Generating Artificial Data for Private Deep Learning Aleksei Triastcyn and Boi Faltings Artificial Intelligence Laboratory Ecole Polytechnique Fédérale de Lausanne Lausanne, Switzerland {aleksei.triastcyn, boi.faltings}@epfl.ch Abstract Sensitive Data Real In this paper, we propose generating artificial data that retain Critic statistical properties of real data as the means of providing Fake privacy for the original dataset. We use generative adversarial networks to draw privacy-preserving artificial data samples ML and derive an empirical method to assess the risk of infor- Artificial Data Labels mation disclosure in a differential-privacy-like way. Our ex- Generator periments show that we are able to generate labelled data of Noise high quality and use it to successfully train and validate su- pervised models. Finally, we demonstrate that our approach significantly reduces vulnerability of such models to model Figure 1: Architecture of our solution. Sensitive data is used inversion attacks. to train a GAN to produce a private artificial dataset, which then can be used by any ML model. 1 Introduction Following recent advancements in deep learning, more and more people and companies get interested in putting their each of them. Moreover, most of these methods assume (im- data in use and employ machine learning (ML) to generate plicitly or explicitly) access to public data of similar nature, a wide range of benefits that span financial, social, medi- which may not be possible in areas like medicine. cal, security, and other aspects. At the same time, however, In contrast, we study the task of privacy-preserving data such models are able to capture a fine level of detail in release, which has many immediate advantages. First, any training data, potentially compromising privacy of individ- ML model could be trained on released data without ad- uals whose features sharply differ from others. Recent re- ditional assumptions. Second, data from different sources search (Fredrikson, Jha, and Ristenpart 2015) suggests that could be easily pooled to build stronger models. Third, even without access to internal model parameters it is possi- released data could be traded on data markets1 , where ble to recover (up to a certain degree) individual examples, anonymisation and protection of sensitive information is one e.g. faces, from the training set. of the biggest obstacles. Finally, data publishing would fa- The latter result is especially disturbing knowing that deep cilitate transparency and reproducibility of research studies. learning models are becoming an integral part of our lives, In particular, we are interested in solving two problems. making its way to phones, smart watches, cars, and appli- First, how to preserve high utility of data for ML algorithms ances. And since these models are often trained on cus- while protecting sensitive information in the dataset. Sec- tomers’ data, such training set recovery techniques endanger ond, how to quantify the risk of recovering private informa- privacy even without access to the manufacturer’s servers tion from the published dataset, and thus, the trained model. where these models are being trained. The main idea of our approach is to use generative adver- One direction to tackle this problem is enforcing privacy sarial networks (GANs) (Goodfellow et al. 2014) to create during training (Abadi et al. 2016; Papernot et al. 2016; artificial datasets to be used in place of real data for train- 2018). We will refer to these techniques as model release ing. This method has a number of advantages over the ear- methods. While these approaches perform well in ML tasks lier work (Abadi et al. 2016; Papernot et al. 2016; 2018; and provide strong privacy guarantees, they are often restric- Bindschaedler, Shokri, and Gunter 2017). First of all, our so- tive. First and foremost, releasing a single trained model lution allows releasing entire datasets, thereby possessing all does not provide much flexibility in the future. For in- the benefits of private data release as opposed to model re- stance, it would significantly reduce possibilities for com- lease. Second, it achieves high accuracy without pre-training bining models trained on data from different sources. Eval- uating a variety of such models and picking the best one is 1 https://www.datamakespossible.com/ also complicated by the need of adjusting private training for value-of-data-2018/dawn-of-data-marketplace on similar public data. Third, it is more intuitive and flexible, a model in a distributed manner by communicating sani- e.g. it does not require a complex distributed architecture. tised updates from participants to a central authority. Such To estimate potential privacy risks, we design an ex post a method, however, yields high privacy losses (Abadi et al. analysis framework for generated data. We use KL diver- 2016; Papernot et al. 2016). An alternative