NeurIPS 2023 Papers

Deep neural networks have been found to be vulnerable in a variety of tasks. Adversarial attacks can manipulate network outputs, resulting in incorrect predictions. Adversarial defense methods aim to improve the adversarial robustness of networks by countering potential attacks. In addition to traditional defense approaches, randomized defense mechanisms have recently received increasing attention from researchers. These methods introduce different types of perturbations during the inference phase to destabilize adversarial attacks. Although promising empirical results have been demonstrated by these approaches, the defense performance is quite sensitive to the randomness parameters, which are always manually tuned without further analysis. On the contrary, we propose incorporating random weights into the optimization to fully exploit the potential of randomized defense. To perform better optimization of randomness parameters, we conduct a theoretical analysis of the connections between randomness parameters and gradient similarity as well as natural performance. From these two aspects, we suggest imposing theoretically-guided constraints on random weights during optimizations, as these weights play a critical role in balancing natural performance and adversarial robustness. We derive both the upper and lower bounds of random weight parameters by considering prediction bias and gradient similarity. In this study, we introduce the Constrained Trainable Random Weight (CTRW), which adds random weight parameters to the optimization and includes a constraint guided by the upper and lower bounds to achieve better trade-offs between natural and robust accuracy. We evaluate the effectiveness of CTRW on several datasets and benchmark convolutional neural networks. Our results indicate that our model achieves a robust accuracy approximately 16% to 17% higher than the baseline model under PGD-20 and 22% to 25% higher on Auto Attack.

Despite the promise of Lipschitz-based methods for provably-robust deep learning with deterministic guarantees, current state-of-the-art results are limited to feed-forward Convolutional Networks (ConvNets) on low-dimensional data, such as CIFAR-10. This paper investigates strategies for expanding certifiably robust training to larger, deeper models. A key challenge in certifying deep networks is efficient calculation of the Lipschitz bound for residual blocks found in ResNet and ViT architectures. We show that fast ways of bounding the Lipschitz constant for conventional ResNets are loose, and show how to address this by designing a new residual block, leading to the *Linear ResNet* (LiResNet) architecture. We then introduce *Efficient Margin MAximization* (EMMA), a loss function that stabilizes robust training by penalizing worst-case adversarial examples from multiple classes simultaneously. Together, these contributions yield new *state-of-the-art* robust accuracy on CIFAR-10/100 and Tiny-ImageNet under $\ell_2$ perturbations. Moreover, for the first time, we are able to scale up fast deterministic robustness guarantees to ImageNet, demonstrating that this approach to robust learning can be applied to real-world applications.

Gradual Domain Adaptation (GDA), in which the learner is provided with additional intermediate domains, has been theoretically and empirically studied in many contexts. Despite its vital role in security-critical scenarios, the adversarial robustness of the GDA model remains unexplored. In this paper, we adopt the effective gradual self-training method and replace vanilla self-training with adversarial self-training (AST). AST first predicts labels on the unlabeled data and then adversarially trains the model on the pseudo-labeled distribution. Intriguingly, we find that gradual AST improves not only adversarial accuracy but also clean accuracy on the target domain. We reveal that this is because adversarial training (AT) performs better than standard training when the pseudo-labels contain a portion of incorrect labels. Accordingly, we first present the generalization error bounds for gradual AST in a multiclass classification setting. We then use the optimal value of the Subset Sum Problem to bridge the standard error on a real distribution and the adversarial error on a pseudo-labeled distribution. The result indicates that AT may obtain a tighter bound than standard training on data with incorrect pseudo-labels. We further present an example of a conditional Gaussian distribution to provide more insights into why gradual AST can improve the clean accuracy for GDA.

