Search

Given a group size m and a sensitive dataset D, group privacy (GP) releases information about D with the guarantee that the adversary cannot infer with high confidence whether the underlying data is D or a neighboring dataset D' that differs from D by m records. GP generalizes the well-established notion of differential privacy (DP) for protecting individuals' privacy; in particular, when m=1, GP reduces to DP. Compared to DP, GP is capable of protecting the sensitive aggregate information of a group of up to m individuals, e.g., the average annual income among members of a yacht club. Despite its longstanding presence in the research literature and its promising applications, GP is often treated as an afterthought, with most approaches first developing a DP mechanism and then using a generic conversion to adapt it for GP, treating the DP solution as a black box. As we point out in the paper, this methodology is suboptimal when the underlying DP solution involves subsampling, e.g., in the classic DP-SGD method for training deep learning models. In this case, the DP-to-GP conversion is overly pessimistic in its analysis, leading to low utility in the published results under GP. Motivated by this, we propose a novel analysis framework that provides tight privacy accounting for subsampled GP mechanisms. Instead of converting a black-box DP mechanism to GP, our solution carefully analyzes and utilizes the inherent randomness in subsampled mechanisms, leading to a substantially improved bound on the privacy loss with respect to GP. The proposed solution applies to a wide variety of foundational mechanisms with subsampling. Extensive experiments with real datasets demonstrate that compared to the baseline convert-from-blackbox-DP approach, our GP mechanisms achieve noise reductions of over an order of magnitude in several practical settings, including deep neural network training., Comment: accepted for publication in Proceedings of VLDB Endowment (PVLDB) 2025

As a solution concept in cooperative game theory, Shapley value is highly recognized in model interpretability studies and widely adopted by the leading Machine Learning as a Service (MLaaS) providers, such as Google, Microsoft, and IBM. However, as the Shapley value-based model interpretability methods have been thoroughly studied, few researchers consider the privacy risks incurred by Shapley values, despite that interpretability and privacy are two foundations of machine learning (ML) models. In this paper, we investigate the privacy risks of Shapley value-based model interpretability methods using feature inference attacks: reconstructing the private model inputs based on their Shapley value explanations. Specifically, we present two adversaries. The first adversary can reconstruct the private inputs by training an attack model based on an auxiliary dataset and black-box access to the model interpretability services. The second adversary, even without any background knowledge, can successfully reconstruct most of the private features by exploiting the local linear correlations between the model inputs and outputs. We perform the proposed attacks on the leading MLaaS platforms, i.e., Google Cloud, Microsoft Azure, and IBM aix360. The experimental results demonstrate the vulnerability of the state-of-the-art Shapley value-based model interpretability methods used in the leading MLaaS platforms and highlight the significance and necessity of designing privacy-preserving model interpretability methods in future studies. To our best knowledge, this is also the first work that investigates the privacy risks of Shapley values., Comment: This work was published in CCS 2022

Graph Convolutional Networks (GCNs) are a popular machine learning model with a wide range of applications in graph analytics, including healthcare, transportation, and finance. Similar to other neural networks, a GCN may memorize parts of the training data through its model weights. Thus, when the underlying graph data contains sensitive information such as interpersonal relationships, a GCN trained without privacy-protection measures could be exploited to extract private data, leading to potential violations of privacy regulations such as GDPR. To defend against such attacks, a promising approach is to train the GCN with differential privacy (DP), which is a rigorous framework that provides strong privacy protection by injecting random noise into the trained model weights. However, training a large graph neural network under DP is a highly challenging task. Existing solutions either introduce random perturbations in the graph topology, which leads to severe distortions of the network's message passing, or inject randomness into each neighborhood aggregation operation, which leads to a high noise scale when the GCN performs multiple levels of aggregations. Motivated by this, we propose GCON, a novel and effective solution for training GCNs with edge differential privacy. The main idea is to (i) convert the GCN training process into a convex optimization problem, and then (ii) apply the classic idea of perturbing the objective function to satisfy DP. Extensive experiments using multiple benchmark datasets demonstrate GCON's consistent and superior performance over existing solutions in a wide variety of settings.