technique sug- gence estimation and Chebyshev’s inequality to find statisti- gested by Papernot et al. (2016), also uses disjoint training cal bounds on expected privacy loss for a dataset in question. sets and builds an ensemble of independently trained teacher Our contributions in this paper are the following: models to transfer knowledge to a student model by labelling • we propose a novel, yet simple, approach for private data public data. This result has been extended in (Papernot et al. release, and to the best of our knowledge, this is the first 2018) to achieve state-of-the-art image classification results practical solution for complex real-world data; in a private setting (with single-digit DP bounds). A differ- ent approach is taken by Abadi et al. (2016). They suggest • we introduce a new framework for statistical estimation using differentially private stochastic gradient descent (DP- of potential privacy loss of the released data; SGD) to train deep learning models in a private manner. This • we show that our method achieves learning performance approach achieves high accuracy while maintaining low DP of model release methods and is resilient to model inver- bounds, but may also require pre-training on public data. sion attacks. A more recent line of research focuses on private data release and providing privacy via generating synthetic The rest of the paper is structured as follows. In Section 2, data (Bindschaedler, Shokri, and Gunter 2017; Huang et al. we give an overview of related work. Section 3 contains 2017; Beaulieu-Jones et al. 2017). In this scenario, DP is some preliminaries. In Section 4, we describe our approach hard to guarantee, and thus, such models either relax the and privacy estimation framework, and discuss its limita- DP requirements or remain limited to simple data. In (Bind- tions. Experimental results and implementation details are schaedler, Shokri, and Gunter 2017), authors use a graph- presented in Section 5, and Section 6 concludes the paper. ical probabilistic model to learn an underlying data distri- bution and transform real data points (seeds) into synthetic 2 Related Work data points, which are then filtered by a privacy test based In recent years, as machine learning applications become a on a plausible deniability criterion. This procedure would be commonplace, a body of work on security of these methods rather expensive for complex data, such as images. Huang et grows at a rapid pace. Several important vulnerabilities and al. (2017) introduce the notion of generative adversarial pri- corresponding attacks on ML models have been discovered, vacy and use GANs to obfuscate real data points w.r.t. pre- raising the need of devising suitable defences. Among the defined private attributes, enabling privacy for more realistic attacks that compromise privacy of training data, model in- datasets. Finally, a natural approach to try is training GANs version (Fredrikson, Jha, and Ristenpart 2015) and member- using DP-SGD (Beaulieu-Jones et al. 2017; Xie et al. 2018; ship inference (Shokri et al. 2017) received high attention. Zhang, Ji, and Wang 2018). However, it proved extremely Model inversion (Fredrikson, Jha, and Ristenpart 2015) difficult to stabilise training √ with the necessary amount of is based on observing the output probabilities of the target noise, which scales as m w.r.t. the number of model pa- model for a given class and performing gradient descent on rameters m. It makes these methods inapplicable to more an input reconstruction. Membership inference (Shokri et al. complex datasets without resorting to unrealistic (at least for 2017) assumes an attacker with access to similar data, which some areas) assumptions, like access to public data from the is used to train a ”shadow” model, mimicking the target, and same distribution. an attack model. The latter predicts if a certain example has Similarly, our approach uses GANs, but data is generated already been seen during training based on its output proba- without real seeds or applying noise to gradients. Instead, we bilities. Note that both attacks can be performed in a black- verify experimentally that out-of-the-box GAN samples can box setting, without access to the model internal parameters. be sufficiently different from real data, and expected privacy To protect privacy while still benefiting from the use of loss is empirically bounded by single-digit numbers. statistics and ML, many techniques have been developed over the years, including k-anonymity (Sweeney 2002), l- diversity (Machanavajjhala et al. 2007), t-closeness (Li, 3 Preliminaries Li, and Venkatasubramanian 2007), and differential privacy This section provides necessary definitions and background. (DP) (Dwork 2006). The latter has been recognised as a rig- Let us commence with approximate differential privacy. orous standard and is widely accepted by the research com- munity. Its