Diffusion-based purification defenses leverage diffusion models to remove crafted perturbations of adversarial examples and achieve state-of-the-art robustness. Recent studies show that even advanced attacks cannot break such defenses effectively, since the purification process induces an extremely deep computational graph which poses the potential problem of gradient obfuscation, high memory cost, and unbounded randomness. In this paper, we propose a unified framework DiffAttack to perform effective and efficient attacks against diffusion-based purification defenses, including both DDPM and score-based approaches. In particular, we propose a deviated-reconstruction loss at intermediate diffusion steps to induce inaccurate density gradient estimation to tackle the problem of vanishing/exploding gradients. We also provide a segment-wise forwarding-backwarding algorithm, which leads to memory-efficient gradient backpropagation. We validate the attack effectiveness of DiffAttack compared with existing adaptive attacks on CIFAR-10 and ImageNet. We show that DiffAttack decreases the robust accuracy of models compared with SOTA attacks by over 20\% on CIFAR-10 under $\ell_\infty$ attack $(\epsilon=8/255)$, and over 10\% on ImageNet under $\ell_\infty$ attack $(\epsilon=4/255)$. We conduct a series of ablations studies, and we find 1) DiffAttack with the deviated-reconstruction loss added over uniformly sampled time steps is more effective than that added over only initial/final steps, and 2) diffusion-based purification with a moderate diffusion length is more robust under DiffAttack.

Densely structured pruning methods utilizing simple pruning heuristics can deliver immediate compression and acceleration benefits with acceptable benign performances. However, empirical findings indicate such naively pruned networks are extremely fragile under simple adversarial attacks. Naturally, we would be interested in knowing if such a phenomenon also holds for carefully designed modern structured pruning methods. If so, then to what extent is the severity? And what kind of remedies are available? Unfortunately, both the questions and the solution remain largely unaddressed: no prior art is able to provide a thorough investigation on the adversarial performance of modern structured pruning methods (spoiler: it is not good), yet the few works that attempt to provide mitigation often do so at various extra costs with only to-be-desired performance. In this work, we answer both questions by fairly and comprehensively investigating the adversarial performance of 10+ popular structured pruning methods. Solution-wise, we take advantage of *Grouped Kernel Pruning (GKP)*'s recent success in pushing densely structured pruning freedom to a more fine-grained level. By mixing up kernel smoothness — a classic robustness-related kernel-level metric — into a modified GKP procedure, we present a one-shot-post-train-weight-dependent GKP method capable of advancing SOTA performance on both the benign and adversarial scale, while requiring no extra (in fact, often less) cost than a standard pruning procedure. Please refer to our [GitHub repository](https://github.com/henryzhongsc/adv_robust_gkp) for code implementation, tool sharing, and model checkpoints.

Data poisoning considers cases when an adversary manipulates the behavior of machine learning algorithms through malicious training data. Existing threat models of data poisoning center around a single metric, the number of poisoned samples. In consequence, if attackers can poison more samples than expected with affordable overhead, as in many practical scenarios, they may be able to render existing defenses ineffective in a short time. To address this issue, we leverage timestamps denoting the birth dates of data, which are often available but neglected in the past. Benefiting from these timestamps, we propose a temporal threat model of data poisoning with two novel metrics, earliness and duration, which respectively measure how long an attack started in advance and how long an attack lasted. Using these metrics, we define the notions of temporal robustness against data poisoning, providing a meaningful sense of protection even with unbounded amounts of poisoned samples when the attacks are temporally bounded. We present a benchmark with an evaluation protocol simulating continuous data collection and periodic deployments of updated models, thus enabling empirical evaluation of temporal robustness. Lastly, we develop and also empirically verify a baseline defense, namely temporal aggregation, offering provable temporal robustness and highlighting the potential of our temporal threat model for data poisoning.

Transfer-based attacks generate adversarial examples on the surrogate model, which can mislead other black-box models without access, making it promising to attack real-world applications. Recently, several works have been proposed to boost adversarial transferability, in which the surrogate model is usually overlooked. In this work, we identify that non-linear layers (e.g., ReLU, max-pooling, etc.) truncate the gradient during backward propagation, making the gradient w.r.t. input image imprecise to the loss function. We hypothesize and empirically validate that such truncation undermines the transferability of adversarial examples. Based on these findings, we propose a novel method called Backward Propagation Attack (BPA) to increase the relevance between the gradient w.r.t. input image and loss function so as to generate adversarial examples with higher transferability. Specifically, BPA adopts a non-monotonic function as the derivative of ReLU and incorporates softmax with temperature to smooth the derivative of max-pooling, thereby mitigating the information loss during the backward propagation of gradients. Empirical results on the ImageNet dataset demonstrate that not only does our method substantially boost the adversarial transferability, but it is also general to existing transfer-based attacks. Code is available at https://github.com/Trustworthy-AI-Group/RPA.

