Privacy-preserving instance encoding aims to encode raw data as feature vectors without revealing their privacy-sensitive information. When designed properly, these encodings can be used for downstream ML applications such as training and inference with limited privacy risk. However, the vast majority of existing instance encoding schemes are based on heuristics and their privacy-preserving properties are only validated empirically against a limited set of attacks. In this paper, we propose a theoretically-principled measure for the privacy of instance encoding based on Fisher information. We show that our privacy measure is intuitive, easily applicable, and can be used to bound the invertibility of encodings both theoretically and empirically.
相關內容
Federated Learning (FL) allows multiple participating clients to train machine learning models collaboratively by keeping their datasets local and only exchanging the gradient or model updates with a coordinating server. Existing FL protocols were shown to be vulnerable to attacks that aim to compromise data privacy and/or model robustness. Recently proposed defenses focused on ensuring either privacy or robustness, but not both. In this paper, we focus on simultaneously achieving differential privacy (DP) and Byzantine robustness for cross-silo FL, based on the idea of learning from history. The robustness is achieved via client momentum, which averages the updates of each client over time, thus reduces the variance of the honest clients and exposes the small malicious perturbations of Byzantine clients that are undetectable in a single round but accumulate over time. In our initial solution DP-BREM, the DP property is achieved via adding noise to the aggregated momentum, and we account for the privacy cost from the momentum, which is different from the conventional DP-SGD that accounts for the privacy cost from gradient. Since DP-BREM assumes a trusted server (who can obtain clients' local models or updates), we further develop the final solution called DP-BREM+, which achieves the same DP and robustness properties as DP-BREM without a trusted server by utilizing secure aggregation techniques, where DP noise is securely and jointly generated by the clients. Our theoretical analysis on the convergence rate and experimental results under different DP guarantees and attack settings demonstrate that our proposed protocols achieve better privacy-utility tradeoff and stronger Byzantine robustness than several baseline methods.
Sampling schemes are fundamental tools in statistics, survey design, and algorithm design. A fundamental result in differential privacy is that a differentially private mechanism run on a simple random sample of a population provides stronger privacy guarantees than the same algorithm run on the entire population. However, in practice, sampling designs are often more complex than the simple, data-independent sampling schemes that are addressed in prior work. In this work, we extend the study of privacy amplification results to more complex, data-dependent sampling schemes. We find that not only do these sampling schemes often fail to amplify privacy, they can actually result in privacy degradation. We analyze the privacy implications of the pervasive cluster sampling and stratified sampling paradigms, as well as provide some insight into the study of more general sampling designs.
There is increasing appetite for analysing populations of network data due to the fast-growing body of applications demanding such methods. While methods exist to provide readily interpretable summaries of heterogeneous network populations, these are often descriptive or ad hoc, lacking any formal justification. In contrast, principled analysis methods often provide results difficult to relate back to the applied problem of interest. Motivated by two complementary applied examples, we develop a Bayesian framework to appropriately model complex heterogeneous network populations, whilst also allowing analysts to gain insights from the data, and make inferences most relevant to their needs. The first application involves a study in Computer Science measuring human movements across a University. The second analyses data from Neuroscience investigating relationships between different regions of the brain. While both applications entail analysis of a heterogeneous population of networks, network sizes vary considerably. We focus on the problem of clustering the elements of a network population, where each cluster is characterised by a network representative. We take advantage of the Bayesian machinery to simultaneously infer the cluster membership, the representatives, and the community structure of the representatives, thus allowing intuitive inferences to be made. The implementation of our method on the human movement study reveals interesting movement patterns of individuals in clusters, readily characterised by their network representative. For the brain networks application, our model reveals a cluster of individuals with different network properties of particular interest in Neuroscience. The performance of our method is additionally validated in extensive simulation studies.
The excessive use of images in social networks, government databases, and industrial applications has posed great privacy risks and raised serious concerns from the public. Even though differential privacy (DP) is a widely accepted criterion that can provide a provable privacy guarantee, the application of DP on unstructured data such as images is not trivial due to the lack of a clear qualification on the meaningful difference between any two images. In this paper, for the first time, we introduce a novel notion of image-aware differential privacy, referred to as DP-image, that can protect user's personal information in images, from both human and AI adversaries. The DP-Image definition is formulated as an extended version of traditional differential privacy, considering the distance measurements between feature space vectors of images. Then we propose a mechanism to achieve DP-Image by adding noise to an image feature vector. Finally, we conduct experiments with a case study on face image privacy. Our results show that the proposed DP-Image method provides excellent DP protection on images, with a controllable distortion to faces.
Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{//github.com/woodyx218/opacus_global_clipping}.