generic formulation, however, makes it hard to Definition 1. A randomised function (mechanism) M : achieve and to quantify potential privacy loss of the already D → R with domain D and range R satisfies (ε, δ)- trained model. To overcome this, we build upon notions of differential privacy if for any two adjacent inputs d, d0 ∈ D empirical DP (Abowd, Schneider, and Vilhuber 2013) and and for any outcome o ∈ R the following holds: on-average KL privacy (Wang, Lei, and Fienberg 2016). Most of the ML-specific literature in the area concentrates Pr [M(d) = o] ≤ eε Pr [M(d0 ) = o] + δ. (1) on the task of privacy-preserving model release. One take on the problem is to distribute training and use disjoint datasets. Definition 2. Privacy loss of a randomised mechanism M : For example, Shokri and Shmatikov (2015) propose to train D → R for inputs d, d0 ∈ D and outcome o ∈ R takes the following form: that generated samples will not repeat the input. To alle- viate this problem, we propose to enforce differential pri- Pr [M(d) = o] vacy on the output of the discriminator (critic). This is done L(M(d)kM(d0 )) = log . (2) Pr [M(d0 ) = o] by employing the Gaussian noise mechanism (Dwork and Definition 3. The Gaussian noise mechanism achieving Roth 2014) at the second-to-last layer: clipping the L2 norm (ε, δ)-DP, for a function f : D → Rm , is defined as of the input and adding Gaussian noise. To be more spe- cific, activations a(x) of the second-to-last layer become M(d) = f (d) + N (0, σ 2 ), (3) ã(x) = a(x)/ max(ka(x)k2 , 1) + N (0; σ 2 ). We refer to this q version of the critic as DP critic. where σ > C 2 log 1.25 Note that if the chosen GAN loss function was directly δ /ε and C is the L2-sensitivity of f . differentiable w.r.t. generator output, i.e. if critic could be For more details on differential privacy and the Gaussian treated as a black box, this modification would enforce the mechanism, we refer the reader to (Dwork and Roth 2014). same DP guarantees on generator parameters, and conse- In our privacy estimation framework, we also use some quently, all generated samples. Unfortunately, this is not the classical notions from probability and information theory. case for practically all existing versions of GANs, including Definition 4. The Kullback–Leibler (KL) divergence be- WGAN-GP (Gulrajani et al. 2017) used in our experiments. tween two continuous probability distributions P and Q with As our evaluation shows, this modification has a number corresponding densities p, q is given by: of advantages. First, it improves diversity of samples and de- Z +∞ creases similarity with real data. Second, it allows to prolong p(x) training, and hence, obtain higher quality samples. Finally, DKL (P kQ) = p(x) log dx. (4) −∞ q(x) in our experiments, it significantly improves the ability of GANs to generate samples conditionally. Note that KL divergence between the distributions of M(d) and M(d0 ) is nothing but the expectation of the pri- 4.2 Privacy Estimation Framework vacy loss random variable E[L(M(d)kM(d0 )) ]. Finally, Chebyshev’s inequality is used to obtain tail Our framework builds upon ideas of empirical DP bounds. In particular, as we expect the distribution to be (EDP) (Abowd, Schneider, and Vilhuber 2013; Schneider asymmetric, we use the version with semi-variances (Berck and Abowd 2015) and on-average KL privacy (Wang, Lei, and Hihn 1982) to get a sharper bound: and Fienberg 2016). The first can be viewed as a measure of sensitivity on posterior distributions of outcomes (Charest 2 1 σ+ and Hou 2017) (in our case, generated data distributions), Pr(x ≥ E[x] + kσ) ≤ , (5) k σ2 2 while the second relaxes DP notion to the case of an average 2 R +∞ user. where σ+ = E[x] p(x)(x − E[x])2 dx is the upper semi- As we don’t have access to exact posterior distributions, a variance. straightforward EDP procedure in our scenario would be the following: (1) train GAN on the original dataset D; (2) re- 4 Our Approach move a random sample from D; (3) re-train GAN on the up- In this section, we describe our solution, its further improve- dated set; (4) estimate probabilities of all outcomes and the ments, and provide details of the privacy estimation frame- maximum privacy loss value; (5) repeat (1)–(4) sufficiently work. We then discuss limitations of the method. More back- many times to approximate ε, δ. ground on privacy can be found in (Dwork and Roth 2014). If the generative model is simple, this procedure can The main idea of our approach is to use artificial data for be used without modification. Otherwise, for models like learning and publishing instead of real (see Figure 1 for a GANs, it becomes prohibitively expensive due to repetitive general workflow). The intuition behind it is the following. re-training (steps (1)–(3)). Another obstacle is estimating the Since it is possible to recover training examples from ML maximum privacy loss value (step (4)). To