Single-step adversarial training (SSAT) has demonstrated the potential to achieve both efficiency and robustness. However, SSAT suffers from catastrophic overfitting (CO), a phenomenon that leads to a severely distorted classifier, making it vulnerable to multi-step adversarial attacks. In this work, we observe that some adversarial examples generated on the SSAT-trained network exhibit anomalous behaviour, that is, although these training samples are generated by the inner maximization process, their associated loss decreases instead, which we named abnormal adversarial examples (AAEs). Upon further analysis, we discover a close relationship between AAEs and classifier distortion, as both the number and outputs of AAEs undergo a significant variation with the onset of CO. Given this observation, we re-examine the SSAT process and uncover that before the occurrence of CO, the classifier already displayed a slight distortion, indicated by the presence of few AAEs. Furthermore, the classifier directly optimizing these AAEs will accelerate its distortion, and correspondingly, the variation of AAEs will sharply increase as a result. In such a vicious circle, the classifier rapidly becomes highly distorted and manifests as CO within a few iterations. These observations motivate us to eliminate CO by hindering the generation of AAEs. Specifically, we design a novel method, termed Abnormal Adversarial Examples Regularization (AAER), which explicitly regularizes the variation of AAEs to hinder the classifier from becoming distorted. Extensive experiments demonstrate that our method can effectively eliminate CO and further boost adversarial robustness with negligible additional computational overhead. Our implementation can be found at https://github.com/tmllab/2023_NeurIPS_AAER.

With the growth of model and data sizes, a broad effort has been made to design pruning techniques that reduce the resource demand of deep learning pipelines, while retaining model performance. In order to reduce both inference and training costs, a prominent line of work uses low-rank matrix factorizations to represent the network weights. Although able to retain accuracy, we observe that low-rank methods tend to compromise model robustness against adversarial perturbations. By modeling robustness in terms of the condition number of the neural network, we argue that this loss of robustness is due to the exploding singular values of the low-rank weight matrices. Thus, we introduce a robust low-rank training algorithm that maintains the network's weights on the low-rank matrix manifold while simultaneously enforcing approximate orthonormal constraints. The resulting model reduces both training and inference costs while ensuring well-conditioning and thus better adversarial robustness, without compromising model accuracy. This is shown by extensive numerical evidence and by our main approximation theorem that shows the computed robust low-rank network well-approximates the ideal full model, provided a highly performing low-rank sub-network exists.

The strategy of ensemble has become popular in adversarial defense, which trains multiple base classifiers to defend against adversarial attacks in a cooperative manner. Despite the empirical success, theoretical explanations on why an ensemble of adversarially trained classifiers is more robust than single ones remain unclear. To fill in this gap, we develop a new error theory dedicated to understanding ensemble adversarial defense, demonstrating a provable 0-1 loss reduction on challenging sample sets in adversarial defense scenarios. Guided by this theory, we propose an effective approach to improve ensemble adversarial defense, named interactive global adversarial training (iGAT). The proposal includes (1) a probabilistic distributing rule that selectively allocates to different base classifiers adversarial examples that are globally challenging to the ensemble, and (2) a regularization term to rescue the severest weaknesses of the base classifiers. Being tested over various existing ensemble adversarial defense techniques, iGAT is capable of boosting their performance by up to 17\% evaluated using CIFAR10 and CIFAR100 datasets under both white-box and black-box attacks.

Adversarial attacks have the potential to mislead deep neural network classifiers by introducing slight perturbations. Developing algorithms that can mitigate the effects of these attacks is crucial for ensuring the safe use of artificial intelligence. Recent studies have suggested that score-based diffusion models are effective in adversarial defenses. However, existing diffusion-based defenses rely on the sequential simulation of the reversed stochastic differential equations of diffusion models, which are computationally inefficient and yield suboptimal results. In this paper, we introduce a novel adversarial defense scheme named ScoreOpt, which optimizes adversarial samples at test-time, towards original clean data in the direction guided by score-based priors. We conduct comprehensive experiments on multiple datasets, including CIFAR10, CIFAR100 and ImageNet. Our experimental results demonstrate that our approach outperforms existing adversarial defenses in terms of both robustness performance and inference speed.

