Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

2108.10934

1

The prevalence of sensitive datasets, such as electronic health records, contributes to a growing concern for violations of an individual’s privacy. In recent years, the notion of Differential Privacy (Dwork et al., 2006) has gained popularity as a privacy metric offering statistical guarantees. This framework bounds how much the likelihood of a randomised algorithm can differ under neighbouring real datasets. We say two datasets DandD0are neighbouring when they differ by at most one observation. A randomised algorithm g:M!R satisfies (;)-differential privacy for ;0 if and only if for all neighbouring datasets D;D0and all subsetsSR , we have Pr(g(D)2S)+ePr(g(D0)2S): The parameter is referred to as the privacy budget; smaller quantities imply more private algorithms.Injecting noise into sensitive data according to this paradigm allows for datasets to be published in a private manner. With the rise of generative modelling approaches, such as Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), there has been a surge of literature proposing generative models for differentially private (DP) synthetic data generation and release (Jordon et al., 2019; Xie et al., 2018;

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

2

2014), there has been a surge of literature proposing generative models for differentially private (DP) synthetic data generation and release (Jordon et al., 2019; Xie et al., 2018; Zhang et al., 2017). These generative models often fail to capture the true underlying distribution of the real data, possibly due to flawed parametric assumptions and the injection of noise into their training and release mechanisms. The constraints imposed by privacy-preservation can lead to significant differences between nature’s true data generating process (DGP) and the induced synthetic DGP (SDGP) (Wilde et al., 2020). This increases the bias of estimators trained on data from the SDGP which reduces their utility. Recent literature has proposed techniques to decrease this bias by modifying the training processes of private algorithms. These approaches are specific to a particular synthetic data generating method (Zhang et al., 2018; Frigerio et al., 2019; Neunhoeffer et al., 2020), or are query-based (Hardt and Rothblum, 2010; Liu et al., 2021) and are thus not generally applicable. Hence, we propose several postprocessing approaches that aid mitigating the bias induced by the DP synthetic data. While there has been extensive research into estimating models directly on protected data without leaking privacy, we

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

3

by the DP synthetic data. While there has been extensive research into estimating models directly on protected data without leaking privacy, we argue that releasing DP synthetic data is crucial for rigorous statistical analysis. This makes providing a framework to debias inference on this an important direction of future research that goes beyond the applicability of any particular DP estimator. Because of the post-processing theorem (Dwork et al., 2014), any function on the DP synthetic data is itself DP. This allows deployment of standard statistical analysis tooling that may otherwise be unavailable for DP estimation. These include 1) exploratory data analysis, 2) model verification and analysis of model diagnostics, 3) private release of (newly developed) models for which no DP analogue has been derived, 4) the computation of con1arXiv:2108.10934v3 [stat.ML] 19 May 2022 fidence intervals of downstream estimators through the nonparametric bootstrap, and 5) the public release of a data set to a research community whose individual requests would otherwise overload the data curator. This endeavour could facilitate the release of data on public platforms like the UCI Machine Learning Repository (Lichman, 2013) or the creation of data competitions, fuelling research growth for specific modelling areas.

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

4

UCI Machine Learning Repository (Lichman, 2013) or the creation of data competitions, fuelling research growth for specific modelling areas. This motivates our main contributions, namely the formulation of multiple approaches to generating DP importance weights that correct for synthetic data’s issues. In particular, this includes: •The bias estimation of an existing DP importance weight estimation method, and the introduction of an unbiased extension with smaller variance (Section 3.3). •An adjustment to DP Stochastic Gradient Descent’s sampling probability and noise injection to facilitate its use in the training of DP-compliant neural networkbased classifiers to estimate importance weights from combinations of real and synthetic data (Section 3.4). •The use of discriminator outputs of DP GANs as importance weights that do not require any additional privacy budget (Section 3.5). •An application of importance weighting to correct for the biases incurred in Bayesian posterior belief updating with synthetic data motivated by the results from (Wilde et al., 2020) and to exhibit our methods’ wide applicability in frequentist and Bayesian contexts (Section 3.1). 2 BACKGROUND Before we proceed, we provide some brief background on bias mitigation in non-private synthetic data generation.

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

5

2 BACKGROUND Before we proceed, we provide some brief background on bias mitigation in non-private synthetic data generation. 2.1 DENSITY RATIOS FOR NON-PRIVATE GANS Since their introduction, GANs have become a popular tool for synthetic data generation in semi-supervised and unsupervised settings. GANs produce realistic synthetic data by trading off the learning of a generator Geto produce synthetic observations, with that of a classifier Dilearning to correctly classify the training and generated data as real or fake. The generator Getakes samples from the prior upu as an input and generates samples Ge(u)2X. The discriminatorDitakes an observation x2Xas input and outputs the probability Di(x)of this observation being drawn from the true DGP. The classification network Didistinguishes between samples from the DGP with label y= 1and distributionpD, and data from the SDGP with label y= 0and distributionpG. Following Bayes’ rule we can show that the output ofDi(x), namely the probabilities bp(y= 1jx)andbp(y= 0jx), can be used for importance weight estimation: bpD(x) bpG(x)=bp(xjy= 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

6

bpD(x) bpG(x)=bp(xjy= 1) bp(xjy= 0)=bp(y= 1jx) bp(y= 0jx)bp(y= 0) bp(y= 1): (1) This observation has been exploited in a stream of literature focusing on importance weighting (IW) based sampling approaches for GANs. Grover et al. (2019) analyse how importance weights of the GAN’s outputs can lead to performance gains; extensions include their proposed usage in rejection sampling on the GAN’s outputs (Azadi et al., 2018), and Metropolis–Hastings sampling from the GAN alongside improvements to the robustness of this sampling via calibration of the discriminator (Turner et al., 2019). To date, no one has leveraged these discriminator-based IW approaches in DP settings where the weights can mitigate the increased bias induced by privatised data models. 2.2 DIFFERENTIAL PRIVACY IN SYNTHETIC DATA GENERATION Private synthetic data generation through DP GANs is built upon the post processing theorem: If Diis(;)- DP, then any composition DiGeis also (;)-DP (Dwork et al.,

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

7

upon the post processing theorem: If Diis(;)- DP, then any composition DiGeis also (;)-DP (Dwork et al., 2014) sinceGedoes not query the protected data. Hence, to train private GANs, we only need to privatise the training of their discriminators, see e.g. Hyland et al. (2018). Xie et al. (2018) propose DPGAN, a Wasserstein GAN which is trained by injecting noise to the gradients of the discriminator’s parameters. In contrast, Jordon et al. (2019) privatise the GAN discriminator by using the Private Aggregation of Teacher Ensembles algorithm. Recently, Torkzadehmahani et al. (2019) proposed DPCGAN as a conditional variant to DPGAN that uses an efficient moments accountant. In contrast, PrivBayes (Zhang et al., 2017) learns a DP Bayesian network and does not rely on a GAN-architecture. Other generative approaches, for instance, include Chen et al. (2018); Acs et al. (2018). See Abay et al. (2018); Fan (2020) for an extensive overview of more DP generative approaches. Differentially private bias mitigation In this paper, we

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

8

extensive overview of more DP generative approaches. Differentially private bias mitigation In this paper, we offer an augmentation to the usual release procedure for synthetic data by leveraging true and estimated importance weights. Most related to our work are the contributions from Elkan (2010) and Ji and Elkan (2013) who train a regularised logistic regression model and assign weights based on the Laplace-noise-contaminated coefficients of the logistic regression. In follow up work, Ji et al. (2014) propose to modify the update step of the Newton-Raphson optimisation algorithm used in fitting the logistic regression classifier to achieve DP. However, neither of these generalise well to more complex and high dimensional settings because of the linearity of the classifier. Further, the authors assume the existence of a public dataset while we consider the case where we first generate DP synthetic data and then weight them a posteriori, providing a generic and universally 2 applicable approach. The benefit of learning a generative model over using public data include on the one hand that there is no requirement for the existence of a public data set, and on the other hand the possibility to generate new data

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

9

model over using public data include on the one hand that there is no requirement for the existence of a public data set, and on the other hand the possibility to generate new data points. This distinction necessitates additional analysis as the privacy budget splits between the budget spent on fitting the SDGP and the budget for estimating the IW approach. Furthermore, we show that the approach from Ji and Elkan (2013) leads to statistically biased estimation and formulate an unbiased extension with improved properties. 3 DIFFERENTIAL PRIVACY AND IMPORTANCE WEIGHTING From a decision theoretic perspective, the goal of statistics is estimating expectations of functions h:X7!R, e.g. loss or utility functions, w.r.t the distribution of future uncertainties xpD. Given data fromfx0 1;:::;x0 NDg=:x0 1:NDi.i.d.pD the data analyst can estimate these expectations consistently via the strong law of large numbers as ExpD(h(x)) 1 NDPND i=1h(x0 i):However, under DP constraints the data analyst is no longer presented with a sample from the true DGPx0

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

10

NDPND i=1h(x0 i):However, under DP constraints the data analyst is no longer presented with a sample from the true DGPx0 1:NDi.i.d.pDbut with a synthetic data sample x1:NG from the SDGP pG. Applying the naive estimator in this scenario biases the downstream tasks as1 NGPNG i=1h(xi)! ExpG(h(x))almost surely. This bias can be mitigated using a standard Monte Carlo method known as importance weighting (IW). Suppose we had access to the weights w(x) :=pD(x) pG(x). IfpG(
)>0 wheneverh(
)pD(
)>0, then IW relies on ExpD[h(x)] =ExpG[w(x)h(x)]: (2) So we have almost surely for x1:NGi.i.d.pGthe convergence IN(hjw) :=1 NGNGX i=1w(xi)h(xi)NG!1!ExpD[h(x)]: 3.1 IMPORTANCE WEIGHTED EMPIRICAL RISK MINIMISATION

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

11

3.1 IMPORTANCE WEIGHTED EMPIRICAL RISK MINIMISATION A downstream task of particular interest is the use of x0 1:NDpDto learn a predictive model, f(
)2F , for the data generating distribution pDbased on empirical risk minimisation. Given a loss function h:FX7!Rcomparing models f(
)2F with observations x2Xand data x0 1:NDpD, the principle of empirical risk minimisation (Vapnik, 1991) states that the optimal bfis given by the minimisation of 1 NDNDX i=1h(f(
);x0 i)ExpD[h(f(
);x)]overf. Maximum likelihood estimation (MLE) is a special case of the above with h(f(
);xi) =logf(xij)for a class of densities fparameterised by . Given synthetic datax1:NGpG, Equation (2)can be used to debias the learning off. Remark 1 (Supplement B.5) .Minimisation of the importance weight adjusted log-likelihood, w(xi) logf(xij),

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

12

learning off. Remark 1 (Supplement B.5) .Minimisation of the importance weight adjusted log-likelihood, w(xi) logf(xij), can be viewed as an M-estimator (e.g. Van der Vaart, 2000) with clear relations to the standard MLE. Bayesian Updating. Wilde et al. (2020) showed that naively conducting Bayesian updating using DP synthetic data without any adjustment could have negative consequences for inference. To show the versatility of our approach and to address the issues they pointed out, we demonstrate how IW can help mitigate this. The posterior distribution for parametergiven ~x0:=x0 1:NDpDis (j~x0)/()NDY i=1f(x0 ij) =() exp NDX i=1logf(x0 ij)! where()denotes the prior distribution for . This posterior is known to learn about model parameter KLD pD:= arg minKLD (pDjjf(
j))(Berk, 1966; Bissiri et al., 2016) where KLD denotes the Kullback-Leibler divergence.