overcome these models (Fredrikson, Jha, and Ristenpart 2015), we need to two issues, we propose the following. limit the exposure of real data during training. While this can First, to avoid re-training, we imitate the removal of ex- be achieved by DP training (e.g. DP-SGD), it would have the amples directly on the generated set D. e We define a similar- limitations mentioned earlier. Moreover, certain attacks can ity metric sim(x, y) between two data points x and y that still be successful if DP bounds are loose (Hitaj, Ateniese, reflects important characteristics of data (see Section 5 for and Pérez-Cruz 2017). Removing real data from the train- details). For every randomly selected real example i, we re- ing process altogether would add another layer of protection move k nearest artificial neighbours to simulate absence of and limit the information leakage to artificial samples. What this example in the training set and obtain D e −i . Our intu- remains to show is that artificial data is sufficiently different ition behind this operation is the following. Removing a real from real. example would result in a lower probability density in the corresponding region of space. If this change is picked up 4.1 Differentially Private Critic by a GAN, which we assume is properly trained (e.g. there Despite the fact that the generator does not have access is no mode collapse), the density of this region in the gen- to real data in the training process, one cannot guarantee erated examples space should also decrease. The number of neighbours k is a hyper-parameter. In our experiments, it is Table 1: Accuracy of student models for non-private base- chosen heuristically by computing KL divergence between line, PATE (Papernot et al. 2016), and our method. the real and artificial data distributions and assuming that all the difference comes from one point. Dataset Non-private PATE Our approach Second, we propose to relax the worst-case privacy loss MNIST 99.2% 98.0% 98.3% bound in step (4) by the expected-case bound, in the same SVHN 92.8% 82.7% 87.7% manner as on-average KL privacy. This relaxation allows us to use a high-dimensional KL divergence estimator (Pérez- Cruz 2008) to obtain the expected privacy loss for every pair Table 2: Empirical privacy parameters: expected privacy loss e −i ). There are two major ad- bound µ and probability γ of exceeding it. of adjacent datasets (D e and D vantages of this estimator: it converges almost surely to the true value of KL divergence; and it does not require inter- Dataset Method µ γ mediate density estimates to converge to the true probabil- WGAN-GP 5.80 ity measures. Also since this estimator uses nearest neigh- MNIST WGAN-GP (DP critic) 5.36 bours to approximate KL divergence, our heuristic described WGAN-GP 13.16 above is naturally linked to the estimation method. SVHN 10−5 WGAN-GP (DP critic) 4.92 Finally, after obtaining sufficiently many samples of dif- ferent pairs (D,e D e −i ), we use Chebyshev’s inequality to WGAN-GP 6.27 CelebA bound the probability γ = Pr(E[L(M(D)kM(D0 )) ] ≥ µ) of WGAN-GP (DP critic) 4.15 the expected privacy loss (Dwork and Rothblum 2016) ex- ceeding a predefined threshold µ. To deal with the problem of insufficiently many samples, one could use a sample ver- 5.1 Experimental Setting sion of inequality (Saw, Yang, and Mo 1984) at the cost of We evaluate our method in two major ways. First, we show looser bounds. that not only is it feasible to train ML models purely on gen- erated data, but it is also possible to achieve high learning 4.3 Limitations performance (Section 5.3). Second, we compute empirical bounds on expected privacy loss and evaluate the effective- Our empirical privacy estimator could be improved in a ness of artificial data against model inversion attacks (Sec- number of ways. For instance, providing worst-case privacy tion 5.4). loss bounds would be largely beneficial. Furthermore, sim- Learning performance experiments are set up as follows: ulating the removal of training examples currently depends on heuristics and the chosen similarity metric, which may 1. Train a generative model (teacher) on the original dataset not lead to representative samples and therefore, poor guar- using only the training split. antees. 