We propose and study a new privacy definition, termed Probably Approximately Correct (PAC) Privacy. PAC Privacy characterizes the information-theoretic hardness to recover sensitive data given arbitrary information disclosure/leakage during/after any processing. Unlike the classic cryptographic definition and Differential Privacy (DP), which consider the adversarial (input-independent) worst case, PAC Privacy is a simulatable metric that quantifies the instance-based impossibility of inference. A fully automatic analysis and proof generation framework is proposed: security parameters can be produced with arbitrarily high confidence via Monte-Carlo simulation for any black-box data processing oracle. This appealing automation property enables analysis of complicated data processing, where the worst-case proof in the classic privacy regime could be loose or even intractable. Moreover, we show that the produced PAC Privacy guarantees enjoy simple composition bounds and the automatic analysis framework can be implemented in an online fashion to analyze the composite PAC Privacy loss even under correlated randomness. On the utility side, the magnitude of (necessary) perturbation required in PAC Privacy is not lower bounded by Theta(\sqrt{d}) for a d-dimensional release but could be O(1) for many practical data processing tasks, which is in contrast to the input-independent worst-case information-theoretic lower bound. Example applications of PAC Privacy are included with comparisons to existing works.
Machine learning (ML) models are costly to train as they can require a significant amount of data, computational resources and technical expertise. Thus, they constitute valuable intellectual property that needs protection from adversaries wanting to steal them. Ownership verification techniques allow the victims of model stealing attacks to demonstrate that a suspect model was in fact stolen from theirs. Although a number of ownership verification techniques based on watermarking or fingerprinting have been proposed, most of them fall short either in terms of security guarantees (well-equipped adversaries can evade verification) or computational cost. A fingerprinting technique, Dataset Inference (DI), has been shown to offer better robustness and efficiency than prior methods. The authors of DI provided a correctness proof for linear (suspect) models. However, in a subspace of the same setting, we prove that DI suffers from high false positives (FPs) -- it can incorrectly identify an independent model trained with non-overlapping data from the same distribution as stolen. We further prove that DI also triggers FPs in realistic, non-linear suspect models. We then confirm empirically that DI in the black-box setting leads to FPs, with high confidence. Second, we show that DI also suffers from false negatives (FNs) -- an adversary can fool DI (at the cost of incurring some accuracy loss) by regularising a stolen model's decision boundaries using adversarial training, thereby leading to an FN. To this end, we demonstrate that black-box DI fails to identify a model adversarially trained from a stolen dataset -- the setting where DI is the hardest to evade. Finally, we discuss the implications of our findings, the viability of fingerprinting-based ownership verification in general, and suggest directions for future work.
In decentralized settings, the shuffle model of differential privacy has emerged as a promising alternative to the classical local model. Analyzing privacy amplification via shuffling is a critical component in both single-message and multi-message shuffle protocols. However, current methods used in these two areas are distinct and specific, making them less convenient for protocol designers and practitioners. In this work, we introduce variation-ratio reduction as a unified framework for privacy amplification analyses in the shuffle model. This framework utilizes total variation bounds of local messages and probability ratio bounds of other users' blanket messages, converting them to indistinguishable levels. Our results indicate that the framework yields tighter bounds for both single-message and multi-message encoders (e.g., with local DP, local metric DP, or multi-message randomizers). Specifically, for a broad range of local randomizers having extremal probability design, our amplification bounds are precisely tight. We also demonstrate that variation-ratio reduction is well-suited for parallel composition in the shuffle model and results in stricter privacy accounting for common sampling-based local randomizers. Our experimental findings show that, compared to existing amplification bounds, our numerical amplification bounds can save up to $30\%$ of the budget for single-message protocols, $75\%$ of the budget for multi-message protocols, and $75\%$-$95\%$ of the budget for parallel composition. Additionally, our implementation for numerical amplification bounds has only $\tilde{O}(n)$ complexity and is highly efficient in practice, taking just $10$ seconds for $n=10^8$ users. The code for our implementation can be found at \url{//github.com/wangsw/PrivacyAmplification}.
User reporting is an essential component of content moderation on many online platforms -- in particular, on end-to-end encrypted (E2EE) messaging platforms where platform operators cannot proactively inspect message contents. However, users' privacy concerns when considering reporting may impede the effectiveness of this strategy in regulating online harassment. In this paper, we conduct interviews with 16 users of E2EE platforms to understand users' mental models of how reporting works and their resultant privacy concerns and considerations surrounding reporting. We find that users expect platforms to store rich longitudinal reporting datasets, recognizing both their promise for better abuse mitigation and the privacy risk that platforms may exploit or fail to protect them. We also find that users have preconceptions about the respective capabilities and risks of moderators at the platform versus community level -- for instance, users trust platform moderators more to not abuse their power but think community moderators have more time to attend to reports. These considerations, along with perceived effectiveness of reporting and how to provide sufficient evidence while maintaining privacy, shape how users decide whether, to whom, and how much to report. We conclude with design implications for a more privacy-preserving reporting system on E2EE messaging platforms.
There is active debate over whether to consider patient race and ethnicity when estimating disease risk. By accounting for race and ethnicity, it is possible to improve the accuracy of risk predictions, but there is concern that their use may encourage a racialized view of medicine. In diabetes risk models, despite substantial gains in statistical accuracy from using race and ethnicity, the gains in clinical utility are surprisingly modest. These modest clinical gains stem from two empirical patterns: first, the vast majority of individuals receive the same screening recommendation regardless of whether race or ethnicity are included in risk models; and second, for those who do receive different screening recommendations, the difference in utility between screening and not screening is relatively small. Our results are based on broad statistical principles, and so are likely to generalize to many other risk-based clinical decisions.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191