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

13

2016) where KLD denotes the Kullback-Leibler divergence. Given only synthetic data ~x:=x1:NGfrom the ‘proposal distribution’ pG, we can use the importance weights defined in Equation (2)to construct the (generalised) posterior distribution IW(j~x)/() exp NGX i=1w(xi) logf(xij)! :(3) In fact, Equation (3)corresponds to a generalised Bayesian posterior (Bissiri et al., 2016) with `IW(xi;) := w(xi) logf(xij), providing a coherent updating of beliefs about parameter KLD pDusing only data from the SDGP. Theorem 1 (Supplement B.6) .The importance weighted Bayesian posterior IW(jx1:NG), defined in Equation (3)forx1:NGi.i.d.pG, admits the same limiting Gaussian distribution as the Bayesian posterior (jx0 1:ND) wherex0 1:NDi.i.d.pD, under regularity conditions as in

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

14

1:ND) wherex0 1:NDi.i.d.pD, under regularity conditions as in (Chernozhukov and Hong, 2003; Lyddon et al., 2018). It is necessary here to acknowledge the existence of methods to directly conduct privatised Bayesian updating (e.g. Dimitrakakis et al., 2014; Foulds et al., 2016; Wang et al., 2015) or M-estimation (Avella-Medina, 2021). We refer the reader Section 1 for why the attention of this paper focuses on downstream tasks for private synthetic data. We consider the application of DP IW to Bayesian updating as a natural example of such a task. 3 3.2 ESTIMATING THE IMPORTANCE WEIGHTS The previous section shows that IW can be used to recalibrate inference for synthetic data. Unfortunately, both the DGPpDand SDGPpGdensities are typically unknown, e.g. due to the intractability of GAN generation, and thus the ‘perfect’ weight w(x)cannot be calculated. Instead, we must rely on estimates of these weights, bw(x). In this section, we show that the existing approach to DP importance weight estimation is biased, and how the data curator can correct it.

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

15

weight estimation is biased, and how the data curator can correct it. Using the same reasoning as in Section 2.1, we argue that any calibrated classification method that learns to distinguish between data from the DGP, labelled thenceforth with y= 1, and from the SDGP, labelled with y= 0, can be used to estimate the likelihood ratio (Sugiyama et al., 2012). Using Equation (1), we compute bw(x) =bp(y= 1jx) bp(y= 0jx)ND NG wherebpare the probabilities estimated by such a classification algorithm. To improve numerical stability, we can also express the log weights as logbw(x) =1(bp(y= 1jx)) + logND NG; where(x) := (1 + exp(x))1is the logistic function and1(bp(y= 1jx))are the logits of the classification method. We will now discuss two such classifiers: logistic regression and neural networks. 3.3 PRIVATISING LOGISTIC REGRESSION DP guarantees for a classification algorithm gcan be achieved by adding noise to the training procedure. The

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

16

DP guarantees for a classification algorithm gcan be achieved by adding noise to the training procedure. The scale of this noise is determined by how much the algorithm differs when one observation of the dataset changes. In more formal terms, the sensitivity of gw.r.t a normj
jis defined by the smallest number S(g)such that for any two neighbouring datasets DandD0it holds that jg(D)g(D0)jS(g): Dwork et al. (2006) show that to ensure the differential privacy ofg, it suffices to add Laplacian noise with standard deviationS(g)=tog. Possibly the simplest classifier gone could use to estimate the importance weights is logistic regression with L2regularisation. It turns out this also has a convenient form for its sensitivity. If the data is scaled to a range from 0to1such thatX[0;1]d, Chaudhuri et al. (2011) show that the L2 sensitivity of the optimal coefficient vector estimated by b in a regularised logistic regression with model bp(y= 1jxi) =(bTxi) =

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

17

in a regularised logistic regression with model bp(y= 1jxi) =(bTxi) = 1 +ebTxi1isS(b) = 2p d=(ND)whereis the coefficient of the L2regularisation term added to the loss during training. For completeness, when the logistic regression contains an intercept parameter, we let xidenote the concatenation of the feature vector and the constant 1. Ji and Elkan (2013) propose to compute DP importance weights by training such an L2regularised logistic classifier on the private and the synthetic data, and perturb the coefficient vector bwith Laplacian noise. For a ddimensional noise vector withji:i:d:Laplace (0;)with = 2p d=(ND)forj2f1;:::;dg, the private regression coefficient is then =b+, akin to adding heteroscedastic noise to the private estimates of the log weights logw(xi) =Txi=bTxi+xi: (4) The resulting privatised importance weights can be shown

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

18

logw(xi) =Txi=bTxi+xi: (4) The resulting privatised importance weights can be shown to lead to statistically biased estimation. Proposition 1 (Supplement B.1) .Letwdenote the importance weights computed by noise perturbing regression coefficients as in Equation (4)(Ji and Elkan, 2013, Algorithm 1). The IS estimator IN(hjw)is biased. Introducing bias on downstream estimators of sensitive information is undesirable as it can lead to an increased expected loss. To address this issue, we propose a way for the data curator to debias the weights after computation. Proposition 2 (Supplement B.2) .Letwdenote the importance weights computed by noise perturbing the regression coefficients as in Equation (4)(Ji and Elkan, 2013, Algorithm 1) where can be sampled from any noise distribution that ensures (;)-differential privacy of . Define b(xi) := 1=Ep[exp Txi
]; and adjusted importance weight w(xi) =w(xi)b(xi) =bw(xi) exp Txi
b(xi):(5)

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

19

and adjusted importance weight w(xi) =w(xi)b(xi) =bw(xi) exp Txi
b(xi):(5) The importance sampling estimator IN(hjw)is unbiased and(;)-DP for Ep[exp Txi
]>0. In Supplement B.2.4, we further show that our approach does not only decrease the bias, but also the variance of the importance weighted estimators. For the case of component-wise independent Laplace perturbationsji:i:d:Laplace (0;), we show that the bias correction term can be computed as b(xi) =dY j=1 12x2 ij
;providedjxijj<1=8j: In practice, e.g. as we observe empirically in Section 4, the optimal choice of the regularisation term is sufficiently 4 large such that <1. Since the data is scaled to a range of 0 to 1 (Chaudhuri et al., 2011), this bias correction method is not limited by the restriction jxijj<1=;8j. If the data

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

20

0 to 1 (Chaudhuri et al., 2011), this bias correction method is not limited by the restriction jxijj<1=;8j. If the data curator still encounters a case where this condition is not fulfilled, they can choose to perturb the weights with Gaussian noise instead, in which case the bias correction term always exists (see Supplement B.2.2). Laplacian perturbations are however preferred as the required noise scale can be expressed analytically without additional optimisation (Balle and Wang, 2018), and as they give stricter privacy guarantees with = 0. Alternatively, unbiased importance weighted estimates can be computed directly by noising the weights instead of the coefficients of the logistic regression. While this procedure removes the bias of the estimates and can also be shown to be consistent, it increases the variance to a greater extent than noising the coefficients does, and is thus only sustainable when small amounts of data are released. Please refer to Supplement A.1 for more details. 3.4 PRIVATISING NEURAL NETWORKS If logistic regression fails to give accurate density ratio estimates, for example because of biases introduced by the

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

21

3.4 PRIVATISING NEURAL NETWORKS If logistic regression fails to give accurate density ratio estimates, for example because of biases introduced by the classifier’s linearity assumptions, a more complex discriminator in the form of a neural network can be trained. We can train DP classification neural networks for the aim of likelihood ratio estimation with stochastic gradient decent (SGD) by clipping the gradients and adding calibrated Gaussian noise at each step of the SGD, see e.g. Abadi et al. (2016). The noised gradients are then added up in a lotbefore the descent step where lots resemble mini-batches. These optimisation algorithms are commonly formulated for the case when the complete dataset is private. However, in our setting, NDobservations are private and NG observations are non-private. Thus, we can define a relaxed version of DP SGD. Algorithm 1 provides an overview of our proposed method. We highlight the modifications to Algorithm 1 from Abadi et al. (2016) in blue. Proposition 3. Each step in the SGD outlined in Algorithm 1 is (;)-differentially private w.r.t the lot 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

22

Proposition 3. Each step in the SGD outlined in Algorithm 1 is (;)-differentially private w.r.t the lot and (O(q);)differentially private w.r.t the full dataset where q=L ND+NGand=q 2 log (1:25 )=. The differential privacy w.r.t a lot follows directly from the observation that the gradients of the synthetic data are already private. Further, the labels of the synthetic data are public knowledge. Lastly, the differential privacy w.r.t the dataset follows from the amplification theorem (Kasiviswanathan et al., 2011), the fact that sampling one particular private observation within a lot of size Lisq=L ND+NG, and the reasoning behind the moment accountant of Abadi et al. (2016). We still clip the gradients of the public datasetAlgorithm 1: Relaxed DP SGD Input: Examplesx1:ND;y1:NDfrom the DGP and xND+1:ND+NG;yND+1:ND+NGfrom the SDGP, loss function L() =1 NG+NDP iL(;xi;yi). Parameters: learning rate t, noise scale , expected lot size

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

23

L() =1 NG+NDP iL(;xi;yi). Parameters: learning rate t, noise scale , expected lot size L, gradient norm bound C. 1Initialise0randomly 2fort2[T]do 3 Construct a random subset Ltf1;:::;ND+NGgby including each index independently at random with probabilityL ND+NG 4 Compute gradient 5 For eachi2Lt, compute gt(xi;yi)
tL(t;xi;yi) 6 Clip gradient 7gt(xi;yi) gt(xi;yi)=max(1;jjgt(xi;yi)jj2 C) 8 Add noise 9 ~gt 1 LP i2Lt(gt(xi;yi) +N(0;2C2I)1(yi=1)), where 1(yi=1)is 1 ifyi= 1and 0 otherwise 10 Descent 11t+1 t+t~gt Output:Tand the overall privacy cost (;)using the moment’s accountant of Abadi et al. (2016) with sampling probability q=L ND+NG.