2. Generate an artificial dataset by the obtained model and We provide bounds on expected privacy loss based on ex use it to train ML models (students). post analysis of the artificial dataset, which is not equiv- 3. Evaluate students on a held-out test set. alent to the traditional formulation of DP and has certain Note that there is no dependency between teacher and stu- limitations (Charest and Hou 2017) (e.g. it only concerns dent models. Moreover, student models are not constrained a given dataset). Nevertheless, it may be useful in the sit- to neural networks and can be implemented as any type of uations where strict privacy guarantees are not required or machine learning algorithm. cannot be achieved by existing methods, or when one wants We choose three commonly used image datasets for our to get a better idea about expected privacy loss rather than experiments: MNIST, SVHN, and CelebA. MNIST is a the highly unlikely worst-case. handwritten digit recognition dataset consisting of 60000 Lastly, all existing limitations of GANs (or generative training examples and 10000 test examples, each example models in general), such as training instability or mode col- is a 28x28 size greyscale image. SVHN is also a digit recog- lapse, will apply to this method. Hence, at the current state nition task, with 73257 images for training and 26032 for of the field, our approach may be difficult to adapt to inputs testing. The examples are coloured 32x32 pixel images of other than image data. Yet, there is still a number of privacy- house numbers from Google Street View. CelebA is a fa- sensitive applications, e.g. medical imaging or facial analy- cial attributes dataset with 202599 images, each of which sis, that could benefit from our technique. And as generative we crop to 128x128 and then downscale to 48x48. methods progress, new uses will be possible. 5.2 Implementation Details 5 Evaluation For our experiments, we use Python and Pytorch frame- work.2 We implement, with some minor modifications, a In this section, we describe the experimental setup and im- Wasserstein GAN with gradient penalty (WGAN-GP) by plementation, and evaluate our method on MNIST (LeCun Gulrajani et al. (2017). More specifically, the critic consists et al. 1998), SVHN (Netzer et al. 2011), and CelebA (Liu et 2 al. 2015) datasets. http://pytorch.org Figure 2: Results of the model inversion attack. Top to bot- tom: real target images, reconstructions from non-private model, our method, and DP model. Figure 3: Privacy-accuracy trade-off curve and correspond- Table 3: Face detection and recognition rates (pairs with dis- ing image reconstructions from a multi-layer perceptron tances below 0.99) for non-private, our method, and DP. trained on artificial MNIST dataset. Non-private Our approach DP presented in Section 4. Specifically, based on recent ideas Detection 63.6% 1.3% 0.0% in image qualitative evaluation, e.g. FID and Inception Recognition 11.0% 0.3% − Score, we compute image features by the Inception V3 net- work (Szegedy et al. 2016) and use inverse distances be- tween features as sim function. We implement the KL diver- of four convolutional layers with SELU (Klambauer et al. gence estimator (Pérez-Cruz 2008) and use k-d trees (Bent- 2017) activations (instead of ReLU) followed by a fully con- ley 1975) for fast nearest neighbour searches. For privacy nected linear layer which outputs a d-dimensional feature evaluation, we implement the model inversion attack. vector (d = 64). For the DP critic, we implement the Gaus- sian noise mechanism (Dwork and Roth 2014) by clipping 5.3 Learning Performance the L2-norm of this feature vector to C = 1 and adding Gaussian noise with σ = 1.5 (we refer to it as DP layer). First, we evaluate the generalisation ability of a student Finally, it is passed through a linear classification layer. The model trained on artificial data. More specifically, we train generator starts with a fully connected linear layer that trans- a student model on generated data and report test classifica- forms noise and labels into a 4096-dimensional feature vec- tion accuracy on a held-out real set. tor which is then passed through a SELU activation and three As noted above, most of the work on privacy-preserving deconvolution layers with SELU activations. The output of ML focuses on model release methods and assumes (explic- the third deconvolution layer is downsampled by max pool- itly or implicitly) access to similar ”public” data in one form ing and normalised with a tanh activation function. or another (Abadi et al. 2016; Papernot et al. 2016; 2018; Similarly to the original paper, we use a classical WGAN Zhang, Ji, and Wang 2018). On the other hand, existing data value function with the gradient penalty that enforces Lips- release solutions struggle with high-dimensional data (Zhu chitz constraint on a critic. We also set the penalty parameter et al. 2017). It limits the choice of methods for comparison. λ = 10 and the number of critic iterations ncritic = 5. Fur- We chose to compare learning performance with the cur- thermore, we modify the architecture to allow for condition- rent state-of-the-art model release technique, PATE by Pa- ing WGAN on class labels. Binarised labels are appended to pernot et al. (2018), which uses a relatively small set of un- the input of the generator and to the linear layer of the critic labelled ”public” data. Since our approach does not require after convolutions. Therefore, the generator can be used to any ”public” data, in order to make the evaluation more ap- create labelled datasets for supervised learning. propriate, we pick the results of PATE corresponding to the Both networks are trained using Adam (Kingma and Ba least number of labelling queries. 