our objective as approximating the curvature restricted to the data manifold. 5. Experiments We conduct experiments to better understand multiple facets of machine-generated text detection; we study the effectiveness of DetectGPT for zero-shot machine-generated text detection compared to prior zero-shot approaches, the impact of distribution shift on zero-shot and supervised detectors, and detection accuracy for the largest publicly-available models. To further characterize factors that impact detection accuracy, we also study the robustness of zero-shot methods to machine-generated text that has been partially revised, the impact of alternative decoding strategies on detection accuracy, and a black-box variant of the detection task. Finally, we analyze more closely DetectGPT’s behavior as the choice of perturbation function, the number of samples used to estimate d(x; p; q), the length of the passage, and the data distribution is varied.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

24

of samples used to estimate d(x; p; q), the length of the passage, and the data distribution is varied. Comparisons. We compare DetectGPT with various existing zero-shot methods for machine-generated text detection that also leverage the predicted token-wise conditional distributions of the source model for detection. These methods correspond to statistical tests based on token log probabilities, token ranks, or predictive entropy (Gehrmann et al., 2019; Solaiman et al., 2019; Ippolito et al., 2020). The first method uses the source model’s average token-wise log probability to determine if a candidate passage is machinegenerated or not; passages with high average log probability are likely to be generated by the model. The second and third methods use the average observed rank or log-rank of the tokens in the candidate passage according to the model’s conditional distributions. Passages with smaller average(log-)rank are likely machine-generated. We also evaluate an entropy-based approach inspired by the hypothesis in Gehrmann et al. (2019) that model-generated texts will be more ‘in-distribution’ for the model, leading to more over-confident (thus lower entropy) predictive distributions.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

25

be more ‘in-distribution’ for the model, leading to more over-confident (thus lower entropy) predictive distributions. Empirically, we find predictive entropy to be positively correlated with passage fake-ness more often that not; therefore, this baseline uses high average entropy in the model’s predictive distribution as a signal that a passage is machinegenerated. While our main focus is on zero-shot detectors as they do not require re-training for new domains or source models, for completeness we perform comparisons to supervised detection models in Section 5.1, using OpenAI’s RoBERTa-based (Liu et al., 2019) GPT-2 detector models,3 which are fine-tuned on millions of samples from various GPT-2 model sizes and decoding strategies. Datasets & metrics Our experiments use six datasets that cover a variety of everyday domains and LLM use-cases. We use news articles from the XSum dataset (Narayan et al., 2018) to represent fake news detection, Wikipedia paragraphs from SQuAD contexts (Rajpurkar et al., 2016) to represent machine-written academic essays, and prompted stories from the Reddit WritingPrompts dataset (Fan et al., 2018) to represent detecting machine-generated creative

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

26

represent machine-written academic essays, and prompted stories from the Reddit WritingPrompts dataset (Fan et al., 2018) to represent detecting machine-generated creative writing submissions. To evaluate robustness to distribution shift, we also use the English and German splits of WMT16 (Bojar et al., 2016) as well as long-form answers written by human experts in the PubMedQA dataset (Jin et al., 2019). Each experiment uses between 150 and 500 examples for evaluation, as noted in the text. For each experiment, we generate the machine-generated text by prompting with the first 30 tokens of the real text (or just the question tokens for the PubMedQA experiments). We measure performance using the area under the receiver operating characteristic curve (AUROC), which can be interpreted as the probability that a classifier correctly ranks a randomly-selected positive (machine-generated) example higher than a randomlyselected negative (human-written) example. All experiments use an equal number of positive and negative examples. 3https://github.com/openai/gpt-2-outputdataset/tree/master/detector 5 Zero-Shot Machine-Generated Text Detection using Probability Curvature RoB-baseRoB-lg LikelihoodRank

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

27

5 Zero-Shot Machine-Generated Text Detection using Probability Curvature RoB-baseRoB-lg LikelihoodRank LogRankDetectGPT(Ours)0.50.60.70.80.91.0 0.9810.997 0.889 0.8000.9150.991XSum GPT-2 Detection RoB-baseRoB-lg LikelihoodRank LogRankDetectGPT(Ours)0.8880.946 0.838 0.7950.8630.957WMT16-en mGPT Detection RoB-baseRoB-lg LikelihoodRank LogRankDetectGPT(Ours)0.50.60.70.80.91.0 0.6040.7130.768 0.6640.7730.836PubMedQA PubMedGPT Detection Supervised Unsupervised RoB-baseRoB-lg LikelihoodRank LogRankDetectGPT(Ours)0.3940.5370.7950.8380.8610.962WMT16-de mGPT Detection 0.0 0.2 0.4 0.6 0.8 1.0

