隐私攻击

We study the problem of counting the number of distinct elements in a dataset subject to the constraint of differential privacy. We consider the challenging setting of person-level DP (a.k.a. user-level DP) where each person may contribute an unbounded number of items and hence the sensitivity is unbounded. Our approach is to compute a bounded-sensitivity version of this query, which reduces to solving a max-flow problem. The sensitivity bound is optimized to balance the noise we must add to privatize the answer against the error of the approximation of the bounded-sensitivity query to the true number of unique elements.

We study indiscriminate poisoning for linear learners where an adversary injects a few crafted examples into the training data with the goal of forcing the induced model to incur higher test error. Inspired by the observation that linear learners on some datasets are able to resist the best known attacks even without any defenses, we further investigate whether datasets can be inherently robust to indiscriminate poisoning attacks for linear learners. For theoretical Gaussian distributions, we rigorously characterize the behavior of an optimal poisoning attack, defined as the poisoning strategy that attains the maximum risk of the induced model at a given poisoning budget. Our results prove that linear learners can indeed be robust to indiscriminate poisoning if the class-wise data distributions are well-separated with low variance and the size of the constraint set containing all permissible poisoning points is also small. These findings largely explain the drastic variation in empirical attack performance of the state-of-the-art poisoning attacks on linear learners across benchmark datasets, making an important initial step towards understanding the underlying reasons some learning tasks are vulnerable to data poisoning attacks.

With the growing demand for the right to be forgotten, there is an increasing need for machine learning models to forget sensitive data and its impact. To address this, the paradigm of selective forgetting (a.k.a machine unlearning) has been extensively studied, which aims to remove the impact of requested data from a well-trained model without retraining from scratch. Despite its significant success, limited attention has been given to the security vulnerabilities of the unlearning system concerning malicious data update requests. Motivated by this, in this paper, we explore the possibility and feasibility of malicious data update requests during the unlearning process. Specifically, we first propose a new class of malicious selective forgetting attacks, which involves a static scenario where all the malicious data update requests are provided by the adversary at once. Additionally, considering the sequential setting where the data update requests arrive sequentially, we also design a novel framework for sequential forgetting attacks, which is formulated as a stochastic optimal control problem. We also propose novel optimization algorithms that can find the effective malicious data update requests. We perform theoretical analyses for the proposed selective forgetting attacks, and extensive experimental results validate the effectiveness of our proposed selective forgetting attacks. The source code is available in the supplementary material.

Large language models (LLMs) are excellent in-context learners. However, the sensitivity of data contained in prompts raises privacy concerns. Our work first shows that these concerns are valid: we instantiate a simple but highly effective membership inference attack against the data used to prompt LLMs. To address this vulnerability, one could forego prompting and resort to fine-tuning LLMs with known algorithms for private gradient descent. However, this comes at the expense of the practicality and efficiency offered by prompting. Therefore, we propose to privately learn to prompt. We first show that soft prompts can be obtained privately through gradient descent on downstream data. However, this is not the case for discrete prompts. Thus, we orchestrate a noisy vote among an ensemble of LLMs presented with different prompts, i.e., a flock of stochastic parrots. The vote privately transfers the flock's knowledge into a single public prompt. We show that LLMs prompted with our private algorithms closely match the non-private baselines. For example, using GPT3 as the base model, we achieve a downstream accuracy of 92.7% on the sst2 dataset with $(\varepsilon=0.147, \delta=10^{-6})$-differential privacy vs. 95.2% for the non-private baseline. Through our experiments, we also show that our prompt-based approach is easily deployed with existing commercial~APIs.

Deep Gradient Leakage (DGL) is a highly effective attack that recovers private training images from gradient vectors. This attack casts significant privacy challenges on distributed learning from clients with sensitive data, where clients are required to share gradients. Defending against such attacks requires but lacks an understanding of when and how privacy leakage happens, mostly because of the black-box nature of deep networks. In this paper, we propose a novel Inversion Influence Function (I$^2$F) that establishes a closed-form connection between the recovered images and the private gradients by implicitly solving the DGL problem. Compared to directly solving DGL, I$^2$F is scalable for analyzing deep networks, requiring only oracle access to gradients and Jacobian-vector products. We empirically demonstrate that I$^2$F effectively approximated the DGL generally on different model architectures, datasets, modalities, attack implementations, and perturbation-based defenses. With this novel tool, we provide insights into effective gradient perturbation directions, the unfairness of privacy protection, and privacy-preferred model initialization. Our codes are provided in https://github.com/illidanlab/inversion-influence-function.