2015) with learning rate 10−4 , β1 = 0, β2 = 0.9, and a Table 1 shows test accuracy for the non-private base- batch size of 64. line model (trained on the real training set), PATE, and our The student network is constructed of two convolu- method. We observe that artificial data allows us to achieve tional layers with ReLU activations, batch normalisation and 98.3% accuracy on MNIST and 87.7% accuracy on SVHN, max pooling, followed by two fully connected layers with which is comparable or better than corresponding results of ReLU, and a softmax output layer. Note that this network PATE. These results demonstrate that our approach does not does not achieve state-of-the-art performance on the used compromise learning performance, and may even improve datasets, but we are primarily interested in evaluating the it, while enabling the full flexibility of data release methods. relative performance drop compared to a non-private model. Additionally, we train a simple logistic regression model To estimate privacy loss, we carry out the procedure on artificial MNIST samples, and obtain 91.69% accuracy, (a) Generated (b) Real (a) Generated (b) Real Figure 4: Generated and closest real examples for SVHN. Figure 5: Generated and closest real examples for CelebA. compared to 92.58% on the original data, confirming that rows depict reconstructed images from a non-private model, student models are not restricted to a specific type. a model trained on GAN samples, and DP model, corre- Furthermore, we observe that one could use artificial data spondingly. One can observe a clear information loss in re- for validation and hyper-parameter tuning. In our experi- constructed images going from non-private model, to arti- ments, correlation coefficients between real and artificial ficial data, to DP. The latter is superior in decoupling the validation losses range from 0.7197 to 0.9972 for MNIST model and the training data, and is a preferred choice in and from 0.8047 to 0.9810 for SVHN. the model release setting and/or if public data is accessible for pre-training. The non-private model, albeit trained with 5.4 Privacy Analysis abundant data (∼200K images) reveals facial features, such Using the privacy estimation framework (see Section 4), we as skin and hair colour, expression, etc. Our method, de- fix the probability γ of exceeding the expected privacy loss spite failing to conceal general shapes in training images (i.e. bound µ in all experiments to 10−5 and compute the corre- faces), seems to achieve a trade-off, hiding most of the spe- sponding µ for each dataset and two versions of WGAN-GP cific features. The obtained reconstructions are either very (vanilla and with DP critic). Table 2 summarises our find- noisy (columns 1, 2, 6, 8), much like DP, or converge to ings. It is worth noting, that our µ should not be viewed as some average feature-less faces (columns 4, 5, 7). an empirical estimation of ε of DP, since the former bounds We also analyse real and reconstructed image pairs using expected privacy loss while the latter–maximum. These two OpenFace (Amos et al. 2016) (see Table 3). It confirms our quantities, however, in our experiments turn out to be similar initial findings: in images reconstructed from a non-private to deep learning DP bounds found in recent literature (Abadi model, faces were detected (recognised) 63.6% (11%) of et al. 2016; Papernot et al. 2018). This may be explained by the time, while for our method, detection succeeded only tight concentration of privacy loss random variable (Dwork in 1.3% of cases and recognition rate was 0.3%, well within and Rothblum 2016) or loose estimation. Additionally, DP state-of-the-art error margins. For DP both rates were at 0%. critic helps to bring down µ values in all cases. To evaluate our privacy estimation method, we look at The lack of theoretical privacy guarantees for our method how the privacy loss bound µ correlates with the success neccesitates assessing the strength of provided protection. of the attack. Figure 3 depicts the privacy-accuracy trade-off We perform this evaluation by running the model inversion curve for an MLP (64-32-10) trained on artificial data. In this attack (Fredrikson, Jha, and Ristenpart 2015) on a student setting, we use a stacked denoising autoencoder to compress model. Note that we also experimented with another well- images to 64-dimensional feature vectors and facilitate the known attack on machine learning models, the membership attack performance. Along the curve, we plot examples of inference (Shokri et al. 2017). However, we did not include the model inversion reconstruction at corresponding points. it in the final evaluation, because of the poor attacker’s per- We see that with growing