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

24

moment’s accountant of Abadi et al. (2016) with sampling probability q=L ND+NG. as their influence will otherwise be overproportional under strong maximum norm assumptions. 3.5 GAN DISCRIMINATOR WEIGHTS The downside of the aforementioned likelihood ratio estimators (Equation (4), Equation (5), Algorithm 1) is that their training requires an additional privacy budget which has to be added to the privacy budget used to learn the SDGP. If we however use a GAN such as DPGAN or PATE-GAN for private synthetic data generation, we can use the GAN’s discriminator for the computation of the importance weights. According to the post processing theorem, these importance weights can be released without requiring an additional privacy budget. In contrast to the weights computed from DP classification networks, this approach is more robust and requires less hyperparameter tuning (confer to Section 4). 4 EXPERIMENTS We demonstrate the benefits of using debiased IW for DP data release with a large-scale experimental study comparing three different SDGPs (DPGAN, DPCGAN, PrivBayes) on six real-world data sets (Iris, TGFB, Boston, 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

25

on six real-world data sets (Iris, TGFB, Boston, Breast, Banknote, MNIST) for two different privacy budgets, 2 f1;6g. We stress that debiasing comes with little overhead 5 0.5 0.0 0.50250500750None 20 00200400True 2.5 0.0 2.5050100150LogReg 2.5 0.0050100BetaNoised 2.5 0.0050100150BetaDebiased 5 00100200300MLP 5 00200400DP-MLP 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00 0.0 0.5 1.00.000.250.500.751.00Figure 1: Kernel density plots of 100 observations sampled from a two dimensional uniform square distribution as SDGP

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

26

0.0 0.5 1.00.000.250.500.751.00Figure 1: Kernel density plots of 100 observations sampled from a two dimensional uniform square distribution as SDGP (bottom left) and a uniform triangle distribution as DGP (second figure in second row). The first row depicts histograms of the computed weights starting with the true importance weights (True). The DP weights were privatised with = 1, and the regularisation was chosen as = 0:1. The second row illustrates the importance weighted synthetic observations. We observe that while BetaDebiased corrects the weights of the logistic regression, the complex nature of the MLPs allows a better modelling of the DGP even in this simple setting. ϵ = 1 ϵ = 6 None LogReg MLP BetaNoised BetaDebiased DP-MLP Discriminator0.40.60.81.0 0.40.60.81.0 IW MethodROC-AUCSynth Model CGAN DPCGAN PRIVBAYES Figure 2: ROC-AUC score distributions calculated via chains of parameters sampled from a Bayesian logistic regression model fit on synthesised Banknote data across 10 seeds. to the actual computations. As we see in Supplement C.2,

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

27

seeds. to the actual computations. As we see in Supplement C.2, the computations of the logistic regression and neural network importance weight estimates take less than one and a half minutes to train, even on MNIST. These weight estimators can be applied to any kind of synthetic data generation model, while the importance weights of the GAN discriminator can be computed in a single line of Python code and do not require any additional concerns regarding the privacy budget. Computation of importance weights After fitting the SDGP on the scaled true data, we weight each synthetic observation with importance weights. Based on the train and the synthetic data, we apply one of the following IW approaches: weights computed from a non-private logisticregression (LogReg), its DP alternative introduced by Ji and Elkan (2013) (BetaNoised), or our debiased proposal (BetaDebiased), and likelihood ratios estimated by a non-private multi-layer perceptron (MLP), or a DP-MLP trained using Algorithm 1. We also compare to the naive estimator using uniform weights without IW (called ’None’). Please refer to Supplement C.1 for more details on the implementation and the hyperparameters used in our experiments.

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

28

uniform weights without IW (called ’None’). Please refer to Supplement C.1 for more details on the implementation and the hyperparameters used in our experiments. In Supplement C.8, we provide a comparison to the experimental results reported by related papers. Because of the large scale of our experimental study, we present only the most important results in this section, and give a complete overview in Supplement C. The code and data for all experiments can be found in the Supplements, and will be made available online. 4.1 TOY EXAMPLE We start our analysis with a simple example to illustrate the benefits of the different weighting schemes. We assume that the synthetic data is sampled from a two-dimensional uniform distribution from 0 to 1 whereas the true data follows a uniform distribution on the lower triangle given by x1+x2<1forx1;x22[0;1]. This illustrative toy example was chosen for a fairer comparison of the logistic regression and the neural network based approaches. As we see in Figure 1, the weighted kernel density estimate (KDE) of BetaDebiased is closer to the LogReg weighted KDE, and also the true KDE compared to the BetaNoised KDE. 4.2 UCI DATA SETS

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

29

of BetaDebiased is closer to the LogReg weighted KDE, and also the true KDE compared to the BetaNoised KDE. 4.2 UCI DATA SETS Datasets and preprocessing We performed additional experiments on four UCI datasets of different characteristics as decribed in Supplement C.1: Iris, Banknote, Boston, and Breast. Similarly to Chaudhuri et al. (2011); Ji and Elkan (2013), we scale all data to a feature range from 0 to 1. We 6 Breast Banknote IW DPGAN DPCGAN PrivBayes DPGAN DPCGAN PrivBayesWST#None 2:36650:0982 1:58530:1333 2:11170:1740 0:47460:0214 0:74420:0333 0:32370:0162 BetaNoised 1:43370:1114 2:22320:2325 1:23220:0823 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

30

BetaDebiased 1:89220:1237 1:99130:3507 1:18250:0933 0:40150:0766 0:46180:0832 0:23690:0061 DP-MLP 1:45700:1492 1:03150:1415 1:21900:0795 0:20350:0427 0:42980:0433 0:04560:0061 Discriminator 1:00070:0004 1:00010:0001 - 0:33820:0399 0:10870:0415 LogReg 1:64510:1168 2:29530:2121 1:46630:1152 0:25080:0432 0:43480:0460 0:23480:0034

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

31

MLP 1:61290:1404 1:07090:1579 1:41410:1216 0:09130:0259 0:38600:0452 0:00210:0004MSE#None 2:06430:2012 4:98281:5701 2:39040:1050 11:02151:8377 19:32433:7708 8:17240:3987 BetaNoised 2:75320:2650 2:50250:3763 2:11440:2400 8:42981:0383 15:28624:0365 5:70010:1885 BetaDebiased 2:83370:3842 2:23241:0446 1:82660:2392 8:35082:3127 12:99095:9024 6:68620:1458 DP-MLP 2:39650:2083 3:88650:6043 2:31300:2195 17:15972:5448 16:46184:1011 3:55190:2895

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

32

Discriminator 1:45910:1837 4:06120:9523 - 12:54712:3124 10:92825:4283 LogReg 2:69340:2667 2:21560:3366 1:53330:2138 8:47601:0406 15:29644:0396 5:67510:1785 MLP 2:39990:2040 3:83430:7032 1:65810:2020 17:93902:4926 15:52114:2147 2:62860:3761MLP ROC-AUC"None 0:63740:0421 0:67910:0966 0:83660:0579 0:85460:0213 0:68630:0436 0:76300:0495 BetaNoised 0:61100:0477 0:65460:0727 0:70760:0983 0:84950:0274 0:60630:0510 0:89430:0173

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

33

BetaDebiased 0:68200:0510 0:71730:0842 0:85570:0765 0:87290:0310 0:58680:1005 0:76320:0517 DP-MLP 0:79420:0404 0:56860:0823 0:73530:0887 0:76970:0419 0:56570:0570 0:89530:0299 Discriminator 0:69920:0839 0:72900:0720 - 0:86950:0167 0:71140:0424 LogReg 0:66310:0469 0:64840:1081 0:76180:1019 0:81720:0327 0:60340:0534 0:91020:0129 MLP 0:77300:0412 0:73580:1017 0:75730:0738 0:82910:0333 0:59740:0627 0:85940:0231 Table 1: Mean and standard error over 10 runs for ( = 1,=N1

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

34

Table 1: Mean and standard error over 10 runs for ( = 1,=N1 De6) on the Breast and Banknote data. Best score out of the private methods is marked in bold. use a train-test split of 80%. In all experiments we fix to N1 D106, and choose 2f1;6g. We refer to Supplement C.7 for a complete overview of the results. Synthetic data generators We used DPCGAN (Torkzadehmahani et al., 2019), DPGAN (Xie et al., 2018), and their corresponding non-DP analogues (CGAN and CGAN) to generate DP synthetic data of the same size as the training data set. Additionally we also consider PrivBayes (Zhang et al., 2017), a DP Bayesian Network, as a potential SDGP. Hyperparameter tuning Note that hyperparameter tuning is essentially non-private, and has to be accounted for in the privacy budget. Since hyperparemeter tuning in a DP setting is an unresolved problem (Liu and Talwar, 2019; Rosenblatt et al., 2020; Papernot and Steinke, 2021), we follow Jordon et al. (2019) and tune the hyperparameters

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

35

Rosenblatt et al., 2020; Papernot and Steinke, 2021), we follow Jordon et al. (2019) and tune the hyperparameters of the underlying baselines on private validation data sets. However, we propose default parameters for our methods. This leads to an over-optimistic presentation of the baseline performance, and a conservative presentation of our extensions. Evaluation metrics In order to show that IW decreases statistical bias, we train a linear prediction model on the synthetic data and approximate its bias. Since the true DGP is not known, we train the same linear predictor on the test data and report the mean squared error (MSE) between the test parameters and the parameters estimated on the SDGP, asMSE. We further analyse the divergence of theweighted SDGP and the DGP in a similar way by computing the Wassertstein (WST) distance w.r.t the test data.As one exemplary supervised downstream task, we consider the training of a linear downstream classifier or regressor on the synthetic data. This downstream predictor is then assessed by the error measured in the parameter vector compared to the parameters learnt using the test set ( beta MSE). As another downstream task, we train a one-hidden-layer 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

