ICML 2023 Papers

We address the problem of stably and efficiently training a deep neural network robust to adversarial perturbations bounded by an $l_1$ norm. We demonstrate that achieving robustness against $l_1$-bounded perturbations is more challenging than in the $l_2$ or $l_\infty$ cases, because adversarial training against $l_1$-bounded perturbations is more likely to suffer from catastrophic overfitting and yield training instabilities. Our analysis links these issues to the coordinate descent strategy used in existing methods. We address this by introducing Fast-EG-$l_1$, an efficient adversarial training algorithm based on Euclidean geometry and free of coordinate descent. Fast-EG-$l_1$ comes with no additional memory costs and no extra hyper-parameters to tune. Our experimental results on various datasets demonstrate that Fast-EG-$l_1$ yields the best and most stable robustness against $l_1$-bounded adversarial attacks among the methods of comparable computational complexity. Code and the checkpoints are available at https://github.com/IVRL/FastAdvL.

Recently, the newly emerged multimodal models, which leverage both visual and linguistic modalities to train powerful encoders, have gained increasing attention. However, learning from a large-scale unlabeled dataset also exposes the model to the risk of potential poisoning attacks, whereby the adversary aims to perturb the model's training data to trigger malicious behaviors in it. In contrast to previous work, only poisoning visual modality, in this work, we take the first step to studying poisoning attacks against multimodal models in both visual and linguistic modalities. Specially, we focus on answering two questions: (1) Is the linguistic modality also vulnerable to poisoning attacks? and (2) Which modality is most vulnerable? To answer the two questions, we propose three types of poisoning attacks against multimodal models. Extensive evaluations on different datasets and model architectures show that all three attacks can achieve significant attack performance while maintaining model utility in both visual and linguistic modalities. Furthermore, we observe that the poisoning effect differs between different modalities. To mitigate the attacks, we propose both pre-training and post-training defenses. We empirically show that both defenses can significantly reduce the attack performance while preserving the model's utility. Our code is available at https://github.com/zqypku/mm_poison/.

We characterize the differential privacy guarantees of privacy mechanisms in the large-composition regime, i.e., when a privacy mechanism is sequentially applied a large number of times to sensitive data. Via exponentially tilting the privacy loss random variable, we derive a new formula for the privacy curve expressing it as a contour integral over an integration path that runs parallel to the imaginary axis with a free real-axis intercept. Then, using the method of steepest descent from mathematical physics, we demonstrate that the choice of saddle-point as the real-axis intercept yields closed-form accurate approximations of the desired contour integral. This procedure---dubbed the saddle-point accountant (SPA)---yields a constant-time accurate approximation of the privacy curve. Theoretically, our results can be viewed as a refinement of both Gaussian Differential Privacy and the moments accountant method found in Rényi Differential Privacy. In practice, we demonstrate through numerical experiments that the SPA provides a precise approximation of privacy guarantees competitive with purely numerical-based methods (such as FFT-based accountants), while enjoying closed-form mathematical expressions.

While graph neural networks (GNNs) dominate the state-of-the-art for exploring graphs in real-world applications, they have been shown to be vulnerable to a growing number of privacy attacks. For instance, link stealing is a well-known membership inference attack (MIA) on edges that infers the presence of an edge in a GNN's training graph. Recent studies on independent and identically distributed data (e.g., images) have empirically demonstrated that individuals from different groups suffer from different levels of privacy risks to MIAs, i.e., uneven vulnerability. However, theoretical evidence of such uneven vulnerability is missing. In this paper, we first present theoretical evidence of the uneven vulnerability of GNNs to link stealing attacks, which lays the foundation for demystifying such uneven risks among different groups of edges. We further demonstrate a group-based attack paradigm to expose the practical privacy harm to GNN users derived from the uneven vulnerability of edges. Finally, we empirically validate the existence of obvious uneven vulnerability on nine real-world datasets (e.g., about 25% AUC difference between different groups in the Credit graph). Compared with existing methods, the outperformance of our group-based attack paradigm confirms that customising different strategies for different groups results in more effective privacy attacks.