The adversarial vulnerability of Deep Neural Networks (DNNs) has been well-known and widely concerned, often under the context of learning top-$1$ attacks (e.g., fooling a DNN to classify a cat image as dog). This paper shows that the concern is much more serious by learning significantly more aggressive ordered top-$K$ clear-box targeted attacks proposed in~\citep{zhang2020learning}. We propose a novel and rigorous quadratic programming (QP) method of learning ordered top-$K$ attacks with low computing cost, dubbed as \textbf{QuadAttac$K$}. Our QuadAttac$K$ directly solves the QP to satisfy the attack constraint in the feature embedding space (i.e., the input space to the final linear classifier), which thus exploits the semantics of the feature embedding space (i.e., the principle of class coherence). With the optimized feature embedding vector perturbation, it then computes the adversarial perturbation in the data space via the vanilla one-step back-propagation. In experiments, the proposed QuadAttac$K$ is tested in the ImageNet-1k classification using ResNet-50, DenseNet-121, and Vision Transformers (ViT-B and DEiT-S). It successfully pushes the boundary of successful ordered top-$K$ attacks from $K=10$ up to $K=20$ at a cheap budget ($1\times 60$) and further improves attack success rates for $K=5$ for all tested models, while retaining the performance for $K=1$.

Membership inference attacks are designed to determine, using black box access to trained models, whether a particular example was used in training or not. Membership inference can be formalized as a hypothesis testing problem. The most effective existing attacks estimate the distribution of some test statistic (usually the model's confidence on the true label) on points that were (and were not) used in training by training many \emph{shadow models}---i.e. models of the same architecture as the model being attacked, trained on a random subsample of data. While effective, these attacks are extremely computationally expensive, especially when the model under attack is large. \footnotetext[0]{ Martin and Shuai are the lead authors, and other authors are ordered alphabetically. \{maberlop,shuat\}@amazon.com} We introduce a new class of attacks based on performing quantile regression on the distribution of confidence scores induced by the model under attack on points that are not used in training. We show that our method is competitive with state-of-the-art shadow model attacks, while requiring substantially less compute because our attack requires training only a single model. Moreover, unlike shadow model attacks, our proposed attack does not require any knowledge of the architecture of the model under attack and is therefore truly ``black-box". We show the efficacy of this approach in an extensive series of experiments on various datasets and model architectures. Our code is available at \href{https://github.com/amazon-science/quantile-mia}{github.com/amazon-science/quantile-mia.}

In federated frequency estimation (FFE), multiple clients work together to estimate the frequency of their local data by communicating with a server, while maintaining the security constraint of $\mathtt{secsum}$ where the server can only access the sum of client-held vectors. For FFE with a single communication round, it is known that count sketch is nearly information-theoretically optimal [Chen et al., 2022]. However, when multiple communication rounds are allowed, we propose a new sketch algorithm that is provably more accurate than a naive adaptation of count sketch. Furthermore, we show that both our sketch algorithm and count sketch can achieve better accuracy when the problem instance is simpler. Therefore, we propose a two-phase approach to enable the use of a smaller sketch size for simpler problems. Finally, we provide mechanisms to make our proposed algorithm differentially private. We verify the performance of our methods through experiments conducted on real datasets.

Differentially private (DP) machine learning algorithms incur many sources of randomness, such as random initialization, random batch subsampling, and shuffling. However, such randomness is difficult to take into account when proving differential privacy bounds because it induces mixture distributions for the algorithm's output that are difficult to analyze. This paper focuses on improving privacy bounds for shuffling models and one-iteration differentially private gradient descent (DP-GD) with random initializations using $f$-DP. We derive a closed-form expression of the trade-off function for shuffling models that outperforms the most up-to-date results based on $(\epsilon,\delta)$-DP. Moreover, we investigate the effects of random initialization on the privacy of one-iteration DP-GD. Our numerical computations of the trade-off function indicate that random initialization can enhance the privacy of DP-GD. Our analysis of $f$-DP guarantees for these mixture mechanisms relies on an inequality for trade-off functions introduced in this paper. This inequality implies the joint convexity of $F$-divergences. Finally, we study an $f$-DP analog of the advanced joint convexity of the hockey-stick divergence related to $(\epsilon,\delta)$-DP and apply it to analyze the privacy of mixture mechanisms.

Federated learning (FL) provides a distributed training paradigm where multiple clients can jointly train a global model without sharing their local data. However, recent studies have shown that FL offers an additional surface for backdoor attacks. For instance, an attacker can compromise a subset of clients and thus corrupt the global model to misclassify an input with a backdoor trigger as the adversarial target. Existing defenses for FL against backdoor attacks usually detect and exclude the corrupted information from the compromised clients based on a static attacker model. However, such defenses are inadequate against dynamic attackers who strategically adapt their attack strategies. To bridge this gap, we model the strategic interactions between the defender and dynamic attackers as a minimax game. Based on the analysis of the game, we design an interactive defense mechanism FedGame. We prove that under mild assumptions, the global model trained with FedGame under backdoor attacks is close to that trained without attacks. Empirically, we compare FedGame with multiple state-of-the-art baselines on several benchmark datasets under various attacks. We show that FedGame can effectively defend against strategic attackers and achieves significantly higher robustness than baselines. Our code is available at: https://github.com/AI-secure/FedGame.