Local differential privacy (LDP) is a strong privacy standard that has been adopted by popular software systems. The main idea is that each individual perturbs their own data locally, and only submits the resulting noisy version to a data aggregator. Although much effort has been devoted to computing various types of aggregates and building machine learning applications under LDP, research on fundamental perturbation mechanisms has not achieved significant improvement in recent years. Towards a more refined result utility, existing works mainly focus on improving the worst-case guarantee. However, this approach does not necessarily promise a better average performance given the fact that the data in practice obey a certain distribution, which is not known beforehand. In this paper, we propose the advanced adaptive additive (AAA) mechanism, which is a distribution-aware approach that addresses the average utility and tackles the classical mean estimation problem. AAA is carried out in a two-step approach: first, as the global data distribution is not available beforehand, the data aggregator selects a random subset of individuals to compute a (noisy) quantized data descriptor; then, the data aggregator collects data from the remaining individuals, which are perturbed in a distribution-aware fashion. The perturbation involved in the latter step is obtained by solving an optimization problem, which is formulated with the data descriptor obtained in the former step and the desired properties of task-determined utilities. We provide rigorous privacy proofs, utility analyses, and extensive experiments comparing AAA with state-of-the-art mechanisms. The evaluation results demonstrate that the AAA mechanism consistently outperforms existing solutions with a clear margin in terms of result utility, on a wide range of privacy constraints and real-world and synthetic datasets.

Diffusion models have recently gained significant attention in both academia and industry due to their impressive generative performance in terms of both sampling quality and distribution coverage. Accordingly, proposals are made for sharing pre-trained diffusion models across different organizations, as a way of improving data utilization while enhancing privacy protection by avoiding sharing private data directly. However, the potential risks associated with such an approach have not been comprehensively examined. In this paper, we take an adversarial perspective to investigate the potential privacy and fairness risks associated with the sharing of diffusion models. Specifically, we investigate the circumstances in which one party (the sharer) trains a diffusion model using private data and provides another party (the receiver) black-box access to the pre-trained model for downstream tasks. We demonstrate that the sharer can execute fairness poisoning attacks to undermine the receiver's downstream models by manipulating the training data distribution of the diffusion model. Meanwhile, the receiver can perform property inference attacks to reveal the distribution of sensitive features in the sharer's dataset. Our experiments conducted on real-world datasets demonstrate remarkable attack performance on different types of diffusion models, which highlights the critical importance of robust data auditing and privacy protection protocols in pertinent applications.

Split Learning (SL) has emerged as a practical and efficient alternative to traditional federated learning. While previous attempts to attack SL have often relied on overly strong assumptions or targeted easily exploitable models, we seek to develop more practical attacks. We introduce SDAR, a novel attack framework against SL with an honest-but-curious server. SDAR leverages auxiliary data and adversarial regularization to learn a decodable simulator of the client's private model, which can effectively infer the client's private features under the vanilla SL, and both features and labels under the U-shaped SL. We perform extensive experiments in both configurations to validate the effectiveness of our proposed attacks. Notably, in challenging but practical scenarios where existing passive attacks struggle to reconstruct the client's private data effectively, SDAR consistently achieves attack performance comparable to active attacks. On CIFAR-10, at the deep split level of 7, SDAR achieves private feature reconstruction with less than 0.025 mean squared error in both the vanilla and the U-shaped SL, and attains a label inference accuracy of over 98% in the U-shaped setting, while existing attacks fail to produce non-trivial results., Comment: 19 pages, 20 figures

Graph neural networks (GNNs) have gained an increasing amount of popularity due to their superior capability in learning node embeddings for various graph inference tasks, but training them can raise privacy concerns. To address this, we propose using link local differential privacy over decentralized nodes, enabling collaboration with an untrusted server to train GNNs without revealing the existence of any link. Our approach spends the privacy budget separately on links and degrees of the graph for the server to better denoise the graph topology using Bayesian estimation, alleviating the negative impact of LDP on the accuracy of the trained GNNs. We bound the mean absolute error of the inferred link probabilities against the ground truth graph topology. We then propose two variants of our LDP mechanism complementing each other in different privacy settings, one of which estimates fewer links under lower privacy budgets to avoid false positive link estimates when the uncertainty is high, while the other utilizes more information and performs better given relatively higher privacy budgets. Furthermore, we propose a hybrid variant that combines both strategies and is able to perform better across different privacy budgets. Extensive experiments show that our approach outperforms existing methods in terms of accuracy under varying privacy budgets., Comment: 17 pages, accepted by ACM CCS 2023 as a conference paper