µ, meaning lower privacy, both formance in our setting (nearly random guess accuracy for model accuracy and reconstruction quality increase. given datasets and models even without any protection). Finally, as an additional measure, we perform visual in- In order to run the attack, we train a student model (a sim- spection of generated examples and corresponding nearest ple multi-layer perceptron with two hidden layers of 1000 neighbours in real data. Figures 4 and 5 depict generated and 300 neurons) in three settings: real data, artificial data and the corresponding most similar real images from SVHN generated by GAN (with DP critic), and real data with differ- and CelebA datasets. We observe that, despite general vi- ential privacy (using DP-SGD with a small ε < 1). As facial sual similarity, generated images differ from real examples recognition is a more privacy-sensitive application, and pro- in details, which is normally more important for privacy. For vides a better visualisation of the attack, we picked CelebA SVHN, digits vary either in shape, colour or surroundings. attribute prediction task to run this experiment. A lot of pairs come from different classes. For CelebA, the Figure 2 shows the results of the model inversion attack. pose and lighting may be similar, but such details as gender, The top row presents the real target images. The following skin colour, facial features are usually significantly different. 6 Conclusions Berck, P., and Hihn, J. M. 1982. Using the semivariance to We investigate the problem of private data release for com- estimate safety-first rules. American Journal of Agricultural plex high-dimensional data. In contrast to commonly stud- Economics 64(2):298–300. ied model release setting, this approach enables important Bindschaedler, V.; Shokri, R.; and Gunter, C. A. 2017. Plau- advantages and applications, such as data pooling from mul- sible deniability for privacy-preserving data synthesis. Pro- tiple sources, simpler development process, and data trading. ceedings of the VLDB Endowment 10(5). We employ generative adversarial networks to produce ar- Charest, A.-S., and Hou, Y. 2017. On the meaning and tificial privacy-preserving datasets. The choice of GANs as limits of empirical differential privacy. Journal of Privacy a generative model ensures scalability and makes the tech- and Confidentiality 7(3):3. nique suitable for real-world data with complex structure. Unlike many prior approaches, our method does not as- Dwork, C., and Roth, A. 2014. The algorithmic founda- sume access to similar publicly available data. In our experi- tions of differential privacy. Foundations and Trends R in ments, we show that student models trained on artificial data Theoretical Computer Science 9(3–4):211–407. can achieve high accuracy on MNIST and SVHN datasets. Dwork, C., and Rothblum, G. N. 2016. Concentrated differ- Moreover, models can also be validated on artificial data. ential privacy. arXiv preprint arXiv:1603.01887. We propose a novel technique for estimating privacy of Dwork, C. 2006. Differential privacy. In 33rd Interna- released data by empirical bounds on expected privacy loss. tional Colloquium on Automata, Languages and Program- We compute privacy bounds for samples from WGAN-GP ming, part II (ICALP 2006), volume 4052, 1–12. Venice, on MNIST, SVHN, and CelebA, and demonstrate that ex- Italy: Springer Verlag. pected privacy loss is bounded by single-digit values. To Fredrikson, M.; Jha, S.; and Ristenpart, T. 2015. Model evaluate provided protection, we run a model inversion at- inversion attacks that exploit confidence information and tack and show that training with GAN reduces information basic countermeasures. In Proceedings of the 22nd ACM leakage (e.g. face detection drops from 63.6% to 1.3%) and SIGSAC Conference on Computer and Communications Se- that attack success correlates with estimated privacy bounds. curity, 1322–1333. ACM. Additionally, we introduce a simple modification to the critic: differential privacy layer. Not only does it improve Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; privacy loss bounds and ensures DP guarantees for the critic Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. output, but it also acts as a regulariser, improving stability of 2014. Generative adversarial nets. In Advances in Neural training, and quality and diversity of generated images. Information Processing Systems, 2672–2680. Considering the rising importance of privacy research and Gulrajani, I.; Ahmed, F.; Arjovsky, M.; Dumoulin, V.; and the lack of good solutions for private data publishing, there is Courville, A. C. 2017. Improved training of wasserstein a lot of potential future work. In particular, a major direction gans. In Advances in Neural Information Processing Sys- of advancing current work would be achieving differential tems, 5769–5779. privacy guarantees for generative models while still