Robin Hood and Matthew Effects -- Differential Privacy Has Disparate Impact on Synthetic Data

Generative models trained using Differential Privacy (DP) are increasingly used to produce and share synthetic data in a privacy-friendly manner. In this paper, we set out to analyze the impact of DP on these models vis-a-vis underrepresented classes and subgroups of data. We do so from two angles: 1) the size of classes and subgroups in the synthetic data, and 2) classification accuracy on them. We also evaluate the effect of various levels of imbalance and privacy budgets. Our experiments, conducted using three state-of-the-art DP models (PrivBayes, DP-WGAN, and PATE-GAN), show that DP results in opposite size distributions in the generated synthetic data. More precisely, it affects the gap between the majority and minority classes and subgroups, either reducing it (a "Robin Hood" effect) or increasing it ("Matthew" effect). However, both of these size shifts lead to similar disparate impacts on a classifier's accuracy, affecting disproportionately more the underrepresented subparts of the data. As a result, we call for caution when analyzing or training a model on synthetic data, or risk treating different subpopulations unevenly, which might also lead to unreliable conclusions.

Measuring Utility and Privacy of Synthetic Genomic Data

Genomic data provides researchers with an invaluable source of information to advance progress in biomedical research, personalized medicine, and drug development. At the same time, however, this data is extremely sensitive, which makes data sharing, and consequently availability, problematic if not outright impossible. As a result, organizations have begun to experiment with sharing synthetic data, which should mirror the real data's salient characteristics, without exposing it. In this paper, we provide the first evaluation of the utility and the privacy protection of five state-of-the-art models for generating synthetic genomic data. First, we assess the performance of the synthetic data on a number of common tasks, such as allele and population statistics as well as linkage disequilibrium and principal component analysis. Then, we study the susceptibility of the data to membership inference attacks, i.e., inferring whether a target record was part of the data used to train the model producing the synthetic dataset. Overall, there is no single approach for generating synthetic genomic data that performs well across the board. We show how the size and the nature of the training dataset matter, especially in the case of generative models. While some combinations of datasets and models produce synthetic data with distributions close to the real data, there often are target data points that are vulnerable to membership inference. Our measurement framework can be used by practitioners to assess the risks of deploying synthetic genomic data in the wild, and will serve as a benchmark tool for researchers and practitioners in the future.

ML-Doctor: Holistic Risk Assessment of Inference Attacks Against Machine Learning Models

Inference attacks against Machine Learning (ML) models allow adversaries to learn information about training data, model parameters, etc. While researchers have studied these attacks thoroughly, they have done so in isolation. We lack a comprehensive picture of the risks caused by the attacks, such as the different scenarios they can be applied to, the common factors that influence their performance, the relationship among them, or the effectiveness of defense techniques. In this paper, we fill this gap by presenting a first-of-its-kind holistic risk assessment of different inference attacks against machine learning models. We concentrate on four attacks - namely, membership inference, model inversion, attribute inference, and model stealing - and establish a threat model taxonomy. Our extensive experimental evaluation conducted over five model architectures and four datasets shows that the complexity of the training dataset plays an important role with respect to the attack's performance, while the effectiveness of model stealing and membership inference attacks are negatively correlated. We also show that defenses like DP-SGD and Knowledge Distillation can only hope to mitigate some of the inference attacks. Our analysis relies on a modular re-usable software, ML-Doctor, which enables ML model owners to assess the risks of deploying their models, and equally serves as a benchmark tool for researchers and practitioners.

Toward Robustness and Privacy in Federated Learning: Experimenting with Local and Central Differential Privacy

Federated Learning (FL) allows multiple participants to collaboratively train machine learning models by keeping their datasets local and exchanging model updates. Recent work has highlighted weaknesses related to robustness and privacy in FL, including backdoor, membership and property inference attacks. In this paper, we investigate whether and how Differential Privacy (DP) can be used to defend against attacks targeting both robustness and privacy in FL. To this end, we present a first-of-its-kind experimental evaluation of Local and Central Differential Privacy (LDP/CDP), assessing their feasibility and effectiveness. We show that both LDP and CDP do defend against backdoor attacks, with varying levels of protection and utility, and overall more effectively than non-DP defenses. They also mitigate white-box membership inference attacks, which our work is the first to show. Neither, however, defend against property inference attacks, prompting the need for further research in this space. Overall, our work also provides a re-usable measurement framework to quantify the trade-offs between robustness/privacy and utility in differentially private FL.

An Overview of Privacy in Machine Learning

Over the past few years, providers such as Google, Microsoft, and Amazon have started to provide customers with access to software interfaces allowing them to easily embed machine learning tasks into their applications. Overall, organizations can now use Machine Learning as a Service (MLaaS) engines to outsource complex tasks, e.g., training classifiers, performing predictions, clustering, etc. They can also let others query models trained on their data. Naturally, this approach can also be used (and is often advocated) in other contexts, including government collaborations, citizen science projects, and business-to-business partnerships. However, if malicious users were able to recover data used to train these models, the resulting information leakage would create serious issues. Likewise, if the inner parameters of the model are considered proprietary information, then access to the model should not allow an adversary to learn such parameters. In this document, we set to review privacy challenges in this space, providing a systematic review of the relevant research literature, also exploring possible countermeasures. More specifically, we provide ample background information on relevant concepts around machine learning and privacy. Then, we discuss possible adversarial models and settings, cover a wide range of attacks that relate to private and/or sensitive information leakage, and review recent results attempting to defend against such attacks. Finally, we conclude with a list of open problems that require more work, including the need for better evaluations, more targeted defenses, and the study of the relation to policy and data protection efforts.

