ICML 2023 Papers

We study a $K$-armed bandit with delayed feedback and intermediate observations. We consider a model, where intermediate observations have a form of a finite state, which is observed immediately after taking an action, whereas the loss is observed after an adversarially chosen delay. We show that the regime of the mapping of states to losses determines the complexity of the problem, irrespective of whether the mapping of actions to states is stochastic or adversarial. If the mapping of states to losses is adversarial, then the regret rate is of order $\sqrt{(K+d)T}$ (within log factors), where $T$ is the time horizon and $d$ is a fixed delay. This matches the regret rate of a $K$-armed bandit with delayed feedback and without intermediate observations, implying that intermediate observations are not helpful. However, if the mapping of states to losses is stochastic, we show that the regret grows at a rate of $\sqrt{\big(K+\min\\{|\mathcal{S}|,d\\}\big)T}$ (within log factors), implying that if the number $|\mathcal{S}|$ of states is smaller than the delay, then intermediate observations help. We also provide refined high-probability regret upper bounds for non-uniform delays, together with experimental validation of our algorithms.

Having reliable specifications is an unavoidable challenge in achieving verifiable correctness, robustness, and interpretability of AI systems. Existing specifications for neural networks are in the paradigm of data as specification. That is, the local neighborhood centering around a reference input is considered to be correct (or robust). While existing specifications contribute to verifying adversarial robustness, a significant problem in many research domains, our empirical study shows that those verified regions are somewhat tight, and thus fail to allow verification of test set inputs, making them impractical for some real-world applications. To this end, we propose a new family of specifications called neural representation as specification. This form of specifications uses the intrinsic information of neural networks, specifically neural activation patterns (NAPs), rather than input data to specify the correctness and/or robustness of neural network predictions. We present a simple statistical approach to mining neural activation patterns. To show the effectiveness of discovered NAPs, we formally verify several important properties, such as various types of misclassifications will never happen for a given NAP, and there is no ambiguity between different NAPs. We show that by using NAP, we can verify a significant region of the input space, while still recalling 84% of the data on MNIST. Moreover, we can push the verifiable bound to 10 times larger on the CIFAR10 benchmark. Thus, we argue that NAPs can potentially be used as a more reliable and extensible specification for neural network verification.

The parameter perturbation attack is a safety threat to deep learning, where small parameter perturbations are made such that the attacked network gives wrong or desired labels of the adversary to specified inputs. However, such attacks could be detected by the user, because the accuracy of the attacked network will reduce and the network cannot work normally. To make the attack more stealthy, in this paper, the adversarial parameter attack is proposed, in which small perturbations to the parameters of the network are made such that the accuracy of the attacked network does not decrease much, but its robustness against adversarial example attacks becomes much lower. As a consequence, the attacked network performs normally on standard samples, but is much more vulnerable to adversarial attacks. The existence of nearly perfect adversarial parameters under $L_\infty$ norm and $L_0$ norm is proved under reasonable conditions. Algorithms are given which can be used to produce high quality adversarial parameters for the commonly used networks trained with various robust training methods, in that the robustness of the attacked networks decreases significantly when they are evaluated using various adversarial attack methods.

We study robustness to test-time adversarial attacks in the regression setting with $\ell_p$ losses and arbitrary perturbation sets. We address the question of which function classes are PAC learnable in this setting. We show that classes of finite fat-shattering dimension are learnable in both the realizable and agnostic settings. Moreover, for convex function classes, they are even properly learnable. In contrast, some non-convex function classes provably require improper learning algorithms. Our main technique is based on a construction of an adversarially robust sample compression scheme of a size determined by the fat-shattering dimension. Along the way, we introduce a novel agnostic sample compression scheme for real-valued functions, which may be of independent interest.

Machine learning models are vulnerable to adversarial perturbations, and a thought-provoking paper by Bubeck and Sellke has analyzed this phenomenon through the lens of over-parameterization: interpolating smoothly the data requires significantly more parameters than simply memorizing it. However, this "universal" law provides only a necessary condition for robustness, and it is unable to discriminate between models. In this paper, we address these gaps by focusing on empirical risk minimization in two prototypical settings, namely, random features and the neural tangent kernel (NTK). We prove that, for random features, the model is not robust for any degree of over-parameterization, even when the necessary condition coming from the universal law of robustness is satisfied. In contrast, for even activations, the NTK model meets the universal lower bound, and it is robust as soon as the necessary condition on over-parameterization is fulfilled. This also addresses a conjecture in prior work by Bubeck, Li and Nagaraj. Our analysis decouples the effect of the kernel of the model from an "interaction matrix", which describes the interaction with the test data and captures the effect of the activation. Our theoretical results are corroborated by numerical evidence on both synthetic and standard datasets (MNIST, CIFAR-10).