36

by the error measured in the parameter vector compared to the parameters learnt using the test set ( beta MSE). As another downstream task, we train a one-hidden-layer MLP on the training data, and report the test prediction error as MLP ROC-AUC for classification tasks, and MLP MSE for regression tasks. Choice of budget split We only present results for = 1 in this section, and refer the reader to Supplement C.7 for further results with = 6. If the weight computation procedure requires a separate privacy budget (e.g. if the weights are computed by a separate MLP or logistic regression), we spend 10% of the -budget on fitting the SDGP and 30% of the-budget on the weight computation; the complete budget can be spent on fitting the SDGP if no weights, or the weights of the discriminator are used. In Supplement C.3, we evaluate a range of different privacy splits on the Breast and Boston data. Results In Tables 1 and 2, we see that the performance of the models mostly improved when weighted with any type of estimated weights. Although the best inference for each data set is nearly always achieved after importance weighting, we notice that there are some rare cases where no

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

37

type of estimated weights. Although the best inference for each data set is nearly always achieved after importance weighting, we notice that there are some rare cases where no importance weighting performs (insignificantly) better. For 7 IW DPGAN PrivBayesWST#None 2:20130:0945 1:39380:0231 BetaNoised 2:09220:0419 1:30090:0338 BetaDebiased 2:09300:0393 1:27050:0290 DP-MLP 2:05420:0184 1:02650:0035 Discriminator 2:01450:0141 LogReg 2:20510:0819 1:40780:0492 MLP 2:03500:0158 1:00720:0009MSE#None 0:18670:0434 0:00110:0002 BetaNoised 0:17610:0948 0:00880:0028 BetaDebiased 0:06670:0188 0:00770:0022 DP-MLP 0:15300:0812 0:00480:0024

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

38

DP-MLP 0:15300:0812 0:00480:0024 Discriminator 0:15670:1825 LogReg 0:07490:0279 0:00370:0016 MLP 0:14760:0804 0:00080:0002MLP MSE#None 1:88510:5262 0:19730:0108 BetaNoised 1:00570:1973 0:22000:0154 BetaDebiased 0:90240:1244 0:21390:0122 DP-MLP 0:94620:1702 0:18770:0174 Discriminator 1:62560:2394 LogReg 1:06060:2648 0:25150:0305 MLP 1:09790:2225 0:16970:0079 Table 2: Mean and standard error over 10 runs for ( = 1, =N1 De6) on the Boston Housing data. Best score out of the private methods is marked in bold. IW MSE# MLP ROC-AUC "

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

39

out of the private methods is marked in bold. IW MSE# MLP ROC-AUC " None 0:66050:0384 0:85020:0386 BetaNoised 0:62470:0184 0:87660:0086 BetaDebiased 0:62400:0179 0:87830:0093 DP-MLP 0:58130:0246 0:86830:0055 Discriminator 0:62420:0140 0:86310:0310 LogReg 0:62340:0183 0:87700:0092 MLP 0:57070:0207 0:87370:0058 Table 3: Mean and standard error over 10 runs with standard errors for (= 9:64;= 60;0001e6)on MNIST. instance, we observe that the SDGP obtained with PrivBayes seems to be close to the true DGP of the Boston Housing data, and that importance weighting is no longer helpful. In settings where the SDGP and the DGP are really close, it is possible that the effects of additional variance induced by

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

40

data, and that importance weighting is no longer helpful. In settings where the SDGP and the DGP are really close, it is possible that the effects of additional variance induced by estimating and privatising the importance weights (where appropriate) cancels out the reduction in bias. This effect might be mitigated with hyperparameter tuning. Further, we note that debiasing the logistic regression weights mainly results in better performance. Even though we experience a slight drop in performance from BetaNoised to BetaDebiased in some rare cases, this can be explained by randomness in the data set as we show in Supplement Table 6 that the weights estimated by BetaDebiased are significantly closer to the true LogReg weights than the importance weights given by BetaNoised. If a GAN is used as SDGP, and the data curator is hesitant to release additional importance weights,the discriminator weights nearly always lead to an improvement in results without requiring additional computations. To further illustrate the practical meaning of debiasing, we have included an exemplary case study in Supplement C.6. 4.3 BAYESIAN UPDATING WITH IW We investigate the effectiveness of IW in a Bayesian learning setting as per Equation 3. We evaluated and compared the performance of these weighted posteriors alongside the

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

41

We investigate the effectiveness of IW in a Bayesian learning setting as per Equation 3. We evaluated and compared the performance of these weighted posteriors alongside the standard non-weighted posterior by applying them to learning the parameters of models for various regression tasks. Figure 2 shows the ROC-AUC scores associated with the Bayesian predictive distribution arising from integration over the posterior of a Bayesian logistic regression model fit on synthesised versions of the Banknote dataset. We observe that the ROC-AUC under PrivBayes’ synthetic data is significantly improved upon across all IW methods, with similar gains made to the median performance under CGAN’s synthetic data. Additionally, most of the methods help in decreasing variability in the results, especially DP-MLP and MLP. See Supplement C.5 for a full specification of the experimental details and for further results from fitting Bayesian linear regression and multinomial logistic regression models on the TGFB and Iris datasets respectively. 4.4 MNIST Additionally, we assessed how IW performs in a highdimensional setting such as a classification task on the MNIST dataset. Since PrivBayes does not scale to large data sets, we only evaluate DPCGAN as possible SDGP. For

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

42

MNIST dataset. Since PrivBayes does not scale to large data sets, we only evaluate DPCGAN as possible SDGP. For this we follow the setup by Torkzadehmahani et al. (2019) for= 9:64and= 60001106. We observe in Table 3 that all IW methods improve upon the state of the art. 5 DISCUSSION In this paper, we investigated importance weighting methods to correct for biases in downstream estimation tasks when using differentially private synthetic data. While classification algorithms can be used to estimate the required importance weights, noise must be added in order to maintain privacy. We presented methods to debias inference based on privatised weights estimated by logistic regression, developed private estimation procedures allowing the complexity of neural networks to be leveraged for weight estimation, and proposed using inbuilt discriminator weights from GAN synthetic data generation to avoid increases to the privacy budget. Following these developments, we advocate that future releases of DP synthetic data are augmented with privatised importance weights to allow researchers to conduct unbiased 8 downstream model estimation. Future work will focus on improved hyperparameter tuning practises to choose the optimal IW approach for the task and dataset at hand. Acknowledgements

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

43

downstream model estimation. Future work will focus on improved hyperparameter tuning practises to choose the optimal IW approach for the task and dataset at hand. Acknowledgements SG is a student of the EPSRC CDT in Modern Statistics and Statistical Machine Learning (EP/S023151/1) and receives funding from the Oxford Radcliffe Scholarship and Novartis. HW is supported by the Feuer International Scholarship in Artificial Intelligence. JJ was funded by the Ayudas Fundación BBV A a Equipos de Investigación Cientifica 2017 and Government of Spain’s Plan Nacional PGC2018101643-B-I00 grants whilst working on this project. SJV is supported by the University of Warwick, University of Warwick and German Resarch Centre for Aritifical Intelligence. CH is supported by The Alan Turing Institute, Health Data Research UK, the Medical Research Council UK, the EPSRC through the Bayes4Health programme Grant EP/R018561/1, and AI for Science and Government UK Research and Innovation (UKRI). References Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings

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

44

McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages 308–318, 2016. Nazmiye Ceren Abay, Yan Zhou, Murat Kantarcioglu, Bhavani Thuraisingham, and Latanya Sweeney. Privacy preserving synthetic data release using deep learning. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages 510–526. Springer, 2018. Gergely Acs, Luca Melis, Claude Castelluccia, and Emiliano De Cristofaro. Differentially private mixture of generative neural networks. IEEE Transactions on Knowledge and Data Engineering , 31(6):1109–1121, 2018. Marco Avella-Medina. Privacy-preserving parametric inference: a case for robust statistics. Journal of the American Statistical Association , 116(534):969–983, 2021. Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. arXiv preprint arXiv:1810.06758 , 2018. Borja Balle and Yu-Xiang Wang. Improving the gaussian mechanism for differential privacy: Analytical calibration

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

45

Borja Balle and Yu-Xiang Wang. Improving the gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. In International Conference on Machine Learning , pages 394–403. PMLR, 2018. Robert H Berk. Limiting behavior of posterior distributions when the model is incorrect. The Annals of Mathematical Statistics , pages 51–58, 1966.Pier Bissiri, Chris Holmes, and Stephen Walker. A general framework for updating belief distributions. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 2016. Avrim Blum, Cynthia Dwork, Frank McSherry, and Kobbi Nissim. Practical privacy: the sulq framework. In Proceedings of the twenty-fourth ACM SIGMOD-SIGACTSIGART symposium on Principles of database systems , pages 128–138, 2005. Kamalika Chaudhuri, Claire Monteleoni, and Anand D Sarwate. Differentially private empirical risk minimization. Journal of Machine Learning Research , 12(3), 2011. Qingrong Chen, Chong Xiang, Minhui Xue, Bo Li, Nikita Borisov, Dali Kaarfar, and Haojin Zhu. Differentially private data generative models. arXiv preprint

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

46

Borisov, Dali Kaarfar, and Haojin Zhu. Differentially private data generative models. arXiv preprint arXiv:1812.02274 , 2018. Victor Chernozhukov and Han Hong. An MCMC approach to classical estimation. Journal of Econometrics , 115(2): 293–346, 2003. Christos Dimitrakakis, Blaine Nelson, Aikaterini Mitrokotsa, and Benjamin IP Rubinstein. Robust and private bayesian inference. In International Conference on Algorithmic Learning Theory , pages 291–305. Springer, 2014. Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Theory of cryptography conference , pages 265–284. Springer, 2006. Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science , 9(3-4):211–407, 2014. Charles Elkan. Preserving privacy in data mining via importance weighting. In International Workshop on Privacy and Security Issues in Data Mining and Machine Learning, pages 15–21. Springer, 2010. Liyue Fan. A survey of differentially private generative

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

47

and Security Issues in Data Mining and Machine Learning, pages 15–21. Springer, 2010. Liyue Fan. A survey of differentially private generative adversarial networks. In The AAAI Workshop on PrivacyPreserving Artificial Intelligence , 2020. James Foulds, Joseph Geumlek, Max Welling, and Kamalika Chaudhuri. On the theory and practice of privacypreserving bayesian data analysis. arXiv preprint arXiv:1603.07294 , 2016. Lorenzo Frigerio, Anderson Santana de Oliveira, Laurent Gomez, and Patrick Duverger. Differentially private generative adversarial networks for time series, continuous, and discrete open data. In IFIP International Conference on ICT Systems Security and Privacy Protection , pages 151–164. Springer, 2019. 9 Hong Ge, Kai Xu, and Zoubin Ghahramani. Turing: a language for flexible probabilistic inference. In International Conference on Artificial Intelligence and Statistics, AISTATS 2018, 9-11 April 2018, Playa Blanca, Lanzarote, Canary Islands, Spain , pages 1682–1690, 2018. URL http://proceedings.mlr.press/ v84/ge18b.html .