Self-supervised learning (SSL) algorithms can produce useful image representations by learning to associate different parts of natural images with one another. However, when taken to the extreme, SSL models can unintendedly memorize specific parts in individual training samples rather than learning semantically meaningful associations. In this work, we perform a systematic study of the unintended memorization of image-specific information in SSL models -- which we refer to as déjà vu memorization. Concretely, we show that given the trained model and a crop of a training image containing only the background (e.g., water, sky, grass), it is possible to infer the foreground object with high accuracy or even visually reconstruct it. Furthermore, we show that déjà vu memorization is common to different SSL algorithms, is exacerbated by certain design choices, and cannot be detected by conventional techniques for evaluating representation quality. Our study of déjà vu memorization reveals previously unknown privacy risks in SSL models, as well as suggests potential practical mitigation strategies.

Privacy and communication constraints are two major bottlenecks in federated learning (FL) and analytics (FA). We study the optimal accuracy of mean and frequency estimation (canonical models for FL and FA respectively) under joint communication and $(\varepsilon, \delta)$-differential privacy (DP) constraints. We consider both the central and the multi-message shuffled DP models. We show that in order to achieve the optimal $\ell_2$ error under $(\varepsilon, \delta)$-DP, it is sufficient for each client to send $\Theta\left( n \min\left(\varepsilon, \varepsilon^2\right)\right)$ bits for FL %{\color{blue}(assuming the dimension $d \gg n \min\left(\varepsilon, \varepsilon^2\right)$)} and $\Theta\left(\log\left( n\min\left(\varepsilon, \varepsilon^2\right) \right)\right)$ bits for FA to the server, where $n$ is the number of participating clients. Without compression, each client needs $O(d)$ bits and $O\left(\log d\right)$ bits for the mean and frequency estimation problems respectively (where $d$ corresponds to the number of trainable parameters in FL or the domain size in FA), meaning that we can get significant savings in the regime $ n \min\left(\varepsilon, \varepsilon^2\right) = o(d)$, which is often the relevant regime in practice. We propose two different ways to leverage compression for privacy amplification and achieve the optimal privacy-communication-accuracy trade-offs. In both cases, each client communicates only partial information about its sample and we show that privacy is amplified by randomly selecting the part contributed by each client. In the first method, the random selection is revealed to the server, which results in a central DP guarantee with optimal privacy-communication-accuracy trade-offs. In the second method, the random data parts from the clients are shuffled by a secure shuffler resulting in a multi-message shuffling scheme with the same optimal trade-offs. As a result, we establish the optimal three-way trade-offs between privacy, communication, and accuracy for both the central DP and multi-message shuffling frameworks.

A well-known numerical bottleneck in the differentially-private stochastic gradient descent (DP-SGD) algorithm is the computation of the gradient norm for each example in a large input batch. When the loss function in DP-SGD is consists of an intermediate linear operation, existing methods in the literature have proposed decompositions of gradients that are amenable to fast norm computations. In this paper, we present a framework that generalizes the above approach to arbitrary (possibly nonlinear) intermediate operations. Moreover, we show that for certain operations, such as fully-connected and embedding layer computations, further improvements to the runtime and storage costs of existing decompositions can be deduced using certain components of our framework. Finally, preliminary numerical experiments are given to demonstrate the substantial effects of the aforementioned improvements.

We study private and robust multi-armed bandits (MABs), where the agent receives Huber's contaminated heavy-tailed rewards and meanwhile needs to ensure differential privacy. We consider both the finite $k$-th raw moment and the finite $k$-th central moment settings for heavy-tailed rewards distributions with $k\ge 2$. We first present its minimax lower bound, characterizing the information-theoretic limit of regret with respect to privacy budget, contamination level, and heavy-tailedness. Then, we propose a meta-algorithm that builds on a private and robust mean estimation sub-routine \texttt{PRM} that essentially relies on reward truncation and the Laplace mechanism. For the above two different heavy-tailed settings, we give corresponding schemes of \texttt{PRM}, which enable us to achieve nearly-optimal regrets. Moreover, our two proposed truncation-based or histogram-based \texttt{PRM} schemes achieve the optimal trade-off between estimation accuracy, privacy and robustness. Finally, we support our theoretical results and show the effectiveness of our algorithms with experimental studies.

We propose a scheme for auditing differentially private machine learning systems with a single training run. This exploits the parallelism of being able to add or remove multiple training examples independently. We analyze this using the connection between differential privacy and statistical generalization, which avoids the cost of group privacy. Our auditing scheme requires minimal assumptions about the algorithm and can be applied in the black-box or white-box setting. We demonstrate the effectiveness of our framework by applying it to DP-SGD, where we can achieve meaningful empirical privacy lower bounds by training only one model. In contrast, standard methods would require training hundreds of models.