Exploiting Unintended Feature Leakage in Collaborative Learning

Collaborative machine learning and related techniques such as federated learning allow multiple participants, each with his own training dataset, to build a joint model by training locally and periodically exchanging model updates. We demonstrate that these updates leak unintended information about participants' training data and develop passive and active inference attacks to exploit this leakage. First, we show that an adversarial participant can infer the presence of exact data points -- for example, specific locations -- in others' training data (i.e., membership inference). Then, we show how this adversary can infer properties that hold only for a subset of the training data and are independent of the properties that the joint model aims to capture. For example, he can infer when a specific person first appears in the photos used to train a binary gender classifier. We evaluate our attacks on a variety of tasks, datasets, and learning configurations, analyze their limitations, and discuss possible defenses.

Building and Measuring Privacy-Preserving Predictive Blacklists

(Withdrawn) Collaborative security initiatives are increasingly often advocated to improve timeliness and effectiveness of threat mitigation. Among these, collaborative predictive blacklisting (CPB) aims to forecast attack sources based on alerts contributed by multiple organizations that might be targeted in similar ways. Alas, CPB proposals thus far have only focused on improving hit counts, but overlooked the impact of collaboration on false positives and false negatives. Moreover, sharing threat intelligence often prompts important privacy, confidentiality, and liability issues. In this paper, we first provide a comprehensive measurement analysis of two state-of-the-art CPB systems: one that uses a trusted central party to collect alerts [Soldo et al., Infocom'10] and a peer-to-peer one relying on controlled data sharing [Freudiger et al., DIMVA'15], studying the impact of collaboration on both correct and incorrect predictions. Then, we present a novel privacy-friendly approach that significantly improves over previous work, achieving a better balance of true and false positive rates, while minimizing information disclosure. Finally, we present an extension that allows our system to scale to very large numbers of organizations.

LOGAN: Membership Inference Attacks Against Generative Models

Generative models estimate the underlying distribution of a dataset to generate realistic samples according to that distribution. In this paper, we present the first membership inference attacks against generative models: given a data point, the adversary determines whether or not it was used to train the model. Our attacks leverage Generative Adversarial Networks (GANs), which combine a discriminative and a generative model, to detect overfitting and recognize inputs that were part of training datasets, using the discriminator's capacity to learn statistical differences in distributions. We present attacks based on both white-box and black-box access to the target model, against several state-of-the-art generative models, over datasets of complex representations of faces (LFW), objects (CIFAR-10), and medical images (Diabetic Retinopathy). We also discuss the sensitivity of the attacks to different training parameters, and their robustness against mitigation strategies, finding that defenses are either ineffective or lead to significantly worse performances of the generative models in terms of training stability and/or sample quality.

Differentially Private Mixture of Generative Neural Networks

Generative models are used in a wide range of applications building on large amounts of contextually rich information. Due to possible privacy violations of the individuals whose data is used to train these models, however, publishing or sharing generative models is not always viable. In this paper, we present a novel technique for privately releasing generative models and entire high-dimensional datasets produced by these models. We model the generator distribution of the training data with a mixture of $k$ generative neural networks. These are trained together and collectively learn the generator distribution of a dataset. Data is divided into $k$ clusters, using a novel differentially private kernel $k$-means, then each cluster is given to separate generative neural networks, such as Restricted Boltzmann Machines or Variational Autoencoders, which are trained only on their own cluster using differentially private gradient descent. We evaluate our approach using the MNIST dataset, as well as call detail records and transit datasets, showing that it produces realistic synthetic samples, which can also be used to accurately compute arbitrary number of counting queries.

MaMaDroid: Detecting Android Malware by Building Markov Chains of Behavioral Models (Extended Version)

As Android becomes increasingly popular, so does malware targeting it, this motivating the research community to propose many different detection techniques. However, the constant evolution of the Android ecosystem, and of malware itself, makes it hard to design robust tools that can operate for long periods of time without the need for modifications or costly re-training. Aiming to address this issue, we set to detect malware from a behavioral point of view, modeled as the sequence of abstracted API calls. We introduce MaMaDroid, a static-analysis based system that abstracts app's API calls to their class, package, or family, and builds a model from their sequences obtained from the call graph of an app as Markov chains. This ensures that the model is more resilient to API changes and the features set is of manageable size. We evaluate MaMaDroid using a dataset of 8.5K benign and 35.5K malicious apps collected over a period of six years, showing that it effectively detects malware (with up to 0.99 F-measure) and keeps its detection capabilities for long periods of time (up to 0.87 F-measure two years after training). We also show that MaMaDroid remarkably improves over DroidAPIMiner, a state-of-the-art detection system that relies on the frequency of (raw) API calls. Aiming to assess whether MaMaDroid's effectiveness mainly stems from the API abstraction or from the sequencing modeling, we also evaluate a variant of it that uses frequency (instead of sequences), of abstracted API calls. We find that it is not as accurate, failing to capture maliciousness when trained on malware samples including API calls that are equally or more frequently used by benign apps.