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

48

2018. URL http://proceedings.mlr.press/ v84/ge18b.html . Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661 , 2014. Aditya Grover, Jiaming Song, Ashish Kapoor, Kenneth Tran, Alekh Agarwal, Eric J Horvitz, and Stefano Ermon. Bias correction of learned generative models using likelihoodfree importance weighting. In Advances in Neural Information Processing Systems , pages 11058–11070, 2019. Moritz Hardt and Guy N Rothblum. A multiplicative weights mechanism for privacy-preserving data analysis. In2010 IEEE 51st Annual Symposium on Foundations of Computer Science , pages 61–70. IEEE, 2010. Stephanie Hyland, Cristóbal Esteban, and Gunnar Rätsch. Real-valued (medical) time series generation with recurrent conditional gans. arXiv , 2018. Zhanglong Ji and Charles Elkan. Differential privacy based on importance weighting. Machine Learning , 93(1):163–

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

49

Zhanglong Ji and Charles Elkan. Differential privacy based on importance weighting. Machine Learning , 93(1):163– 183, 2013. Zhanglong Ji, Xiaoqian Jiang, Shuang Wang, Li Xiong, and Lucila Ohno-Machado. Differentially private distributed logistic regression using private and public data. BMC medical genomics , 7(1):1–10, 2014. James Jordon, Jinsung Yoon, and Mihaela van der Schaar. Pate-gan: Generating synthetic data with differential privacy guarantees. In International Conference on Learning Representations , 2019. Shiva Prasad Kasiviswanathan, Homin K Lee, Kobbi Nissim, Sofya Raskhodnikova, and Adam Smith. What can we learn privately? SIAM Journal on Computing , 40(3): 793–826, 2011. BJK Kleijn, AW Van der Vaart, et al. The Bernstein-vonMises theorem under misspecification. Electronic Journal of Statistics , 6:354–381, 2012. Siem Jan Koopman, Neil Shephard, and Drew Creal. Testing the assumptions behind importance sampling. Journal of Econometrics , 149(1):2–11, 2009.

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

50

the assumptions behind importance sampling. Journal of Econometrics , 149(1):2–11, 2009. Tomasz J Kozubowski and Krzysztof Podgórski. LogLaplace distributions. International Mathematical Journal , 3(4):467–495, 2003.Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. InArtificial Intelligence and Statistics , pages 623–631. PMLR, 2017. Moshe Lichman. UCI machine learning repository, 2013. Jingcheng Liu and Kunal Talwar. Private selection from private candidates. In Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing , pages 298–309, 2019. Terrance Liu, Giuseppe Vietri, Thomas Steinke, Jonathan Ullman, and Steven Wu. Leveraging public data for practical private query release. In International Conference on Machine Learning , pages 6968–6977. PMLR, 2021. Simon P Lyddon, Chris Holmes, and Stephen Walker. General Bayesian updating and the loss-likelihood bootstrap. Biometrika , 2018.

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

51

Simon P Lyddon, Chris Holmes, and Stephen Walker. General Bayesian updating and the loss-likelihood bootstrap. Biometrika , 2018. Marcel Neunhoeffer, Zhiwei Steven Wu, and Cynthia Dwork. Private post-GAN boosting. arXiv preprint arXiv:2007.11934 , 2020. Nicolas Papernot and Thomas Steinke. Hyperparameter tuning with renyi differential privacy. arXiv preprint arXiv:2110.03620 , 2021. Lucas Rosenblatt, Xiaoyan Liu, Samira Pouyanfar, Eduardo de Leon, Anuj Desai, and Joshua Allen. Differentially Private Synthetic Data: Applied Evaluations and Enhancements. arXiv , Nov 2020. URL https://arxiv. org/abs/2011.05537v1 . Masashi Sugiyama, Taiji Suzuki, and Takafumi Kanamori. Density Ratio Estimation in Machine Learning . Cambridge University Press, 2012. Reihaneh Torkzadehmahani, Peter Kairouz, and Benedict Paten. Dp-cgan: Differentially private synthetic data and label generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages 0–0, 2019.

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

52

label generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages 0–0, 2019. Ryan Turner, Jane Hung, Eric Frank, Yunus Saatchi, and Jason Yosinski. Metropolis–Hastings generative adversarial networks. In International Conference on Machine Learning , pages 6345–6353. PMLR, 2019. Aad W Van der Vaart. Asymptotic Statistics , volume 3. Cambridge University Press, 2000. Vladimir Vapnik. Principles of risk minimization for learning theory. Advances in neural information processing systems , 4, 1991. Aki Vehtari, Daniel Simpson, Andrew Gelman, Yuling Yao, and Jonah Gabry. Pareto smoothed importance sampling. arXiv preprint arXiv:1507.02646 , 2015. 10 Yu-Xiang Wang, Stephen Fienberg, and Alex Smola. Privacy for free: Posterior sampling and stochastic gradient monte carlo. In International Conference on Machine Learning , pages 2493–2502. PMLR, 2015. Harrison Wilde, Jack Jewson, Sebastian V ollmer, and Chris Holmes. Foundations of Bayesian learning from synthetic data. arXiv preprint arXiv:2011.08299 , 2020.

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

53

Holmes. Foundations of Bayesian learning from synthetic data. arXiv preprint arXiv:2011.08299 , 2020. Liyang Xie, Kaixiang Lin, Shu Wang, Fei Wang, and Jiayu Zhou. Differentially private generative adversarial network. arXiv preprint arXiv:1802.06739 , 2018. Jun Zhang, Graham Cormode, Cecilia M Procopiuc, Divesh Srivastava, and Xiaokui Xiao. PrivBayes: Private data release via Bayesian networks. ACM Transactions on Database Systems (TODS) , 42(4):1–41, 2017. Xinyang Zhang, Shouling Ji, and Ting Wang. Differentially private releasing via deep generative model. arXiv preprint arXiv:1801.01594 , 2018. 11 A ADDITIONAL MATERIAL A.1 UNBIASED IMPORTANCE WEIGHTING BY OUTPUT PERTURBATION A simple approach to ensure DP of an algorithm is to add noise (Dwork et al., 2006) to its output, that is the estimated importance weights of the synthetic data. We establish general results under which such a noise perturbation of an unbiased non-private weights algorithm bw(x)preserves the unbiasedness of IS estimation.

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

54

non-private weights algorithm bw(x)preserves the unbiasedness of IS estimation. Theorem 2. Let2(h)=Ndenote the variance of the IS estimate IN(hjw)defined in Equation (2). Then the IS estimator IN(hjw)using noise perturbed importance weights w(xi) =bw(xi) +i, whereiare i.i.d. and E[exp(i)] = 1 , is unbiased and has variance 2(h)=Nwhere 2(h) =2(h) +Var[exp()]EpG[(bw(x)h(x))2]: (6) We refer the reader to Supplement B.3 for the proof. In the following we will analyse how the noise has to be chosen to ensure DP. Corollary 1. The IS estimator with importance weights defined by logw(xi) =bTxi+i (7) foriLaplace (log(12);)and=2p d ND<1 is(NS;0)-differentially private. It is further unbiased and for <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

55

d ND<1 is(NS;0)-differentially private. It is further unbiased and for <1 2has variance as defined in equation 6: Var[exp()] = exp(2 log(12)) 1 1421 (12)2! : Note that privacy budget is additive. If we want to release NSDP weights, we thus have to scale the noise proportional to NS. Although this approach increases the variance of the estimator, it remains unbiased. A limitation of this approach is that  < 1=2. Alternatively, Blum et al. (2005) show that adding Gaussian noise 0 N(0;2 2S(f)2log2 )to an algorithm fensures (;)-DP for >0. From our analysis it follows that we could adjust Corollary 1 as follows. Corollary 2. The IS estimator with importance weights defined by logw(xi) =bTxi+0 i for0 iN( 2 2; 2)and =s 8d (ND)2log2 

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

56

i for0 iN( 2 2; 2)and =s 8d (ND)2log2  is(NS;)-differentially private with >0and<1. It is further unbiased and has variance as defined in equation 6 with Var[exp(0)] = 2: This result trivially extends to the case of 1with accordingly adjusted noise scales following results from Balle and Wang (2018). Sources of Bias and Variance. This analysis gives us insights on two sources of bias and variance. The first one is the bias and/or variance introduced by privatising the weights. The estimator of Ji and Elkan (2013) is biased but as a result adds noise with a smaller variance, whereas to be unbiased by noising the weights we have to pay a price of increasing the variance, e.g., by adding more noise or by releasing fewer samples. The second source is the bias and variance introduced by estimating the weights through the classifier. The importance weighting procedure is only unbiased when we know exactly how to estimate the true weights. Using a logistic regression to estimate these cannot reasonably be considered as unbiased

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

57

how to estimate the true weights. Using a logistic regression to estimate these cannot reasonably be considered as unbiased for any complicated data. However, using an arbitrarily complex classifier such as a classification neural network could arguably be considered as less biased at estimating the density ratio if it converges, but possibly increases the variance of the estimators due to the increased number of parameters to learn. Please refer to Table 11 in Supplement C.8 for some experimental results. 12 A.2 POST-PROCESSING OF LIKELIHOOD RATIOS The performance of importance weighting can suffer from a heavy right tailed distribution of the likelihood ratio estimates which increases the variance of downstream estimators. A simple remedy is tempering: for a 2[0;1]the weights fbw(xi)gi2f1;:::;NGgare less extreme. Alternatively, Vehtari et al. (2015) propose Pareto smoothed IS (PSIS). This procedure requires to fit a generalised Pareto distribution to the upper tail of the distribution of the simulated importance ratios. Their algorithm does not only post-hoc stabilise IS, but also reports a warning when the estimated shape parameter of the Pareto distribution exceeds a certain

