NeurIPS 2023 Papers

Mixup refers to interpolation-based data augmentation, originally motivated as a way to go beyond empirical risk minimization (ERM). Its extensions mostly focus on the definition of interpolation and the space (input or feature) where it takes place, while the augmentation process itself is less studied. In most methods, the number of generated examples is limited to the mini-batch size and the number of examples being interpolated is limited to two (pairs), in the input space. We make progress in this direction by introducing MultiMix, which generates an arbitrarily large number of interpolated examples beyond the mini-batch size and interpolates the entire mini-batch in the embedding space. Effectively, we sample on the entire convex hull of the mini-batch rather than along linear segments between pairs of examples. On sequence data, we further extend to Dense MultiMix. We densely interpolate features and target labels at each spatial location and also apply the loss densely. To mitigate the lack of dense labels, we inherit labels from examples and weight interpolation factors by attention as a measure of confidence. Overall, we increase the number of loss terms per mini-batch by orders of magnitude at little additional cost. This is only possible because of interpolating in the embedding space. We empirically show that our solutions yield significant improvement over state-of-the-art mixup methods on four different benchmarks, despite interpolation being only linear. By analyzing the embedding space, we show that the classes are more tightly clustered and uniformly spread over the embedding space, thereby explaining the improved behavior.

We examine the relationship between learnability and robust learnability for the problem of distribution learning. We show that learnability implies robust learnability if the adversary can only perform additive contamination (and consequently, under Huber contamination), but not if the adversary is allowed to perform subtractive contamination. Thus, contrary to other learning settings (e.g., PAC learning of function classes), realizable learnability does not imply agnostic learnability. We also explore related implications in the context of compression schemes and differentially private learnability.

As large-scale training regimes have gained popularity, the use of pretrained models for downstream tasks has become common practice in machine learning. While pretraining has been shown to enhance the performance of models in practice, the transfer of robustness properties from pretraining to downstream tasks remains poorly understood. In this study, we demonstrate that the robustness of a linear predictor on downstream tasks can be constrained by the robustness of its underlying representation, regardless of the protocol used for pretraining. We prove (i) a bound on the loss that holds independent of any downstream task, as well as (ii) a criterion for robust classification in particular. We validate our theoretical results in practical applications, show how our results can be used for calibrating expectations of downstream robustness, and when our results are useful for optimal transfer learning. Taken together, our results offer an initial step towards characterizing the requirements of the representation function for reliable post-adaptation performance.

Distributional robustness is a promising framework for training deep learning models that are less vulnerable to adversarial examples and data distribution shifts. Previous works have mainly focused on exploiting distributional robustness in the data space. In this work, we explore an optimal transport-based distributional robustness framework in model spaces. Specifically, we examine a model distribution within a Wasserstein ball centered on a given model distribution that maximizes the loss. We have developed theories that enable us to learn the optimal robust center model distribution. Interestingly, our developed theories allow us to flexibly incorporate the concept of sharpness awareness into training, whether it's a single model, ensemble models, or Bayesian Neural Networks, by considering specific forms of the center model distribution. These forms include a Dirac delta distribution over a single model, a uniform distribution over several models, and a general Bayesian Neural Network. Furthermore, we demonstrate that Sharpness-Aware Minimization (SAM) is a specific case of our framework when using a Dirac delta distribution over a single model, while our framework can be seen as a probabilistic extension of SAM. To validate the effectiveness of our framework in the aforementioned settings, we conducted extensive experiments, and the results reveal remarkable improvements compared to the baselines.

Automated content filtering and moderation is an important tool that allows online platforms to build striving user communities that facilitate cooperation and prevent abuse. Unfortunately, resourceful actors try to bypass automated filters in a bid to post content that violate platform policies and codes of conduct. To reach this goal, these malicious actors may obfuscate policy violating images (e.g., overlay harmful images by carefully selected benign images or visual patterns) to prevent machine learning models from reaching the correct decision. In this paper, we invite researchers to tackle this specific issue and present a new image benchmark. This benchmark, based on ImageNet, simulates the type of obfuscations created by malicious actors. It goes beyond Image-Net-C and ImageNet-C-bar by proposing general, drastic, adversarial modifications that preserve the original content intent. It aims to tackle a more common adversarial threat than the one considered by lp-norm bounded adversaries. We evaluate 33 pretrained models on the benchmark and train models with different augmentations, architectures and training methods on subsets of the obfuscations to measure generalization. Our hope is that this benchmark will encourage researchers to test their models and methods and try to find new approaches that are more robust to these obfuscations.