To improve the robustness of deep classifiers against adversarial perturbations, many approaches have been proposed, such as designing new architectures with better robustness properties (e.g., Lipschitz-capped networks), or modifying the training process itself (e.g., min-max optimization, constrained learning, or regularization). These approaches, however, might not be effective at increasing the margin in the input (feature) space. In this paper, we propose a differentiable regularizer that is a lower bound on the distance of the data points to the classification boundary. The proposed regularizer requires knowledge of the model's Lipschitz constant along certain directions. To this end, we develop a scalable method for calculating guaranteed differentiable upper bounds on the Lipschitz constant of neural networks accurately and efficiently. The relative accuracy of the bounds prevents excessive regularization and allows for more direct manipulation of the decision boundary. Furthermore, our Lipschitz bounding algorithm exploits the monotonicity and Lipschitz continuity of the activation layers, and the resulting bounds can be used to design new layers with controllable bounds on their Lipschitz constant. Experiments on the MNIST, CIFAR-10, and Tiny-ImageNet data sets verify that our proposed algorithm obtains competitively improved results compared to the state-of-the-art.

What spatial frequency information do humans and neural networks use to recognize objects? In neuroscience, critical band masking is an established tool that can reveal the frequency-selective filters used for object recognition. Critical band masking measures the sensitivity of recognition performance to noise added at each spatial frequency. Existing critical band masking studies show that humans recognize periodic patterns (gratings) and letters by means of a spatial-frequency filter (or "channel") that has a frequency bandwidth of one octave (doubling of frequency). Here, we introduce critical band masking as a task for network-human comparison and test 14 humans and 76 neural networks on 16-way ImageNet categorization in the presence of narrowband noise. We find that humans recognize objects in natural images using the same one-octave-wide channel that they use for letters and gratings, making it a canonical feature of human object recognition. Unlike humans, the neural network channel is very broad, 2-4 times wider than the human channel. This means that the network channel extends to frequencies higher and lower than those that humans are sensitive to. Thus, noise at those frequencies will impair network performance and spare human performance. Adversarial and augmented-image training are commonly used to increase network robustness and shape bias. Does this training align network and human object recognition channels? Three network channel properties (bandwidth, center frequency, peak noise sensitivity) correlate strongly with shape bias (51% variance explained) and robustness of adversarially-trained networks (66% variance explained). Adversarial training increases robustness but expands the channel bandwidth even further beyond the human bandwidth. Thus, critical band masking reveals that the network channel is more than twice as wide as the human channel, and that adversarial training only makes it worse. Networks with narrower channels might be more robust.

Research on adversarial robustness is primarily focused on image and text data. Yet, many scenarios in which lack of robustness can result in serious risks, such as fraud detection, medical diagnosis, or recommender systems often do not rely on images or text but instead on tabular data. Adversarial robustness in tabular data poses two serious challenges. First, tabular datasets often contain categorical features, and therefore cannot be tackled directly with existing optimization procedures. Second, in the tabular domain, algorithms that are not based on deep networks are widely used and offer great performance, but algorithms to enhance robustness are tailored to neural networks (e.g. adversarial training). In this paper, we tackle both challenges. We present a method that allows us to train adversarially robust deep networks for tabular data and to transfer this robustness to other classifiers via universal robust embeddings tailored to categorical data. These embeddings, created using a bilevel alternating minimization framework, can be transferred to boosted trees or random forests making them robust without the need for adversarial training while preserving their high accuracy on tabular data. We show that our methods outperform existing techniques within a practical threat model suitable for tabular data.