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

58

stabilise IS, but also reports a warning when the estimated shape parameter of the Pareto distribution exceeds a certain threshold. Similarly, Koopman et al. (2009) propose a test to detect whether importance weights have finite variance. In both warnings, there are certain characteristics of the DGP which are not captured by the SDGP and the resulting IS estimates are likely to be unstable. This warning can thus be understood as a general indicator for unsuitable proposal distributions. For large shape parameters the data owner should not release the SDGP. It is also computationally more efficient than comparable distribution divergences such as maximum mean discrepancy or Wasserstein distance. We must also consider that unlike traditional IS where the importance weights are known (at least up to normalisation), here they are being estimated from data, providing further motivation for regularisation. Aside from unstable likelihood ratios, the computed importance weights can suffer from the inability of the classification method to correctly capture the density ratios. To mitigate this problematic, Turner et al. (2019) propose post-calibration of the likelihood ratios in a non-private setting. If we can assume that the data analyst has access to a small dataset of the DGP,

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

59

the likelihood ratios in a non-private setting. If we can assume that the data analyst has access to a small dataset of the DGP, as e.g. in Wilde et al. (2020), we can make use of post-calibration methods, such as beta calibration (Kull et al., 2017). In Table 12 in Supplement C.8, we experimentally extend the results of Vehtari et al. (2015) and Kull et al. (2017) and show that PSIS and -calibration also improve upon the performance of the un-processed importance weights in a DP setting, especially for larger datasets. Note that the post-processing was only applied on the weights from the GAN discriminator to extend the results others have already proven. B PROOFS B.1 PROPOSITION 1: BIAS AND VARIANCE OF ALGORITHM 1 OF JI & ELKAN (2013) Consider Ji and Elkan (2013) Algorithm 1, where under the assumption thatp(y=1) p(y=0)ND NS= 1, the unprivatised importance weights are estimated using logistic regression bw(xi) =p(y= 1jxi) p(y= 0jxi)= exp bTxi ;

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

60

bw(xi) =p(y= 1jxi) p(y= 0jxi)= exp bTxi ; and then the privacy preserving process adds noise to the bcoefficients of this logistic regression =b+with Laplace (2p d=(ND)), a vector of length d, to generate privatised estimates of the importance weights w(xi) = exp Txi = exp bTxi
exp (xi): (8) The following proposition proves that w(xi)is abiased estimate ofbw(xi), the consequences being that if the ‘true’ importance weight really is given by a logistic regression then the procedure of Ji and Elkan (2013) will be biased. Proposition 1. Letwdenote the importance weights computed by noise perturbing the regression coefficients as in Equation (8)(Ji and Elkan, 2013, Algorithm 1). The importance sampling estimator IN(hjw)is biased. Proof. Firstly, we show that w(xi)is not an unbiased estimate of bw(xi) E[w(xi)] =Eh exp bTxi

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

61

E[w(xi)] =Eh exp bTxi
exp (xi)i =E[bw(xi)
exp (xi)] 6=bw(xi): As a consequence, we show that even if the true density ratio can be captured by a logistic regression, i.e. there exists 0 such thatpD(x) pG(x)= exp T 0x
, then the importance sampling estimator IN(hjw) =1 NNX i=1w(xi)h(xi); xipG(
); 13 withw(
)calculated using ‘privatised’ =0+,distributed as above, is a biased estimate of EpD[h(x)]. Indeed, we have Ex1:NpG" 1 NNX i=1w(xi)h(xi)# =Ex1:NpG" 1 NNX i=1exp T 0xi

exp (xi)h(xi)# =1 NNX i=1ExipG[w(xi)
exp (xi)h(xi)] =1 NNX

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

62

=1 NNX i=1ExipG[w(xi)
exp (xi)h(xi)] =1 NNX i=1ExipD[exp (xi)h(xi)] 6=ExipD[h(xi)]: The proof of Proposition 1 provides several insights on what is required for an unbiased estimator. The fact that the bias depends explicitly on the observation suggests either 1) asking the data curator to debias the noise given the synthetic data they are about to release or 2) adding noise to the weights themselves rather to the process of how they are calculated. Ji and Elkan (2013) compute the variance of the estimator =b+whereLaplace (4(d+1)d (ND)2)as Var() =Var(b) +Var() =Var(b) +4(d+ 1)d (ND)2: They show that the asymptotic variance of importance sampling with the unperturbed weights obtained from the logistic regressionwlogreg can be upper bounded by Var(IN(h;wlogreg )) =TVar(b)=TdId ND2 with =P

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

63

Var(IN(h;wlogreg )) =TVar(b)=TdId ND2 with =P xi;xj2DeT 0(xi+xj)(h(xi)h(xj)) (xixj) P xi;xj2EeT 0(xi+xj); where0optimises the loss function of a logistic regression on fixed Gand the true distribution of D. The asymptotic variance of the importance sampling estimator with the weights w logreg from the logistic regression with parameter is then Var(IN(h;w logreg )) =TVar()=T(dId ND2+4(d+ 1)d (ND)2): B.2 PROPOSITION 2: DEBIASING OF JI & ELKAN (2013) As prescribed by Ji and Elkan (2013) Algorithm 1, consider importance weights w(xi) = exp Txi = exp bTxi
exp Txi
: (9)

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

64

w(xi) = exp Txi = exp bTxi
exp Txi
: (9) for privacy preserved bcoefficients of this logistic regression =b+withLaplace (2p d=(ND)), a vector of lengthd. Proposition 1 proved that using w(
)resulted in biased expectation estimation. However, Proposition 2 demonstrates that we can debias this in closed form. Proposition 2. Letwdenote the importance weights computed by noise perturbing the regression coefficients as in Equation (9)(Ji and Elkan, 2013, Algorithm 1) with p. Define b(xi) := 1=Ep[exp Txi
]; and adjusted importance weight w(xi) =w(xi)
b(xi) =bw(xi)
exp Txi

b(xi): 14 The importance sampling estimator IN(hjw)is unbiased and (;0)-differentially private. The variance of estimator IN(hjw)has the following decomposition Varp

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

65

IN(hjw)has the following decomposition Varp G[IN(hjw)] =2(h) N+ 11 N c(h): with 2(h) =2(h) +ExpG h(x)2bw(x)2Varp[b(x) exp(x)] ; 2(h) =VarxpG[h(x)bw(x)]; (10) c(h) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0)1 : Proof. Consider (x1;:::;xN;)i:i:dp G, i.e.xii:i:dpG,i= 1;:::;N andpand IN(hjw) =1 NNX i=1h(xi)bw(xi) exp Txi
b(xi); then Ep

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

66

NNX i=1h(xi)bw(xi) exp Txi
b(xi); then Ep G[IN(hjw)] =ExpG(x)Ep[h(x)bw(x) exp Tx
b(x)] =ExpG(x)[h(x)bw(x)b(x)Ep[exp Tx
]] =ExpG(x)[h(x)bw(x)] =ExpD(x)[h(x)] and as a result IN(hjw)is an unbiased estimator of ExpD(x)[h(x)]. The variance of estimator IN(hjw)is given by Varp G[IN(hjw)] =1 N2NX i=1Varp G[h(xi)w(xi)] +2 N2NX i=1X j<iCovp G[h(xi)w(xi);h(xj)w(xj)] =2(h) N+ 11 N c(h): (11)

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

67

=2(h) N+ 11 N c(h): (11) where the weights are dependent under p Gbecauseis not sampled independently for each xi, it is only sampled once. The terms making up (11) are 2(h) =Varp G[h(x)bw(x) exp(x)b(x)] =Ep Gh (h(x)bw(x) exp(x)b(x))2i Ep G[h(x)bw(x) exp(x)b(x)]2 =ExpG h(x)2bw(x)2Ep b(x)2exp(x)2 EpG[h(x)bw(x)]2 =ExpG h(x)2bw(x)2 Varp[b(x) exp(x)] + 1
 EpG[h(x)bw(x)]2 =2(h) +ExpG h(x)2bw(x)2Varp[b(x) exp(x)]

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

68

=2(h) +ExpG h(x)2bw(x)2Varp[b(x) exp(x)] ; withEp[b(x) exp(x)] = 1 by construction and 2(h)defined in (10), and c(h) =Covp G h(x)bw(x) exp Tx
b(x);h(x0)bw(x0) exp Tx0
b(x0) =Ex;x0pG;p h(x)bw(x) exp Tx
b(x)
h(x0)bw(x0) exp Tx0
b(x0) Ex;p G h(x)bw(x) exp Tx
b(x)
Ex0;p G h(x0)bw(x0) exp Tx0
b(x0) : ByE exp Tx
b(x)

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

69

Tx0
b(x0) : ByE exp Tx
b(x) = 1, andx;x0iidpGthe second term simplifies to Exp G h(x)bw(x) exp Tx
b(x)
Ex0p G h(x0)bw(x0) exp Tx0
b(x0) =ExpG[h(x)bw(x)]2: 15 The first term can be simplified as Ex;x0pG;p h(x)bw(x) exp Tx
b(x)
h(x0)bw(x0) exp Tx0
b(x0) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0)Ep exp T(x+x0)
 =Ex;x0pG

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

70

exp T(x+x0)
 =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0)1 +ExpG[h(x)bw(x)]Ex0pG[h(x0)bw(x0)](indep.) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0)1 +ExpG[h(x)bw(x)]2: As a result c(h) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0)1 B.2.1 Special Case 1: Laplace Noise

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

71

b(x+x0)1 B.2.1 Special Case 1: Laplace Noise Recall thatxiandared-dimensional vectors with d1. For i.i.d.j,j= 1;:::;d E exp Txi
 =E2 4exp0 @dX j=1jxij1 A3 5 =E2 4dY j=1exp (jxij)3 5 =dY j=1E[exp (jxij)];(independence) which is the moment generating function for random variable jevaluated at t=xij. Now forjiidL(;) dY j=1E[exp (jxij)] =dY j=1exp (xij) 12x2 ij;forjxijj<1=8j =exp Pd j=1xij Qd j=1 12x2 ij
;forjxijj<1=8j: as a result b(xi) =Qd j=1 12x2 ij
exp