Deep neural networks have strong capabilities of memorizing the underlying training data, which can be a serious privacy concern. An effective solution to this problem is to train models with differential privacy, which provides rigorous privacy guarantees by injecting random noise to the gradients. This paper focuses on the scenario where sensitive data are distributed among multiple participants, who jointly train a model through federated learning (FL), using both secure multiparty computation (MPC) to ensure the confidentiality of each gradient update, and differential privacy to avoid data leakage in the resulting model. A major challenge in this setting is that common mechanisms for enforcing DP in deep learning, which inject real-valued noise, are fundamentally incompatible with MPC, which exchanges finite-field integers among the participants. Consequently, most existing DP mechanisms require rather high noise levels, leading to poor model utility. Motivated by this, we propose Skellam mixture mechanism (SMM), an approach to enforce DP on models built via FL. Compared to existing methods, SMM eliminates the assumption that the input gradients must be integer-valued, and, thus, reduces the amount of noise injected to preserve DP. Further, SMM allows tight privacy accounting due to the nice composition and sub-sampling properties of the Skellam distribution, which are key to accurate deep learning with DP. The theoretical analysis of SMM is highly non-trivial, especially considering (i) the complicated math of differentially private deep learning in general and (ii) the fact that the mixture of two Skellam distributions is rather complex, and to our knowledge, has not been studied in the DP literature. Extensive experiments on various practical settings demonstrate that SMM consistently and significantly outperforms existing solutions in terms of the utility of the resulting model.

Nowadays, differential privacy (DP) has become a well-accepted standard for privacy protection, and deep neural networks (DNN) have been immensely successful in machine learning. The combination of these two techniques, i.e., deep learning with differential privacy, promises the privacy-preserving release of high-utility models trained with sensitive data such as medical records. A classic mechanism for this purpose is DP-SGD, which is a differentially private version of the stochastic gradient descent (SGD) optimizer commonly used for DNN training. Subsequent approaches have improved various aspects of the model training process, including noise decay schedule, model architecture, feature engineering, and hyperparameter tuning. However, the core mechanism for enforcing DP in the SGD optimizer remains unchanged ever since the original DP-SGD algorithm, which has increasingly become a fundamental barrier limiting the performance of DP-compliant machine learning solutions. Motivated by this, we propose DPIS, a novel mechanism for differentially private SGD training that can be used as a drop-in replacement of the core optimizer of DP-SGD, with consistent and significant accuracy gains over the latter. The main idea is to employ importance sampling (IS) in each SGD iteration for mini-batch selection, which reduces both sampling variance and the amount of random noise injected to the gradients that is required to satisfy DP. Integrating IS into the complex mathematical machinery of DP-SGD is highly non-trivial. DPIS addresses the challenge through novel mechanism designs, fine-grained privacy analysis, efficiency enhancements, and an adaptive gradient clipping optimization. Extensive experiments on four benchmark datasets, namely MNIST, FMNIST, CIFAR-10 and IMDb, demonstrate the superior effectiveness of DPIS over existing solutions for deep learning with differential privacy.

This paper investigates the problem of forecasting multivariate aggregated human mobility while preserving the privacy of the individuals concerned. Differential privacy, a state-of-the-art formal notion, has been used as the privacy guarantee in two different and independent steps when training deep learning models. On one hand, we considered \textit{gradient perturbation}, which uses the differentially private stochastic gradient descent algorithm to guarantee the privacy of each time series sample in the learning stage. On the other hand, we considered \textit{input perturbation}, which adds differential privacy guarantees in each sample of the series before applying any learning. We compared four state-of-the-art recurrent neural networks: Long Short-Term Memory, Gated Recurrent Unit, and their Bidirectional architectures, i.e., Bidirectional-LSTM and Bidirectional-GRU. Extensive experiments were conducted with a real-world multivariate mobility dataset, which we published openly along with this paper. As shown in the results, differentially private deep learning models trained under gradient or input perturbation achieve nearly the same performance as non-private deep learning models, with loss in performance varying between $0.57\%$ to $2.8\%$. The contribution of this paper is significant for those involved in urban planning and decision-making, providing a solution to the human mobility multivariate forecast problem through differentially private deep learning models., Comment: Final version accepted in the journal Neural Computing and Applications