Hand-crafted image quality metrics, such as PSNR and SSIM, are commonly used to evaluate model privacy risk under reconstruction attacks. Under these metrics, reconstructed images that are determined to resemble the original one generally indicate more privacy leakage. Images determined as overall dissimilar, on the other hand, indicate higher robustness against attack. However, there is no guarantee that these metrics well reflect human opinions, which offers trustworthy judgement for model privacy leakage. In this paper, we comprehensively study the faithfulness of these hand-crafted metrics to human perception of privacy information from the reconstructed images. On 5 datasets ranging from natural images, faces, to fine-grained classes, we use 4 existing attack methods to reconstruct images from many different classification models and, for each reconstructed image, we ask multiple human annotators to assess whether this image is recognizable. Our studies reveal that the hand-crafted metrics only have a weak correlation with the human evaluation of privacy leakage and that even these metrics themselves often contradict each other. These observations suggest risks of current metrics in the community. To address this potential risk, we propose a learning-based measure called SemSim to evaluate the Semantic Similarity between the original and reconstructed images. SemSim is trained with a standard triplet loss, using an original image as an anchor, one of its recognizable reconstructed images as a positive sample, and an unrecognizable one as a negative. By training on human annotations, SemSim exhibits a greater reflection of privacy leakage on the semantic level. We show that SemSim has a significantly higher correlation with human judgment compared with existing metrics. Moreover, this strong correlation generalizes to unseen datasets, models and attack methods. We envision this work as a milestone for image quality evaluation closer to the human level. The project webpage can be accessed at https://sites.google.com/view/semsim.

In high-stakes domains such as legal, banking, hiring, and healthcare, learning models frequently rely on sensitive user information for inference, necessitating the complete set of features. This not only poses significant privacy risks for individuals but also demands substantial human effort from organizations to verify information accuracy. This study asks whether it is necessary to use all input features for accurate predictions at inference time. The paper demonstrates that, in a personalized setting, individuals may only need to disclose a small subset of features without compromising decision-making accuracy. The paper also provides an efficient sequential algorithm to determine the appropriate attributes for each individual to provide. Evaluations across various learning tasks show that individuals can potentially report as little as 10\% of their information while maintaining the same accuracy level as a model that employs the full set of user information.

We study design of black-box model extraction attacks that can *send minimal number of queries from* a *publicly available dataset* to a target ML model through a predictive API with an aim *to create an informative and distributionally equivalent replica* of the target. First, we define *distributionally equivalent* and *Max-Information model extraction* attacks, and reduce them into a variational optimisation problem. The attacker sequentially solves this optimisation problem to select the most informative queries that simultaneously maximise the entropy and reduce the mismatch between the target and the stolen models. This leads to *an active sampling-based query selection algorithm*, Marich, which is *model-oblivious*. Then, we evaluate Marich on different text and image data sets, and different models, including CNNs and BERT. Marich extracts models that achieve $\sim 60-95\%$ of true model's accuracy and uses $\sim 1,000 - 8,500$ queries from the publicly available datasets, which are different from the private training datasets. Models extracted by Marich yield prediction distributions, which are $\sim2-4\times$ closer to the target's distribution in comparison to the existing active sampling-based attacks. The extracted models also lead to 84-96$\%$ accuracy under membership inference attacks. Experimental results validate that Marich is *query-efficient*, and capable of performing task-accurate, high-fidelity, and informative model extraction.

We propose a new family of label randomizers for training _regression_ models under the constraint of label differential privacy (DP). In particular, we leverage the trade-offs between bias and variance to construct better label randomizers depending on a privately estimated prior distribution over the labels. We demonstrate that these randomizers achieve state-of-the-art privacy-utility trade-offs on several datasets, highlighting the importance of reducing bias when training neural networks with label DP. We also provide theoretical results shedding light on the structural properties of the optimal unbiased randomizers.

Model distillation is frequently proposed as a technique to reduce the privacy leakage of machine learning. These empirical privacy defenses rely on the intuition that distilled ``student'' models protect the privacy of training data, as they only interact with this data indirectly through a ``teacher'' model. In this work, we design membership inference attacks to systematically study the privacy provided by knowledge distillation to both the teacher and student training sets. Our new attacks show that distillation alone provides only limited privacy across a number of domains. We explain the success of our attacks on distillation by showing that membership inference attacks on a private dataset can succeed even if the target model is never queried on any actual training points, but only on inputs whose predictions are highly influenced by training data. Finally, we show that our attacks are strongest when student and teacher sets are similar, or when the attacker can poison the teacher set.