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

72

as a result b(xi) =Qd j=1 12x2 ij
exp Pd j=1xij;withjxijj<1=8j (12) 16 The variance Of interest to the performance of such an approach are the terms Varp b(xi) exp(Txi) =b(xi)2Varp exp(Txi) =b(xi)2 Ep exp(Txi)2 Ep exp(Txi)2 =b(xi)2 Ep exp(2Txi) Egp exp(Txi)2 =Qd j=1 12x2 ij
2 exp 2Pd j=1xij0 @exp 2Pd j=1xij Qd j=1 14b2x2 ij
exp 2Pd j=1xij Qd j=1 12x2

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

73

14b2x2 ij
exp 2Pd j=1xij Qd j=1 12x2 ij
21 A =dY j=1 12x2 ij
2 14b2x2 ij
1 withjxijj<1=28j, and b(x)b(x0) b(x+x0)1 =Qd j=1(12x2 j) exp(Pd j=1xj)Qd j=1 12x02 j exp(Pd j=1x0 j) Qd j=1(12(xj+x0 j)2) exp(Pd j=1(xj+x0 j))1;withjxjj;jx0 jjandjxj+x0 jj<1=8j =Qd j=1 12x2 j
 12x02 j Qd j=1 12(xj+x0

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

74

12x2 j
 12x02 j Qd j=1 12(xj+x0 j)2
1: B.2.2 Special Case 2: Gaussian Noise Recall thatxiandared-dimensional vectors with d1. The reciprocal of the bias correction 1 b(xi)=E[exp Txi
]; is the moment generating function of random variable Txievaluated at t= 1. Now ifjiidN(;2),j= 1;:::;d , then Txi=dX j=1jxijN(dX j=1xij;2dX j=1x2 ij) and therefore E exp Txi
 = exp0 @dX j=1xij+1 22dX j=1x2 ij1 A: The variance Of interest to the performance of such an approach are the terms Varp b(xi) exp(Txi) =b(xi)2Varp exp(Txi) =b(xi)2

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

75

b(xi) exp(Txi) =b(xi)2Varp exp(Txi) =b(xi)2 Ep exp(2Txi) Ep exp(Txi)2 = exp0 @2dX j=1xij2dX j=1x2 ij1 A0 @exp0 @2dX j=1xij+ 22dX j=1x2 ij1 A exp0 @2dX j=1xij+2dX j=1x2 ij1 A1 A = exp0 @2dX j=1x2 ij1 A1 17 and b(x)b(x0) b(x+x0)1 =exp Pd j=1xj1 22Pd j=1x2 j exp Pd j=1x0 j1 22Pd j=1x02 j exp Pd

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

76

exp Pd j=1x0 j1 22Pd j=1x02 j exp Pd j=1(xj+x0 j)1 22Pd j=1(xj+x0 j)21 = exp0 @1 22dX j=1n (xj+x0 j)2x2 jx02 jo1 A1 = exp0 @2dX j=1xjx0 j1 A1 B.2.3 Differential Privacy The differential privacy of the approach follows from the post-processing theorem: since the synthetic data x1;:::;xNGis already privatised, the corresponding weights w(x1);:::;w(xNG)are(;)differentially private, and the adversary can be assumed to know which differential privacy mechanism is used (Balle and Wang, 2018), the data curator can debias the weights without any additional privacy budget. B.2.4 Variance Comparison of Debiasing Ji & Elkan (2013)

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

77

weights without any additional privacy budget. B.2.4 Variance Comparison of Debiasing Ji & Elkan (2013) Ji and Elkan (2013) provide bounds for the asymptotic variance of their privatised estimator. Here, we investigate the finite sample variance of their (biased) method and compare it with the finite variance of our unbiased estimator form Proposition 2. Note that we do not consider self-normalised IW while this is an implicit assumption made by Ji and Elkan (2013). The variance of estimator IN(hjw), wherewis defined in Equation (9), is given by Varp G[IN(hjw)] =1 N2NX i=1Varp G[h(xi)w(xi)] +2 N2NX i=1X j<iCovp G[h(xi)w(xi);h(xj)w(xj)] =2(h) N+ 11 N c(h): where,x;x0p G. The term2(h)is 2(h) =Varp G h(x)bw(x) exp(Tx)

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

78

G. The term2(h)is 2(h) =Varp G h(x)bw(x) exp(Tx) =Ep Gh h(x)bw(x) exp(Tx)
2i Ep G h(x)bw(x) exp(Tx)2 =ExpG h(x)2bw(x)2Ep exp(Tx)2 ExpG(x)h(x)bw(x) b(x)2 =ExpG h(x)2bw(x)2 Varp exp(Tx) +1 b(x)2 ExpG(x)h(x)bw(x) b(x)2 =ExpG h(x)2bw(x)2Varp exp(Tx) +VarxpG(x)h(x)bw(x) b(x) : Further,c(h)is c(h) =Covp

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

79

b(x) : Further,c(h)is c(h) =Covp G h(x)bw(x) exp Tx
;h(x0)bw(x0) exp Tx0
 =Ex;x0p G h(x)bw(x) exp Tx

h(x0)bw(x0) exp Tx0
 Exp G h(x)bw(x) exp Tx

Ex0p G h(x0)bw(x0) exp Tx0
 ; where firstly, Exp G h(x)bw(x) exp Tx

Ex0p G h(x0)bw(x0) exp Tx0
 =ExpG(x)h(x)bw(x) b(x)2 ; 18 and Ex;x0p G h(x)bw(x) exp Tx

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

80

b(x)2 ; 18 and Ex;x0p G h(x)bw(x) exp Tx

h(x0)bw(x0) exp Tx0
 =Ex;x0pG h(x)bw(x)h(x0)bw(x0)Ep exp T(x+x0)
 =Ex;x0pG h(x)bw(x)h(x0)bw(x0)1 b(x+x0) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)1 b(x+x0)1 b(x)b(x0) +Ex;x0pGh(x)bw(x) b(x)h(x0)bw(x0) b(x0) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)1 b(x+x0)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

81

h(x)bw(x)h(x0)bw(x0)1 b(x+x0)1 b(x)b(x0) +ExpGh(x)bw(x) b(x) Ex0pGh(x0)bw(x0) b(x0) (indep.) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)1 b(x+x0)1 b(x)b(x0) +ExpGh(x)bw(x) b(x)2 as a result c(h) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)1 b(x+x0)1 b(x)b(x0) =Ex;x0pGh(x)bw(x) b(x)h(x0)bw(x0) b(x0)b(x)b(x0) b(x+x0)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

82

b(x0)b(x)b(x0) b(x+x0)1 : Comparisons after debiasing: We can compare the variance of IN(hjw)with the previously evaluated variance of IN(hjw)as follows Varp G[IN(hjw)] =2(h) N+ 11 N c(h): Varp G[IN(hjw)] =2(h) N+ 11 N c(h): with 2(h) =ExpG h(x)2bw(x)2Varp[b(x) exp(x)] +VarxpG(x)[h(x)bw(x)] 2(h) =ExpG h(x)2bw(x)2Varp exp(Tx) +VarxpG(x)h(x)bw(x) b(x) and c(h) =Ex;x0pG

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

83

b(x) and c(h) =Ex;x0pG h(x)bw(x)h(x0)bw(x0)b(x)b(x0) b(x+x0)1 c(h) =Ex;x0pGh(x)bw(x) b(x)h(x0)bw(x0) b(x0)b(x)b(x0) b(x+x0)1 : Comparison for the introduction of Laplace noise: From Equation (12), under jL(0;)we have that b(xi) =pY j=1 12x2 ij
;withjxijj<1=8j: 19 The condition that jxijj<1=ensures that 0 12x2 ij
1;8j )0b(x) =pY j=1 12x2 j
1 As a result, Varg b(x) exp(Tx) Varg

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

84

j
1 As a result, Varg b(x) exp(Tx) Varg exp(Tx) ;8x andh(x)bw(x)h(x)bw(x) b(x);8x which provides that 2(h)2(h) andc(h)c(h) )Varp G[IN(hjw)]Varp G[IN(hjw)]: (13) Not only does debiasing remove bias, it also makes the estimator’s variance smaller. B.3 THEOREM 2: NOISY IMPORTANCE SAMPLING For privacy purposes, we want to be able to noise the importance weights as in logw(x) = logbw(x) +;forgdrawn from a noise distribution (14) but we would like to still preserve the consistency properties of importance sampling estimates. To achieve this, we expand the original target in importance sampling as follows p D(x;) =pD(x) exp()g()

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

85

To achieve this, we expand the original target in importance sampling as follows p D(x;) =pD(x) exp()g() where2Rwill correspond to some additive noise on the log weights, and g()is a probability density on Rsuch that by assumptionZ exp()g()d= 1; So, in particular, this implies thatZ p D(x;)d=pD(x): Now, we can use a proposal density p G(x;) =pG(x)g()targetingp D(x;)and the resulting importance weight is indeed w(x;) =p D(x;) p G(x;)=bw(x) exp(); i.e. the importance weight in this extended space is a noisy version of the original weight bw(x). We thus have EpD[h(x)] =EpG[h(x)bw(x)] =Ep G[h(x)w(x;)] =Ep G[h(x)bw(x) exp()]:

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

86

=Ep G[h(x)w(x;)] =Ep G[h(x)bw(x) exp()]: It follows that for i.i.d. (xi;i)p G, i.e.xipGandig, then IN(hjw) =1 NNX i=1h(xi)bw(xi) exp(i) 20 is an unbiased and consistent estimator of EpD[h(x)]. Its variance is Var[IN(hjw)] =1 NVarp D[h(x)bw(x) exp()] =2(h) N: By the variance decomposition formula, we have 2(h) =Varp D[h(x)bw(x) exp()] =Eg[exp()]2VarpG[h(x)bw(x)] +Varg[exp()]EpG (h(x)bw(x))2 (variance decomposition formula) =2(h) +Varg[exp()]EpG[(h(x)bw(x))2];