This paper investigates the problem of collecting multidimensional data throughout time (i.e., longitudinal studies) for the fundamental task of frequency estimation under Local Differential Privacy (LDP) guarantees. Contrary to frequency estimation of a single attribute, the multidimensional aspect demands particular attention to the privacy budget. Besides, when collecting user statistics longitudinally, privacy progressively degrades. Indeed, the "multiple" settings in combination (i.e., many attributes and several collections throughout time) impose several challenges, for which this paper proposes the first solution for frequency estimates under LDP. To tackle these issues, we extend the analysis of three state-of-the-art LDP protocols (Generalized Randomized Response -- GRR, Optimized Unary Encoding -- OUE, and Symmetric Unary Encoding -- SUE) for both longitudinal and multidimensional data collections. While the known literature uses OUE and SUE for two rounds of sanitization (a.k.a. memoization), i.e., L-OUE and L-SUE, respectively, we analytically and experimentally show that starting with OUE and then with SUE provides higher data utility (i.e., L-OSUE). Also, for attributes with small domain sizes, we propose Longitudinal GRR (L-GRR), which provides higher utility than the other protocols based on unary encoding. Last, we also propose a new solution named Adaptive LDP for LOngitudinal and Multidimensional FREquency Estimates (ALLOMFREE), which randomly samples a single attribute to be sent with the whole privacy budget and adaptively selects the optimal protocol, i.e., either L-GRR or L-OSUE. As shown in the results, ALLOMFREE consistently and considerably outperforms the state-of-the-art L-SUE and L-OUE protocols in the quality of the frequency estimates., Comment: Final version accepted in the journal Digital Communications and Networks. Version of Record: https://doi.org/10.1016/j.dcan.2022.07.003

With local differential privacy (LDP), users can privatize their data and thus guarantee privacy properties before transmitting it to the server (a.k.a. the aggregator). One primary objective of LDP is frequency (or histogram) estimation, in which the aggregator estimates the number of users for each possible value. In practice, when a study with rich content on a population is desired, the interest is in the multiple attributes of the population, that is to say, in multidimensional data ($d \geq 2$). However, contrary to the problem of frequency estimation of a single attribute (the majority of the works), the multidimensional aspect imposes to pay particular attention to the privacy budget. This one can indeed grow extremely quickly due to the composition theorem. To the authors' knowledge, two solutions seem to stand out for this task: 1) splitting the privacy budget for each attribute, i.e., send each value with $\frac{\epsilon}{d}$-LDP (Spl), and 2) random sampling a single attribute and spend all the privacy budget to send it with $\epsilon$-LDP (Smp). Although Smp adds additional sampling error, it has proven to provide higher data utility than the former Spl solution. However, we argue that aggregators (who are also seen as attackers) are aware of the sampled attribute and its LDP value, which is protected by a "less strict" $e^{\epsilon}$ probability bound (rather than $e^{\epsilon/d}$). This way, we propose a solution named Random Sampling plus Fake Data (RS+FD), which allows creating uncertainty over the sampled attribute by generating fake data for each non-sampled attribute; RS+FD further benefits from amplification by sampling. We theoretically and experimentally validate our proposed solution on both synthetic and real-world datasets to show that RS+FD achieves nearly the same or better utility than the state-of-the-art Smp solution.

InstaHide is a state-of-the-art mechanism for protecting private training images, by mixing multiple private images and modifying them such that their visual features are indistinguishable to the naked eye. In recent work, however, Carlini et al. show that it is possible to reconstruct private images from the encrypted dataset generated by InstaHide. Nevertheless, we demonstrate that Carlini et al.'s attack can be easily defeated by incorporating data augmentation into InstaHide. This leads to a natural question: is InstaHide with data augmentation secure? In this paper, we provide a negative answer to this question, by devising an attack for recovering private images from the outputs of InstaHide even when data augmentation is present. The basic idea is to use a comparative network to identify encrypted images that are likely to correspond to the same private image, and then employ a fusion-denoising network for restoring the private image from the encrypted ones, taking into account the effects of data augmentation. Extensive experiments demonstrate the effectiveness of the proposed attack in comparison to Carlini et al.'s attack., Comment: 15 pages