preserv- Hitaj, B.; Ateniese, G.; and Pérez-Cruz, F. 2017. Deep ing high utility of generated data. A step in another direction models under the gan: information leakage from collabo- would be to improve the privacy estimation framework, e.g. rative deep learning. In Proceedings of the 2017 ACM by bounding maximum privacy loss, or finding a more prin- SIGSAC Conference on Computer and Communications Se- cipled way of sampling from outcome distributions. curity, 603–618. ACM. References Huang, C.; Kairouz, P.; Chen, X.; Sankar, L.; and Rajagopal, R. 2017. Context-aware generative adversarial privacy. En- Abadi, M.; Chu, A.; Goodfellow, I.; McMahan, H. B.; tropy 19(12):656. Mironov, I.; Talwar, K.; and Zhang, L. 2016. Deep learning with differential privacy. In Proceedings of the 2016 ACM Kingma, D., and Ba, J. 2015. Adam: A method for stochas- SIGSAC Conference on Computer and Communications Se- tic optimization. In Proceedings of the 3rd International curity, 308–318. ACM. Conference for Learning Representations. Abowd, J. M.; Schneider, M. J.; and Vilhuber, L. 2013. Dif- Klambauer, G.; Unterthiner, T.; Mayr, A.; and Hochreiter, ferential privacy applications to bayesian and linear mixed S. 2017. Self-normalizing neural networks. In Advances in model estimation. Journal of Privacy and Confidentiality Neural Information Processing Systems, 972–981. 5(1):4. LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Amos, B.; Ludwiczuk, B.; Satyanarayanan, M.; et al. 2016. Gradient-based learning applied to document recognition. Openface: A general-purpose face recognition library with Proceedings of the IEEE 86(11):2278–2324. mobile applications. Li, N.; Li, T.; and Venkatasubramanian, S. 2007. t- Beaulieu-Jones, B. K.; Wu, Z. S.; Williams, C.; and Greene, closeness: Privacy beyond k-anonymity and l-diversity. In C. S. 2017. Privacy-preserving generative deep neural net- Data Engineering, 2007. ICDE 2007. IEEE 23rd Interna- works support clinical data sharing. bioRxiv 159756. tional Conference on, 106–115. IEEE. Bentley, J. L. 1975. Multidimensional binary search trees Liu, Z.; Luo, P.; Wang, X.; and Tang, X. 2015. Deep learning used for associative searching. Communications of the ACM face attributes in the wild. In Proceedings of International 18(9):509–517. Conference on Computer Vision (ICCV). Machanavajjhala, A.; Kifer, D.; Gehrke, J.; and Venkita- subramaniam, M. 2007. l-diversity: Privacy beyond k- anonymity. ACM Transactions on Knowledge Discovery from Data (TKDD) 1(1):3. Netzer, Y.; Wang, T.; Coates, A.; Bissacco, A.; Wu, B.; and Ng, A. Y. 2011. Reading digits in natural images with unsu- pervised feature learning. In NIPS workshop on deep learn- ing and unsupervised feature learning, volume 2011, 5. Papernot, N.; Abadi, M.; Erlingsson, Ú.; Goodfellow, I.; and Talwar, K. 2016. Semi-supervised knowledge transfer for deep learning from private training data. arXiv preprint arXiv:1610.05755. Papernot, N.; Song, S.; Mironov, I.; Raghunathan, A.; Tal- war, K.; and Erlingsson, Ú. 2018. Scalable private learning with pate. arXiv preprint arXiv:1802.08908. Pérez-Cruz, F. 2008. Kullback-leibler divergence estimation of continuous distributions. In Information Theory, 2008. ISIT 2008. IEEE International Symposium on, 1666–1670. IEEE. Saw, J. G.; Yang, M. C.; and Mo, T. C. 1984. Chebyshev inequality with estimated mean and variance. The American Statistician 38(2):130–132. Schneider, M. J., and Abowd, J. M. 2015. A new method for protecting interrelated time series with bayesian prior distri- butions and synthetic data. Journal of the Royal Statistical Society: Series A (Statistics in Society) 178(4):963–975. Shokri, R., and Shmatikov, V. 2015. Privacy-preserving deep learning. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 1310–1321. ACM. Shokri, R.; Stronati, M.; Song, C.; and Shmatikov, V. 2017. Membership inference attacks against machine learning models. In Security and Privacy (SP), 2017 IEEE Sympo- sium on, 3–18. IEEE. Sweeney, L. 2002. K-anonymity: A model for protecting privacy. Int. J. Uncertain. Fuzziness Knowl.-Based Syst. 10(5):557–570. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; and Wojna, Z. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2818–2826. Wang, Y.-X.; Lei, J.; and Fienberg, S. E. 2016. On-average kl-privacy and its equivalence to generalization for max- entropy mechanisms. In International Conference on Pri- vacy in Statistical Databases, 121–134. Springer. Xie, L.; Lin, K.; Wang, S.; Wang, F.; and Zhou, J. 2018. Differentially private generative adversarial network. arXiv preprint arXiv:1802.06739. Zhang, X.; Ji, S.; and Wang, T. 2018. Differentially pri- vate releasing via deep generative model. arXiv preprint arXiv:1801.01594. Zhu, T.; Li, G.; Zhou, W.; and Philip, S. Y. 2017. Dif- ferentially private data publishing and analysis: a survey. IEEE Transactions on Knowledge and Data Engineering 29(8):1619–1638.