This paper studies online nonstochastic control problems with adversarial and static constraints. We propose online nonstochastic control algorithms that achieve both sublinear regret and sublinear adversarial constraint violation while keeping static constraint violation minimal against the optimal constrained linear control policy in hindsight. To establish the results, we introduce an online convex optimization with memory framework under adversarial and static constraints, which serves as a subroutine for the constrained online nonstochastic control algorithms. This subroutine also achieves the state-of-the-art regret and constraint violation bounds for constrained online convex optimization problems, which is of independent interest. Our experiments demonstrate the proposed control algorithms are adaptive to adversarial constraints and achieve smaller cumulative costs and violations. Moreover, our algorithms are less conservative and achieve significantly smaller cumulative costs than the state-of-the-art algorithm.

3D point clouds serve as a crucial data representation in numerous real-world applications such as autonomous driving, robotics, and medical imaging. While the advancements in deep learning have spurred the utilization of 3D point clouds, deep models are notoriously vulnerable to adversarial attacks. Various defense solutions have been proposed to build robust models against adversarial attacks. In this work, we pinpoint a major limitation of the leading empirical defense, adversarial training, when applied to 3D point cloud models: gradient obfuscation, which significantly hampers robustness against potent attacks. To bridge the gap, we propose PointDP, a purification strategy that leverages diffusion models to defend against 3D adversarial attacks. Since PointDP does not rely on predefined adversarial examples for training, it can defend against a variety of threats. We conduct a comprehensive evaluation of PointDP across six representative 3D point cloud architectures, employing sixteen strong and adaptive attacks to manifest its foundational robustness. Our evaluation shows that PointDP achieves significantly better (i.e., 12.6%-40.3%) adversarial robustness than state-of-the-art methods under strong attacks bounded by different $\ell_p$ norms.

Generalizable Neural Radiance Fields (GNeRF) are one of the most promising real-world solutions for novel view synthesis, thanks to their cross-scene generalization capability and thus the possibility of instant rendering on new scenes. While adversarial robustness is essential for real-world applications, little study has been devoted to understanding its implication on GNeRF. We hypothesize that because GNeRF is implemented by conditioning on the source views from new scenes, which are often acquired from the Internet or third-party providers, there are potential new security concerns regarding its real-world applications. Meanwhile, existing understanding and solutions for neural networks' adversarial robustness may not be applicable to GNeRF, due to its 3D nature and uniquely diverse operations. To this end, we present NeRFool, which to the best of our knowledge is the first work that sets out to understand the adversarial robustness of GNeRF. Specifically, NeRFool unveils the vulnerability patterns and important insights regarding GNeRF's adversarial robustness. Built upon the above insights gained from NeRFool, we further develop NeRFool$^+$, which integrates two techniques capable of effectively attacking GNeRF across a wide range of target views, and provide guidelines for defending against our proposed attacks. We believe that our NeRFool/NeRFool$^+$ lays the initial foundation for future innovations in developing robust real-world GNeRF solutions. Our codes are available at: https://github.com/GATECH-EIC/NeRFool.

Recently, Diffusion Models (DMs) boost a wave in AI for Art yet raise new copyright concerns, where infringers benefit from using unauthorized paintings to train DMs and generate novel paintings in a similar style. To address these emerging copyright violations, in this paper, we are the first to explore and propose to utilize adversarial examples for DMs to protect human-created artworks. Specifically, we first build a theoretical framework to define and evaluate the adversarial examples for DMs. Then, based on this framework, we design a novel algorithm to generate these adversarial examples, named AdvDM, which exploits a Monte-Carlo estimation of adversarial examples for DMs by optimizing upon different latent variables sampled from the reverse process of DMs. Extensive experiments show that the generated adversarial examples can effectively hinder DMs from extracting their features. Therefore, our method can be a powerful tool for human artists to protect their copyright against infringers equipped with DM-based AI-for-Art applications. The code of our method is available on GitHub: https://github.com/mist-project/mist.git.