Deep neural networks are known to be vulnerable to adversarial attacks (AA). For an image recognition task, this means that a small perturbation of the original can result in the image being misclassified. Design of such attacks as well as methods of adversarial training against them are subject of intense research. We re-cast the problem using techniques of Wasserstein distributionally robust optimization (DRO) and obtain novel contributions leveraging recent insights from DRO sensitivity analysis. We consider a set of distributional threat models. Unlike the traditional pointwise attacks, which assume a uniform bound on perturbation of each input data point, distributional threat models allow attackers to perturb inputs in a non-uniform way. We link these more general attacks with questions of out-of-sample performance and Knightian uncertainty. To evaluate the distributional robustness of neural networks, we propose a first-order AA algorithm and its multistep version. Our attack algorithms include Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) as special cases. Furthermore, we provide a new asymptotic estimate of the adversarial accuracy against distributional threat models. The bound is fast to compute and first-order accurate, offering new insights even for the pointwise AA. It also naturally yields out-of-sample performance guarantees. We conduct numerical experiments on CIFAR-10, CIFAR-100, ImageNet datasets using DNNs on RobustBench to illustrate our theoretical results. Our code is available at https://github.com/JanObloj/W-DRO-Adversarial-Methods.

Neural networks are known to be susceptible to adversarial samples: small variations of natural examples crafted to deliberately mislead the models. While they can be easily generated using gradient-based techniques in digital and physical scenarios, they often differ greatly from the actual data distribution of natural images, resulting in a trade-off between strength and stealthiness. In this paper, we propose a novel framework dubbed Diffusion-Based Projected Gradient Descent (Diff-PGD) for generating realistic adversarial samples. By exploiting a gradient guided by a diffusion model, Diff-PGD ensures that adversarial samples remain close to the original data distribution while maintaining their effectiveness. Moreover, our framework can be easily customized for specific tasks such as digital attacks, physical-world attacks, and style-based attacks. Compared with existing methods for generating natural-style adversarial samples, our framework enables the separation of optimizing adversarial loss from other surrogate losses (e.g. content/smoothness/style loss), making it more stable and controllable. Finally, we demonstrate that the samples generated using Diff-PGD have better transferability and anti-purification power than traditional gradient-based methods.

The susceptibility of modern machine learning classifiers to adversarial examples has motivated theoretical results suggesting that these might be unavoidable. However, these results can be too general to be applicable to natural data distributions. Indeed, humans are quite robust for tasks involving vision. This apparent conflict motivates a deeper dive into the question: Are adversarial examples truly unavoidable? In this work, we theoretically demonstrate that a key property of the data distribution -- concentration on small-volume subsets of the input space -- determines whether a robust classifier exists. We further demonstrate that, for a data distribution concentrated on a union of low-dimensional linear subspaces, utilizing structure in data naturally leads to classifiers that enjoy data-dependent polyhedral robustness guarantees, improving upon methods for provable certification in certain regimes.

The transferability of adversarial perturbations provides an effective shortcut for black-box attacks. Targeted perturbations have greater practicality but are more difficult to transfer between models. In this paper, we experimentally and theoretically demonstrated that neural networks trained on the same dataset have more consistent performance in High-Sample-Density-Regions (HSDR) of each class instead of low sample density regions. Therefore, in the target setting, adding perturbations towards HSDR of the target class is more effective in improving transferability. However, density estimation is challenging in high-dimensional scenarios. Further theoretical and experimental verification demonstrates that easy samples with low loss are more likely to be located in HSDR. Perturbations towards such easy samples in the target class can avoid density estimation for HSDR location. Based on the above facts, we verified that adding perturbations to easy samples in the target class improves targeted adversarial transferability of existing attack methods. A generative targeted attack strategy named Easy Sample Matching Attack (ESMA) is proposed, which has a higher success rate for targeted attacks and outperforms the SOTA generative method. Moreover, ESMA requires only $5\%$ of the storage space and much less computation time comparing to the current SOTA, as ESMA attacks all classes with only one model instead of seperate models for each class. Our code is available at https://github.com/gjq100/ESMA

Out-of-distribution (OOD) generalization has attracted increasing research attention in recent years, due to its promising experimental results in real-world applications. Interestingly, we find that existing OOD generalization methods are vulnerable to adversarial attacks. This motivates us to study OOD adversarial robustness. We first present theoretical analyses of OOD adversarial robustness in two different complementary settings. Motivated by the theoretical results, we design two algorithms to improve the OOD adversarial robustness. Finally, we conduct experiments to validate the effectiveness of our proposed algorithms.