Transformer-based architectures have recently exhibited remarkable successes across different domains beyond just powering large language models. However, existing approaches typically focus on predictive accuracy and computational cost, largely ignoring certain other practical issues such as robustness to contaminated samples. In this paper, by re-interpreting the self-attention mechanism as a non-parametric kernel density estimator, we adapt classical robust kernel density estimation methods to develop novel classes of transformers that are resistant to adversarial attacks and data contamination. We first propose methods that down-weight outliers in RKHS when computing the self-attention operations. We empirically show that these methods produce improved performance over existing state-of-the-art methods, particularly on image data under adversarial attacks. Then we leverage the median-of-means principle to obtain another efficient approach that results in noticeably enhanced performance and robustness on language modeling and time series classification tasks. Our methods can be combined with existing transformers to augment their robust properties, thus promising to impact a wide variety of applications.

Deep neural networks (DNNs) have demonstrated vulnerabilities to adversarial examples, which raises concerns about their reliability in safety-critical applications. While the majority of existing methods generate adversarial examples by making small modifications to the entire image, recent research has proposed a practical alternative known as adversarial patches. Adversarial patches have shown to be highly effective in causing DNNs to misclassify by distorting a localized area (patch) of the image. However, existing methods often produce clearly visible distortions since they do not consider the visibility of the patch. To address this, we propose a novel method for constructing adversarial patches that approximates the appearance of the area it covers. We achieve this by using a set of semi-transparent, RGB-valued circles, drawing inspiration from the computational art community. We utilize an evolutionary strategy to optimize the properties of each shape, and employ a simulated annealing approach to optimize the patch's location. Our approach achieves better or comparable performance to state-of-the-art methods on ImageNet DNN classifiers while achieving a lower $l_2$ distance from the original image. By minimizing the visibility of the patch, this work further highlights the vulnerabilities of DNNs to adversarial patches.

Transfer-based attack adopts the adversarial examples generated on the surrogate model to attack various models, making it applicable in the physical world and attracting increasing interest. Recently, various adversarial attacks have emerged to boost adversarial transferability from different perspectives. In this work, inspired by the observation that flat local minima are correlated with good generalization, we assume and empirically validate that adversarial examples at a flat local region tend to have good transferability by introducing a penalized gradient norm to the original loss function. Since directly optimizing the gradient regularization norm is computationally expensive and intractable for generating adversarial examples, we propose an approximation optimization method to simplify the gradient update of the objective function. Specifically, we randomly sample an example and adopt a first-order procedure to approximate the curvature of the second-order Hessian matrix, which makes computing more efficient by interpolating two Jacobian matrices. Meanwhile, in order to obtain a more stable gradient direction, we randomly sample multiple examples and average the gradients of these examples to reduce the variance due to random sampling during the iterative process. Extensive experimental results on the ImageNet-compatible dataset show that the proposed method can generate adversarial examples at flat local regions, and significantly improve the adversarial transferability on either normally trained models or adversarially trained models than the state-of-the-art attacks. Our codes are available at: https://github.com/Trustworthy-AI-Group/PGN.

Adversarial training has become the de-facto standard method for improving the robustness of models against adversarial examples. However, robust overfitting remains a significant challenge, leading to a large gap between the robustness on the training and test datasets. To understand and improve robust generalization, various measures have been developed, including margin, smoothness, and flatness-based measures. In this study, we present a large-scale analysis of robust generalization to empirically verify whether the relationship between these measures and robust generalization remains valid in diverse settings. We demonstrate when and how these measures effectively capture the robust generalization gap by comparing over 1,300 models trained on CIFAR-10 under the $L_\infty$ norm and further validate our findings through an evaluation of more than 100 models from RobustBench across CIFAR-10, CIFAR-100, and ImageNet. We hope this work can help the community better understand adversarial robustness and motivate the development of more robust defense methods against adversarial attacks.