Federated learning (FL) is an emerging paradigm that enables multiple organizations to jointly train a model without revealing their private data to each other. This paper studies {\it vertical} federated learning, which tackles the scenarios where (i) collaborating organizations own data of the same set of users but with disjoint features, and (ii) only one organization holds the labels. We propose Pivot, a novel solution for privacy preserving vertical decision tree training and prediction, ensuring that no intermediate information is disclosed other than those the clients have agreed to release (i.e., the final tree model and the prediction output). Pivot does not rely on any trusted third party and provides protection against a semi-honest adversary that may compromise $m-1$ out of $m$ clients. We further identify two privacy leakages when the trained decision tree model is released in plaintext and propose an enhanced protocol to mitigate them. The proposed solution can also be extended to tree ensemble models, e.g., random forest (RF) and gradient boosting decision tree (GBDT) by treating single decision trees as building blocks. Theoretical and experimental analysis suggest that Pivot is efficient for the privacy achieved., Comment: Proc. VLDB Endow. 13(11): 2090-2103 (2020)

Local differential privacy (LDP) is a recently proposed privacy standard for collecting and analyzing data, which has been used, e.g., in the Chrome browser, iOS and macOS. In LDP, each user perturbs her information locally, and only sends the randomized version to an aggregator who performs analyses, which protects both the users and the aggregator against private information leaks. Although LDP has attracted much research attention in recent years, the majority of existing work focuses on applying LDP to complex data and/or analysis tasks. In this paper, we point out that the fundamental problem of collecting multidimensional data under LDP has not been addressed sufficiently, and there remains much room for improvement even for basic tasks such as computing the mean value over a single numeric attribute under LDP. Motivated by this, we first propose novel LDP mechanisms for collecting a numeric attribute, whose accuracy is at least no worse (and usually better) than existing solutions in terms of worst-case noise variance. Then, we extend these mechanisms to multidimensional data that can contain both numeric and categorical attributes, where our mechanisms always outperform existing solutions regarding worst-case noise variance. As a case study, we apply our solutions to build an LDP-compliant stochastic gradient descent algorithm (SGD), which powers many important machine learning tasks. Experiments using real datasets confirm the effectiveness of our methods, and their advantages over existing solutions., Comment: 12-Page Full Paper in Proceedings of the 2019 IEEE International Conference on Data Engineering (ICDE). arXiv admin note: text overlap with arXiv:1606.05053

It is challenging for stochastic optimizations to handle large-scale sensitive data safely. Recently, Duchi et al. proposed private sampling strategy to solve privacy leakage in stochastic optimizations. However, this strategy leads to robustness degeneration, since this strategy is equal to the noise injection on each gradient, which adversely affects updates of the primal variable. To address this challenge, we introduce a robust stochastic optimization under the framework of local privacy, which is called Privacy-pREserving StochasTIc Gradual lEarning (PRESTIGE). PRESTIGE bridges private updates of the primal variable (by private sampling) with the gradual curriculum learning (CL). Specifically, the noise injection leads to the issue of label noise, but the robust learning process of CL can combat with label noise. Thus, PRESTIGE yields "private but robust" updates of the primal variable on the private curriculum, namely an reordered label sequence provided by CL. In theory, we reveal the convergence rate and maximum complexity of PRESTIGE. Empirical results on six datasets show that, PRESTIGE achieves a good tradeoff between privacy preservation and robustness over baselines.

Differential privacy is a promising privacy-preserving paradigm for statistical query processing over sensitive data. It works by injecting random noise into each query result, such that it is provably hard for the adversary to infer the presence or absence of any individual record from the published noisy results. The main objective in differentially private query processing is to maximize the accuracy of the query results, while satisfying the privacy guarantees. Previous work, notably \cite{LHR+10}, has suggested that with an appropriate strategy, processing a batch of correlated queries as a whole achieves considerably higher accuracy than answering them individually. However, to our knowledge there is currently no practical solution to find such a strategy for an arbitrary query batch; existing methods either return strategies of poor quality (often worse than naive methods) or require prohibitively expensive computations for even moderately large domains. Motivated by this, we propose the \emph{Low-Rank Mechanism} (LRM), the first practical differentially private technique for answering batch queries with high accuracy, based on a \emph{low rank approximation} of the workload matrix. We prove that the accuracy provided by LRM is close to the theoretical lower bound for any mechanism to answer a batch of queries under differential privacy. Extensive experiments using real data demonstrate that LRM consistently outperforms state-of-the-art query processing solutions under differential privacy, by large margins., Comment: VLDB 2012