In data poisoning attacks, an adversary tries to change a model's prediction by adding, modifying, or removing samples in the training data. Recently, *ensemble-based* approaches for obtaining *provable* defenses against data poisoning have been proposed where predictions are done by taking a majority vote across multiple base models. In this work, we show that merely considering the majority vote in ensemble defenses is wasteful as it does not effectively utilize available information in the logits layers of the base models. Instead, we propose *Run-Off Election (ROE)*, a novel aggregation method based on a two-round election across the base models: In the first round, models vote for their preferred class and then a second, *Run-Off* election is held between the top two classes in the first round. Based on this approach, we propose DPA+ROE and FA+ROE defense methods based on Deep Partition Aggregation (DPA) and Finite Aggregation (FA) approaches from prior work. We evaluate our methods on MNIST, CIFAR-10, and GTSRB and obtain improvements in certified accuracy by up to $3\%$-$4\%$. Also, by applying ROE on a boosted version of DPA, we gain improvements around $12\%$-$27\%$ comparing to the current state-of-the-art, establishing **a new state-of-the-art** in (pointwise) certified robustness against data poisoning. In many cases, our approach outperforms the state-of-the-art, even when using 32 times less computational power.

A fundamental issue in machine learning is the robustness of the model with respect to changes in the input. In natural language processing, models typically contain a first embedding layer, transforming a sequence of tokens into vector representations. While the robustness with respect to changes of continuous inputs is well-understood, the situation is less clear when considering discrete changes, for instance replacing a word by another in an input sentence. Our work formally proves that popular embedding schemes, such as concatenation, TF-IDF, and Paragraph Vector (a.k.a. doc2vec), exhibit robustness in the Hölder or Lipschitz sense with respect to the Hamming distance. We provide quantitative bounds for these schemes and demonstrate how the constants involved are affected by the length of the document. These findings are exemplified through a series of numerical examples.

This paper proposes a new loss function for adversarial training. Since adversarial training has difficulties, e.g., necessity of high model capacity, focusing on important data points by weighting cross-entropy loss has attracted much attention. However, they are vulnerable to sophisticated attacks, e.g., Auto-Attack. This paper experimentally reveals that the cause of their vulnerability is their small margins between logits for the true label and the other labels. Since neural networks classify the data points based on the logits, logit margins should be large enough to avoid flipping the largest logit by the attacks. Importance-aware methods do not increase logit margins of important samples but decrease those of less-important samples compared with cross-entropy loss. To increase logit margins of important samples, we propose switching one-vs-the-rest loss (SOVR), which switches from cross-entropy to one-vs-the-rest loss for important samples that have small logit margins. We prove that one-vs-the-rest loss increases logit margins two times larger than the weighted cross-entropy loss for a simple problem. We experimentally confirm that SOVR increases logit margins of important samples unlike existing methods and achieves better robustness against Auto-Attack than importance-aware methods.

Adversarial detection aims to determine whether a given sample is an adversarial one based on the discrepancy between natural and adversarial distributions. Unfortunately, estimating or comparing two data distributions is extremely difficult, especially in high-dimension spaces. Recently, the gradient of log probability density (a.k.a., score) w.r.t. the sample is used as an alternative statistic to compute. However, we find that the score is sensitive in identifying adversarial samples due to insufficient information with one sample only. In this paper, we propose a new statistic called expected perturbation score (EPS), which is essentially the expected score of a sample after various perturbations. Specifically, to obtain adequate information regarding one sample, we perturb it by adding various noises to capture its multi-view observations. We theoretically prove that EPS is a proper statistic to compute the discrepancy between two samples under mild conditions. In practice, we can use a pre-trained diffusion model to estimate EPS for each sample. Last, we pro- pose an EPS-based adversarial detection (EPS- AD) method, in which we develop EPS-based maximum mean discrepancy (MMD) as a metric to measure the discrepancy between the test sample and natural samples. We also prove that the EPS-based MMD between natural and adversarial samples is larger than that among natural samples. Extensive experiments show the superior adversarial detection performance of our EPS-AD.