We study reinforcement learning in episodic inhomogeneous MDPs with adversarial full-information rewards and the unknown transition kernel. We consider the linear mixture MDPs whose transition kernel is a linear mixture model and choose the \emph{dynamic regret} as the performance measure. Denote by $d$ the dimension of the feature mapping, $H$ the horizon, $K$ the number of episodes, $P_T$ the non-stationary measure, we propose a novel algorithm that enjoys an $\widetilde{\mathcal{O}}\big(\sqrt{d^2 H^3K} + \sqrt{H^4(K+P_T)(1+P_T)}\big)$ dynamic regret under the condition that $P_T$ is known, which improves previously best-known dynamic regret for adversarial linear mixture MDP and adversarial tabular MDPs. We also establish an $\Omega\big(\sqrt{d^2 H^3 K} + \sqrt{H K (H+P_T)}\big)$ lower bound, indicating our algorithm is \emph{optimal} in $K$ and $P_T$. Furthermore, when the non-stationary measure $P_T$ is unknown, we design an online ensemble algorithm with a meta-base structure, which is proved to achieve an $\widetilde{\mathcal{O}}\big(\sqrt{d^2 H^3K} + \sqrt{H^4(K+P_T)(1+P_T) + H^2 S_T^2}\big)$ dynamic regret and here $S_T$ is the expected switching number of the best base-learner. The result can be optimal under certain regimes.

Lipschitz bandit is a variant of stochastic bandits that deals with a continuous arm set defined on a metric space, where the reward function is subject to a Lipschitz constraint. In this paper, we introduce a new problem of Lipschitz bandits in the presence of adversarial corruptions where an adaptive adversary corrupts the stochastic rewards up to a total budget $C$. The budget is measured by the sum of corruption levels across the time horizon $T$. We consider both weak and strong adversaries, where the weak adversary is unaware of the current action before the attack, while the strong one can observe it. Our work presents the first line of robust Lipschitz bandit algorithms that can achieve sub-linear regret under both types of adversary, even when the total budget of corruption $C$ is unrevealed to the agent. We provide a lower bound under each type of adversary, and show that our algorithm is optimal under the strong case. Finally, we conduct experiments to illustrate the effectiveness of our algorithms against two classic kinds of attacks.

Real-world data is complex and often consists of objects that can be decomposed into multiple entities (e.g. images into pixels, graphs into interconnected nodes). Randomized smoothing is a powerful framework for making models provably robust against small changes to their inputs - by guaranteeing robustness of the majority vote when randomly adding noise before classification. Yet, certifying robustness on such complex data via randomized smoothing is challenging when adversaries do not arbitrarily perturb entire objects (e.g. images) but only a subset of their entities (e.g. pixels). As a solution, we introduce hierarchical randomized smoothing: We partially smooth objects by adding random noise only on a randomly selected subset of their entities. By adding noise in a more targeted manner than existing methods we obtain stronger robustness guarantees while maintaining high accuracy. We initialize hierarchical smoothing using different noising distributions, yielding novel robustness certificates for discrete and continuous domains. We experimentally demonstrate the importance of hierarchical smoothing in image and node classification, where it yields superior robustness-accuracy trade-offs. Overall, hierarchical smoothing is an important contribution towards models that are both - certifiably robust to perturbations and accurate.

The visual object category reports of artificial neural networks (ANNs) are notoriously sensitive to tiny, adversarial image perturbations. Because human category reports (aka human percepts) are thought to be insensitive to those same small-norm perturbations -- and locally stable in general -- this argues that ANNs are incomplete scientific models of human visual perception. Consistent with this, we show that when small-norm image perturbations are generated by standard ANN models, human object category percepts are indeed highly stable. However, in this very same "human-presumed-stable" regime, we find that robustified ANNs reliably discover low-norm image perturbations that strongly disrupt human percepts. These previously undetectable human perceptual disruptions are massive in amplitude, approaching the same level of sensitivity seen in robustified ANNs. Further, we show that robustified ANNs support precise perceptual state interventions: they guide the construction of low-norm image perturbations that strongly alter human category percepts toward specific prescribed percepts. In sum, these contemporary models of biological visual processing are now accurate enough to guide strong and precise interventions on human perception.