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

28

0.0 0.2 0.4 0.6 0.8 1.0 Detection Method0.00.20.40.60.81.0Detection AUROC Figure 4. Supervised machine-generated text detection models trained on large datasets of real and generated texts perform as well as or better than DetectGPT on in-distribution (top row) text. However, zero-shot methods work out-of-the-box for new domains (bottom row) such as PubMed medical texts and German news data from WMT16. For these domains, supervised detectors fail due to excessive distribution shift. Hyperparameters. The key hyperparameters of DetectGPT are the fraction of words masked for perturbation, the length of the masked spans, the model used for mask filling, and the sampling hyperparameters for the maskfilling model. Using BERT (Devlin et al., 2019) masked language modeling as inspiration, we use 15% as the mask rate. We performed a small sweep over masked span lengths off2;5;10gon a held-out set of XSum data, finding 2 to perform best. We use these settings for all experiments, without re-tuning . We use T5-3B for almost all experiments,

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

29

except for GPT-NeoX and GPT-3 experiments, where compute resources allowed for the larger T5-11B model; we also use mT5-3B instead of T5-3B for the WMT multilingual experiment. We do not tune the hyperparameters for the mask filling model, sampling directly with temperature 1. 5.1. Main Results We first present two groups of experiments to evaluate DetectGPT along with existing methods for zero-shot and supervised detection on models from 1.5B to 175B parameters. Zero-shot machine-generated text detection. We present the comparison of different zero-shot detection methods in Table 1. In these experiments, model samples are generated by sampling from the raw conditional distribution with temperature 1. DetectGPT most improves average detection accuracy for XSum stories (0.1 AUROC improvement) and SQuAD Wikipedia contexts (0.05 AUROC improvement). While it also performs accurate detection for WritingPrompts, the performance of all methods tends to increase,PMQA XSum WritingP Avg. RoB-base 0.64 / 0.58 0.92 / 0.74 0.92 / 0.81 0.77 RoB-large 0.71 / 0.64 0.92 /0.88 0.91 / 0.88 0.82

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

30

RoB-large 0.71 / 0.64 0.92 /0.88 0.91 / 0.88 0.82 logp(x) 0.64 / 0.55 0.76 / 0.61 0.88 / 0.67 0.69 DetectGPT 0.84 /0.77 0.84 / 0.84 0.87 / 0.84 0.83 Table 2. DetectGPT detects generations from GPT-3 and Jurassic-2 Jumbo (175B models from OpenAI and AI21 Labs) with average AUROC on-par with supervised models trained specifically for machine-generated text detection. For more ‘typical’ text, such as news articles, supervised methods perform strongly. The GPT3 AUROC appears first in each column, the Jurassic-2 AUROC appears second (i.e., after the slash). and the average margin of improvement is narrow.4For 14 of the 15 combinations of dataset and model, DetectGPT provides the most accurate detection performance, with a 0.06 AUROC improvement on average. Log-rank thresholding proves to be a consistently stronger baseline than log probability thresholding, although it requires slightly more information (full predicted logits), which are not always available in public APIs.

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723

2301.11305

31

probability thresholding, although it requires slightly more information (full predicted logits), which are not always available in public APIs. Comparison with supervised detectors. While our experiments generally focus on zero-shot detection, some works have evaluated the detection performance of supervised methods (typically fine-tuned transformers) for detecting machine-generated text. In this section, we explore several domains to better understand the relative strengths of supervised and zero-shot detectors. The results are presented in Figure 4, using 200 samples from each dataset for evaluation. We find that supervised detectors can provide similar detection performance to DetectGPT on in-distribution data like English news, but perform significantly worse than zeroshot methods in the case of English scientific writing and fail altogether for German writing. This finding echoes past work showing that language models trained for machinegenerated text detection overfit to their training data (source model, decoding strategy, topic, language, etc.; Uchendu et al. (2020); Ippolito et al. (2020); Jawahar et al. (2020)). In contrast, zero-shot methods generalize relatively easily to new languages and domains; DetectGPT’s performance in particular is mostly unaffected by the change in language

2301.11305

DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature

The increasing fluency and widespread usage of large language models (LLMs) highlight the desirability of corresponding tools aiding detection of LLM-generated text. In this paper, we identify a property of the structure of an LLM's probability function that is useful for such detection. Specifically, we demonstrate that text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function. Leveraging this observation, we then define a new curvature-based criterion for judging if a passage is generated from a given LLM. This approach, which we call DetectGPT, does not require training a separate classifier, collecting a dataset of real or generated passages, or explicitly watermarking generated text. It uses only log probabilities computed by the model of interest and random perturbations of the passage from another generic pre-trained language model (e.g., T5). We find DetectGPT is more discriminative than existing zero-shot methods for model sample detection, notably improving detection of fake news articles generated by 20B parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95 AUROC for DetectGPT. See https://ericmitchell.ai/detectgpt for code, data, and other project information.

http://arxiv.org/pdf/2301.11305

ICML 2023

cs.CL

20230126

20230723