Bayesian inference usually requires running potentially costly inference procedures separately for every new observation. In contrast, the idea of amortized Bayesian inference is to initially invest computational cost in training an inference network on simulated data, which can subsequently be used to rapidly perform inference (i.e., to return estimates of posterior distributions) for new observations. This approach has been applied to many real-world models in the sciences and engineering, but it is unclear how robust the approach is to adversarial perturbations in the observed data. Here, we study the adversarial robustness of amortized Bayesian inference, focusing on simulation-based estimation of multi-dimensional posterior distributions. We show that almost unrecognizable, targeted perturbations of the observations can lead to drastic changes in the predicted posterior and highly unrealistic posterior predictive samples, across several benchmark tasks and a real-world example from neuroscience. We propose a computationally efficient regularization scheme based on penalizing the Fisher information of the conditional density estimator, and show how it improves the adversarial robustness of amortized Bayesian inference.

Many problems, such as online ad display, can be formulated as online bipartite matching. The crucial challenge lies in the nature of sequentially-revealed online item information, based on which we make irreversible matching decisions at each step. While numerous expert online algorithms have been proposed with bounded worst-case competitive ratios, they may not offer satisfactory performance in average cases. On the other hand, reinforcement learning (RL) has been applied to improve the average performance, but it lacks robustness and can perform arbitrarily poorly. In this paper, we propose a novel RL-based approach to edge-weighted online bipartite matching with robustness guarantees (LOMAR), achieving both good average-case and worst-case performance. The key novelty of LOMAR is a new online switching operation which, based on a judicious condition to hedge against future uncertainties, decides whether to follow the expert's decision or the RL decision for each online item. We prove that for any $\rho\in[0,1]$, LOMAR is $\rho$-competitive against any given expert online algorithm. To improve the average performance, we train the RL policy by explicitly considering the online switching operation. Finally, we run empirical experiments to demonstrate the advantages of LOMAR compared to existing baselines.

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, there is an emerging interest in adversarially training a classifier with a rejection option (also known as a selective classifier) for boosting adversarial robustness. While rejection can incur a cost in many applications, existing studies typically associate zero cost with rejecting perturbed inputs, which can result in the rejection of numerous slightly-perturbed inputs that could be correctly classified. In this work, we study adversarially-robust classification with rejection in the stratified rejection setting, where the rejection cost is modeled by rejection loss functions monotonically non-increasing in the perturbation magnitude. We theoretically analyze the stratified rejection setting and propose a novel defense method -- Adversarial Training with Consistent Prediction-based Rejection (CPR) -- for building a robust selective classifier. Experiments on image datasets demonstrate that the proposed method significantly outperforms existing methods under strong adaptive attacks. For instance, on CIFAR-10, CPR reduces the total robust loss (for different rejection losses) by at least 7.3% under both seen and unseen attacks.

We attack the state-of-the-art Go-playing AI system KataGo by training adversarial policies against it, achieving a >97% win rate against KataGo running at superhuman settings. Our adversaries do not win by playing Go well. Instead, they trick KataGo into making serious blunders. Our attack transfers zero-shot to other superhuman Go-playing AIs, and is comprehensible to the extent that human experts can implement it without algorithmic assistance to consistently beat superhuman AIs. The core vulnerability uncovered by our attack persists even in KataGo agents adversarially trained to defend against our attack. Our results demonstrate that even superhuman AI systems may harbor surprising failure modes. Example games are available https://goattack.far.ai/.

Deep networks for computer vision are not reliable when they encounter adversarial examples. In this paper, we introduce a framework that uses the dense intrinsic constraints in natural images to robustify inference. By introducing constraints at inference time, we can shift the burden of robustness from training to testing, thereby allowing the model to dynamically adjust to each individual image's unique and potentially novel characteristics at inference time. Our theoretical results show the importance of having dense constraints at inference time. In contrast to existing single-constraint methods, we propose to use equivariance, which naturally allows dense constraints at a fine-grained level in the feature space. Our empirical experiments show that restoring feature equivariance at inference time defends against worst-case adversarial perturbations. The method obtains improved adversarial robustness on four datasets (ImageNet, Cityscapes, PASCAL VOC, and MS-COCO) on image recognition, semantic segmentation, and instance segmentation tasks.