Recently, Montasser at al. (2019) showed that finite VC dimension is not sufficient for proper adversarially robust PAC learning. In light of this hardness, there is a growing effort to study what type of relaxations to the adversarially robust PAC learning setup can enable proper learnability. In this work, we initiate the study of proper learning under relaxations of the worst-case robust loss. We give a family of robust loss relaxations under which VC classes are properly PAC learnable with sample complexity close to what one would require in the standard PAC learning setup. On the other hand, we show that for an existing and natural relaxation of the worst-case robust loss, finite VC dimension is not sufficient for proper learning. Lastly, we give new generalization guarantees for the adversarially robust empirical risk minimizer.

Deep neural networks (DNNs) have been shown to be vulnerable to adversarial attacks -- subtle, perceptually indistinguishable perturbations of inputs that change the response of the model. In the context of vision, we hypothesize that an important contributor to the robustness of human visual perception is constant exposure to low-fidelity visual stimuli in our peripheral vision. To investigate this hypothesis, we develop RBlur, an image transform that simulates the loss in fidelity of peripheral vision by blurring the image and reducing its color saturation based on the distance from a given fixation point. We show that compared to DNNs trained on the original images, DNNs trained on images transformed by RBlur are substantially more robust to adversarial attacks, as well as other, non-adversarial, corruptions, achieving up to 25% higher accuracy on perturbed data.

Finding classifiers robust to adversarial examples is critical for their safe deployment. Determining the robustness of the best possible classifier under a given threat model for a fixed data distribution and comparing it to that achieved by state-of-the-art training methods is thus an important diagnostic tool. In this paper, we find achievable information-theoretic lower bounds on robust loss in the presence of a test-time attacker for *multi-class classifiers on any discrete dataset*. We provide a general framework for finding the optimal $0-1$ loss that revolves around the construction of a conflict hypergraph from the data and adversarial constraints. The prohibitive cost of this formulation in practice leads us to formulate other variants of the attacker-classifier game that more efficiently determine the range of the optimal loss. Our valuation shows, for the first time, an analysis of the gap to optimal robustness for classifiers in the multi-class setting on benchmark datasets.

Deep neural networks (DNNs) are vulnerable to adversarial attacks, which lead to incorrect predictions. In black-box settings, transfer attacks can be conveniently used to generate adversarial examples. However, such examples tend to overfit the specific architecture and feature representations of the source model, resulting in poor attack performance against other target models. To overcome this drawback, we propose a novel model modification-based transfer attack: Blurred-Dilated method (BD) in this paper. In summary, BD works by reducing downsampling while introducing BlurPool and dilated convolutions in the source model. Then BD employs the modified source model to generate adversarial samples. We think that BD can more comprehensively preserve the feature information than the original source model. It thus enables more thorough destruction of the image features, which can improve the transferability of the generated adversarial samples. Extensive experiments on the ImageNet dataset show that adversarial examples generated by BD achieve significantly higher transferability than the state-of-the-art baselines. Besides, BD can be conveniently combined with existing black-box attack techniques to further improve their performance.

We study the consistency of surrogate risks for robust binary classification. It is common to learn robust classifiers by adversarial training, which seeks to minimize the expected $0$-$1$ loss when each example can be maliciously corrupted within a small ball. We give a simple and complete characterization of the set of surrogate loss functions that are \emph{consistent}, i.e., that can replace the $0$-$1$ loss without affecting the minimizing sequences of the original adversarial risk, for any data distribution. We also prove a quantitative version of adversarial consistency for the $\rho$-margin loss. Our results reveal that the class of adversarially consistent surrogates is substantially smaller than in the standard setting, where many common surrogates are known to be consistent.

We study robust adversarial training of two-layer neural networks as a bi-level optimization problem. In particular, for the inner loop that implements the adversarial attack during training using projected gradient descent (PGD), we propose maximizing a \emph{lower bound} on the $0/1$-loss by reflecting a surrogate loss about the origin. This allows us to give a convergence guarantee for the inner-loop PGD attack. Furthermore, assuming the data is linearly separable, we provide precise iteration complexity results for end-to-end adversarial training, which holds for any width and initialization. We provide empirical evidence to support our theoretical results.