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

87

(variance decomposition formula) =2(h) +Varg[exp()]EpG[(h(x)bw(x))2]; asEg[exp()] = 1 by assumption and Var[IN(hjw)] =1 NVarpG[h(x)bw(x)]. The variance of our estimator is inflated as expected by the introduction of noise. B.4 COROLLARY 1 AND 2: DIFFERENTIAL PRIVACY OF LOG-LAPLACE NOISED IMPORTANCE WEIGHTS Following Kozubowski and Podgórski (2003), the (symmetric) log-Laplace distribution is the distribution of random variable xsuch thaty= log(x)has a Laplace density with location parameter and scale. The density of a log-Laplace (;) random variable is fX(xj;) =1 21 xexp 1 jlogxj : Note this is recovered from the asymmetric log-Laplace in Kozubowski and Podgórski (2003) with ==1 . Kozubowski

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

88

: Note this is recovered from the asymmetric log-Laplace in Kozubowski and Podgórski (2003) with ==1 . Kozubowski and Podgórski (2003) further provide forms for the expectation and variance of the log-Laplace distribution as E[X] =exp() 12for<1; (15) Var[X] = exp(2) 1 1421 (12)2! for<1 2: Next we wish to investigate the differential privacy provided by using the Laplace mechanism (Dwork et al., 2006) to noise importance weights. Adding Laplace noise to the log-weights, as in Equation (14), is equivalent to multiplying the importance weights by log-Laplace noise. In order for the importance sampling to remain unbiased, the log-Laplace noise must have expectation 1. From Equation (15) this will be the case for all <1if we set= log 12
. A binary logistic-regression classifier specifies class probabilities bp(y= 1jx;b) =1 1 + exp

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

89

A binary logistic-regression classifier specifies class probabilities bp(y= 1jx;b) =1 1 + exp xb;bp(y= 0jx;b) =exp xb 1 + exp xb: We denote by z1:NGthe private data sampled from the DGP, and by x1:NDthe synthetic data sampled from the SDGP. Let z0 1:NG be the neighboring data set of z1:NG. The importance weights estimated by such a classifier become bw(xijx1:NG;z1:ND) =~p(yi= 1jxi;^(x1:NG;z1:ND)) ~p(yi= 0jxi;^(x1:NG;z1:ND))ND NG =1 1 + exp xi^(x1:NG;z1:ND)1 + exp xi^(x1:NG;z1:ND) exp xi^(x1:NG;z1:ND)ND NG = exp xi^(x1:NG;z1:ND)ND NG;

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

90

NG = exp xi^(x1:NG;z1:ND)ND NG; 21 and as a result logbw(xijx1:NG;z1:ND)logbw(xijx1:NG;z0 1:ND) =xi^(x1:NG;z1:ND) + logND NG xi^(x1:NG;z0 1:ND) + logND NG =xi^(x1:NG;z1:ND)xi^(x1:NG;z0 1:ND) =pX j=1xij ^(x1:NG;z1:ND)j^(x1:NG;z0 1:ND)j jxijdX j=1 ^(x1:NG;z1:ND)j^(x1:NG;z0 1:ND)j 2p d ND

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

91

1:ND)j 2p d ND if the features are minmax scaled using the sensitivity computed by Chaudhuri et al. (2011). B.5 REMARK 1: THE IMPORTANCE-WEIGHTED LIKELIHOOD AND M-ESTIMATION Remark 1. Minimisation of the importance weight adjusted log-likelihood, w(xi) logf(xij), can be viewed as an M-estimator with clear relations to the standard MLE. Remark 1 of the paper points out the the connection between the Minimisation of the importance weight adjusted log-likelihood, `IW(x;) :=w(xi) logf(xij)and the standard maximum likelihood estimator which can be seen through the lens of M-estimation. We exemplify this below. Following Van der Vaart (2000), the M-estimate of parameter  h:= arg max ExpD[h(;x)] is given by ^(n) h:= arg max nX i=1h(;xi): The estimator ^(n) his consistent and is asymptotically normal, i.e. pn ^(n)

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

92

The estimator ^(n) his consistent and is asymptotically normal, i.e. pn ^(n) h h D!N 0;~V( h) where ~V() := E r2 h(;x)
1
Var[rh(;x)]
E r2 h(;x)
1: M-estimators generalises the case of MLE under model misspecification and the variance calculation collapses to the standard inverse Fisher’s information if the likelihood is correctly specified for the DGP. The minimiser of the importance weight adjusted log-likelihood can be considered an M-estimate with the following form ^(n) IW= arg maxf`IW(x;)g= arg maxfw(x) logf(x;)g: As a result, given x1:nPGthe covariance of the asymptotic Gaussian distribution for ^(n) IWsimplifies to, ~VIW( IW) = EpG 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

93

IWsimplifies to, ~VIW( IW) = EpG r2 `IW(x; IW)
1
VarpG[r`IW(x; IW)]
EpG r2 `IW(x; IW)
1 = EpD r2 `0(x; 0)
1
VarpG[r`IW(x; IW)]
EpD r2 `0(x; 0)
1 = EpD r2 `0(x; 0)
1
EpGh (r`IW(x; IW)) (r`IW(x; IW))Ti
EpD r2 `0(x; 0)
1 where VarpG[r`IW(x; IW)] = EpGh (r`IW(x;

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

94

where VarpG[r`IW(x; IW)] = EpGh (r`IW(x; IW)) (r`IW(x; IW))Ti because at the maximiser  IW EpG[r`IW(x; IW)] = 0 22 Further we can write the variance of the minimiser of the importance weight adjusted log-likelihood in terms of the variance of the standard MLE given the same number of observations x1:nPDas follows: ~VIW( IW) ~V0( 0)=EpGh (r`IW(x; IW)) (r`IW(x; IW))Ti EpDh (r`0(x; 0)) (r`0(x; 0))Ti =EpDh w(x) (r`0(x; IW)) (r`0(x; IW))Ti EpDh (r`0(x; 0)) (r`0(x; 0))Ti:

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

95

IW))Ti EpDh (r`0(x; 0)) (r`0(x; 0))Ti: We can then use such notions to produce an idea of the effective sample size of synthetic data. B.5.1 The Effective Sample Size of Synthetic Data When constructing traditional Importance Sampling estimates it is typical to talk about the ‘effective sample’ size of the sample from the proposal density. The effective sample size is the number of independent samples from the true target that gives an unbiased estimator with the same variance as the importance sampling estimator using NGsamples from the proposal density. When using importance weights to adjust the likelihood for Bayesian updating we are not directly seeking to estimate an expectation, but minimize an (expected) loss to produce a parameter estimate. Analogously, in this scenario we define the effective sample size of the synthetic data as the number of samples, N(e) G, from true DGP PDthat would provide an unbiased maximum likelihood estimate (MLE) with the same variance as the Importance-Weighted MLE (IW-MLE), i.e. N(e) G:=n n:Vh ^(NG) IWi=Vh

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

96

N(e) G:=n n:Vh ^(NG) IWi=Vh ^(n) 0io ; where the function Vcorresponds to the asymptotic variance of that estimator, and j
jis a norm summary of the matrix values covariance of the estimator. Given the asymptotic analysis presented above for the importance-weighted likelihood we have that N(e) G=0 @pNG~V ^(n) 0 ~V ^(NG) IW1 A2 (16) where ~V ^(n) 0 ~V ^(NG) IW=EpDh bw(x) (r`0(x; IW)) (r`0(x; IW))Ti EpDh (r`0(x; 0)) (r`0(x; 0))Ti =EpGh

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

97

(r`0(x; 0)) (r`0(x; 0))Ti =EpGh (r`IW(x; IW)) (r`IW(x; IW))Ti EpGh bw(x) (r`0(x; 0)) (r`0(x; 0))Ti: We note that for multidimensional parameter vectors the V’s are covariance matrices and therefore we need to take a scalar summary using the normj
jof these matrices in order to provide an integer effective sample size N(e) G. Faced with a similar problem Lyddon et al. (2018) consider the matrix trace for example. Lastly, given a sample x1:NGPGthe effective sample size can be estimated by using empirical expectations ~V ^(n) 0 ~V ^(NG) IW1 NGPNG i=1 r`IW(xi;^(n) 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

98

NGPNG i=1 r`IW(xi;^(n) IW) r`IW(xi;^(n) IW)T 1 NGPNG i=1bw(xi) r`0(xi;^(n) IW) r`0(xi;^(n) IW)T: B.6 THEOREM 1: ASYMPTOTIC POSTERIOR DISTRIBUTION OF IMPORTANCE WEIGHTED BAYESIAN UPDATING Section 3.1 of the paper considers the importance weighted Bayesian updating as a special case of general Bayesian updating where the loss function is specifically chosen to account for the fact that inference is being done with samples from pGwhile trying to approximate pD. We henceforth write IW(jfxigi2f1;:::;NGg)/() exp NGX i=1bw(xi) logf(xij)! =() exp NGX i=1`IW(xi;)! ; 23

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

99

=() exp NGX i=1`IW(xi;)! ; 23 for`IW(xi;) :=bw(xi) logf(xij)andbw(xi) =pD(xi)=pG(xi). The next theorem shows that such a posterior given observations frompGhas the same asymptotic distribution as the standard Bayes posterior given samples from pDwould have, and therefore we consider this posterior to be asymptotically calibrated. We give here the formal statement of Theorem 1. BelowD! denotes convergence in distribution. Theorem 1. Let the regular conditions in (Chernozhukov and Hong, 2003; Lyddon et al., 2018) hold. Consider ^(N) IW := arg min2PN i=1`IW(xi;),xii.i.d.pGand^(N) 0:= arg min2PN i=1`0(xi;),xii.i.d.pDwhere`0(x;) :=logf(x;). Then both ^(N) 0and^(N) IWare consistent estimates of 

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

100

Then both ^(N) 0and^(N) IWare consistent estimates of  0:= arg min2R `0(x;)dPD(x). Moreover there exists a non-singular matrix J1such that we have under the importance weighted Bayesian posterior IW(jx1:N) p N ^(N) IW D!N 0;J1
; almost surely w.r.t. x1:11while under the standard Bayesian posterior (jx1:N) p N ^(N) 0 D!N 0;J1
; almost surely w.r.t. x1:1. Proof. Firstly, define  IW:= arg min 2Z `IW(x;)dPG(x); JIW() :=Z r2 `IW(x;)dPG(x): Then Chernozhukov and Hong (2003); Lyddon et al. (2018) show that under regularity conditions the following asymptotic result holds p N

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