Graph Contrastive Learning (GCL) has attracted considerable interest due to its impressive node representation learning capability. Despite the wide application of GCL techniques, little attention has been paid to the security of GCL. In this paper, we systematically study the vulnerability of GCL in the presence of malicious backdoor adversaries. In particular, we propose *GCBA*, the first backdoor attack for graph contrastive learning. GCBA incorporates three attacks: poisoning, crafting, and natural backdoor, each targeting one stage of the GCL pipeline. We formulate our attacks as optimization problems and solve them with a novel discrete optimization technique to overcome the discrete nature of graph-structured data. By extensively evaluating GCBA on multiple datasets and GCL methods, we show that our attack can achieve high attack success rates while preserving stealthiness. We further consider potential countermeasures to our attack and conclude that existing defenses are insufficient to mitigate GCBA. We show that as a complex paradigm involving data and model republishing, GCL is vulnerable to backdoor attacks, and specifically designed defenses are needed to mitigate the backdoor attacks on GCL.

Federated learning (FL) aims to perform privacy-preserving machine learning on distributed data held by multiple data owners. To this end, FL requires the data owners to perform training locally and share the gradients or weight updates (instead of the private inputs) with the central server, which are then securely aggregated over multiple data owners. Although aggregation by itself does not offer provable privacy protection, prior work suggested that if the batch size is sufficiently large the aggregation may be secure enough. In this paper, we propose the Cocktail Party Attack (CPA) that, contrary to prior belief, is able to recover the private inputs from gradients/weight updates aggregated over as many as 1024 samples. CPA leverages the crucial insight that aggregate gradients from a fully connected (FC) layer is a linear combination of its inputs, which allows us to frame gradient inversion as a blind source separation (BSS) problem. We adapt independent component analysis (ICA)---a classic solution to the BSS problem---to recover private inputs for FC and convolutional networks, and show that CPA significantly outperforms prior gradient inversion attacks, scales to ImageNet-sized inputs, and works on large batch sizes of up to 1024.

We study the relationship between two desiderata of algorithms in statistical inference and machine learning---differential privacy and robustness to adversarial data corruptions. Their conceptual similarity was first observed by Dwork and Lei (STOC 2009), who observed that private algorithms satisfy robustness, and gave a general method for converting robust algorithms to private ones. However, all general methods for transforming robust algorithms into private ones lead to suboptimal error rates. Our work gives the first black-box transformation that converts any adversarially robust algorithm into one that satisfies pure differential privacy. Moreover, we show that for any low-dimensional estimation task, applying our transformation to an optimal robust estimator results in an optimal private estimator. Thus, we conclude that for any low-dimensional task, the optimal error rate for $\varepsilon$-differentially private estimators is essentially the same as the optimal error rate for estimators that are robust to adversarially corrupting $1/\varepsilon$ training samples. We apply our transformation to obtain new optimal private estimators for several high-dimensional statistical tasks, including Gaussian linear regression and PCA. Finally, we present an extension of our transformation that leads to approximately differentially private algorithms whose error does not depend on the range of the output space, which is impossible under pure differential privacy.

We study the problem of histogram estimation under user-level differential privacy, where the goal is to preserve the privacy of *all* entries of any single user. We consider the heterogeneous scenario where the quantity of data can be different for each user. In this scenario, the amount of noise injected into the histogram to obtain differential privacy is proportional to the maximum user contribution, which can be amplified by few outliers. One approach to circumvent this would be to bound (or limit) the contribution of each user to the histogram. However, if users are limited to small contributions, a significant amount of data will be discarded. In this work, we propose algorithms to choose the best user contribution bound for histogram estimation under both bounded and unbounded domain settings. When the size of the domain is bounded, we propose a user contribution bounding strategy that almost achieves a two-approximation with respect to the best contribution bound in hindsight. For unbounded domain histogram estimation, we propose an algorithm that is logarithmic-approximation with respect to the best contribution bound in hindsight. This result holds without any distribution assumptions on the data. Experiments on both real and synthetic datasets verify our theoretical findings and demonstrate the effectiveness of our algorithms. We also show that clipping bias introduced by bounding user contribution may be reduced under mild distribution assumptions, which can be of independent interest.