Wasserstein distributional robustness of neural networks

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 multi-step 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 the CIFAR-10 dataset using DNNs on RobustBench to illustrate our theoretical results. Our code is available at https://github.com/JanObloj/W-DRO-Adversarial-Methods.

Paper

References (61)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 1Lzk7/10 · confidence 4/52023-07-04

Summary

This paper presents a novel adversarial attack methodology premised on Wasserstein distributionally robust optimization. This approach serves as a generalized technique that encompasses many established adversarial attack methods such as FGSM and TRADES. To ensure computational tractability, the authors introduce a first-order approximation of the attack loss. This approximation is particularly suited for gradient-based optimization methodologies, offering a much faster computational speed compared to SOTA techniques such as AutoAttack. Theoretical evidence is provided for certified bounds on adversarial accuracy and out-of-sample performance. Experiments on the CIFAR-10 dataset are conducted to demonstrate the effectiveness of the proposed method.

Strengths

+ The proposed ReDLR loss (10) is both simple and effective. It is built on clear intuition, that the attack should perturb images classified far from the decision boundary more aggressively while leaving misclassified images unchanged. The theoretical analysis and guarantees provided further validate the proposed approach. + The method is presented using robust proofs and in-depth analysis. I didn't notice any substantial unjustifiable assumptions or logical inaccuracies. The section on "Bounds on Out-of-Sample Performance" is particularly intriguing. Despite substantial parts of the proof being left in the appendix due to page limits, the analysis still provides invaluable insights for future research on empirical distributions of test sets and broader applications of Wasserstein balls. + The empirical results demonstrate notable improvements over robust baseline methods like DLR. It is worth noting that while the proposed ReDLR method may appear very similar to DLR at first glance, the empirical results show significant improvements, attesting to the effectiveness of the theoretical analysis. + The paper is well-structured and clearly written. Even the first part of the method section could serve as a comprehensive introduction to DLR-based methods, making it approachable and comprehensible even for novices in the field.

Weaknesses

- One weakness of the paper is that the number of experiments provided is limited (only results on CIFAR-10 with a limited number of baselines compared). However, considering the major contribution of the paper is to provide a comprehensive theoretical analysis and the existing results already demonstrate the effectiveness of the method, I would say the weakness of experiments does not affect the overall quality of the paper.

Questions

- I personally found the intuition stated in line 171-173 is quite important to understand the method and should be further stressed, both in the introduction and the earlier subsections of the method. - S^c seems to first appeare in line 215, but no clear definition is provided. - How strong is Assumption 4.1? Though intuitively for classification problems it should be held, it would be great if the authors provide some experimental analysis (with several commonly used NNs), or pinpoint some references.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

See Weakness and Questions.

Reviewer Q9Xf8/10 · confidence 3/52023-07-06

Summary

The paper focuses on the issue of distributional robustness in deep neural networks and introduces a novel approach to address it. The authors present a tractable distributional attack within a Wasserstein ball, which allows attackers to perturb inputs in a non-uniform manner. They propose a first-order approximation of the problem and introduce a new loss function called ReDLR. This framework encompasses the Fast Gradient Sign Method (FGSM) attack as a special case. The authors also provide tractable bounds for adversarial accuracy against distributional threat models, which offer insights even for pointwise attacks. To validate their approach, experiments are conducted on the CIFAR-10 dataset using state-of-the-art deep neural networks on RobustBench, demonstrating the effectiveness of the attack and the tightness of the bounds.

Strengths

The strengths of the paper are: -The introduction of a tractable attack for Wasserstein Distributionally Robust Optimization (WDRO) stands as a significant contribution, as does the adversarial accuracy boundary. Both offer key advancements to the current understanding and applications of deep learning models under adversarial conditions. -The paper is underpinned by robust mathematical foundations, and its place in relation to the existing body of work in the field is clearly and precisely articulated, demonstrating a deep and accurate understanding of the state-of-the-art.

Weaknesses

-The guarantees provided in the paper are asymptotic in nature and depend on the quality of the first-order approximation of the problem. This introduces some limitations and potential uncertainty in the practical applicability of the results.

Questions

The paper demonstrates strong mathematical foundations and is very well written with a very accurate state of the art. The contributions, particularly the formation of a tractable framework for Wasserstein Distributionally Robust Optimization (WDRO) with asymptotic guarantees, are significant. However, certain concerns arise: -The proposed attack and adversarial accuracy boundaries heavily rely on the validity of the first-order approximation. Therefore, one query would be how the first-order estimation scales with the dimension of the input space. Specifically, does this approach retain its efficacy when applied to larger vision problems, such as those found in ImageNet (224x224x3)? -The use of the term 'certified bound' seems somewhat misleading as it actually refers to an asymptotic result. The proposed bounds, as they currently stand, cannot provide certification for a given radius (even a small one) as depicted in appendix Figure E1. Furthermore, the lower bound appears to be loose in the experiments. The approach proposed in the appendix to tighten the bound sometimes leads to overestimation, rendering it unsuitable for certification without additional guarantees. Therefore, how could this issue be addressed for a more accurate and reliable certification?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations

Reviewer 5QWe4/10 · confidence 4/52023-07-07

Summary

The paper proposes a new method combining Wasserstein distance, DRO, and adversarial attack. The novelty is that the method computes the adversarial attack bounded by the Wasserstein distance. The paper provides a bound on the performance of out-of-domain samples. It applies the new adversarial attack on CIFAR10 dataset. The numerical results show that the attack is effective on models that are robust to pointwise attacks, and computing attacks is much faster than other methods.

Strengths

1. The paper proposes an interesting direction by combining DRO and adversarial attacks. It is clear in the writing and presenting methods. 2. The paper gives some theoretical results on the proposed methods, like lower bound on $R_{\delta}$. The paper also shows numerical results comparing the lower bound from different methods. 3. The paper provides numerical results to show that the adversarial attack is effective at attacking and fast to compute.

Weaknesses

1. The paper misses some references and does not compare the method with the existing DRO adversarial attack methods, like [A]. [B, C] also have a similar approach by combining DRO with adversarial attacks. In the area of traditional attacks, the paper can compare with more methods, like CW attack in [D]. [A] Sinha, A., Namkoong, H., Volpi, R. and Duchi, J., 2017. Certifying some distributional robustness with principled adversarial training. arXiv preprint arXiv:1710.10571. [B] Volpi, R., Namkoong, H., Sener, O., Duchi, J.C., Murino, V. and Savarese, S., 2018. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31. [C] Hua, X., Xu, H., Blanchet, J. and Nguyen, V.A., 2022, December. Human imperceptible attacks and applications to improve fairness. In 2022 Winter Simulation Conference (WSC) (pp. 2641-2652). IEEE. [D] Carlini, N. and Wagner, D., 2017, May. Towards evaluating the robustness of neural networks. In 2017 ieee symposium on security and privacy (sp) (pp. 39-57). Ieee. 2. The paper only does experiments on CIFAR10 dataset. It would be helpful to see results on other datasets.

Questions

1. For the computation time in Table 2, do the authors compare with PGD? PGD is a simple gradient-based method, so the computation time should be very small. Other methods, like FGSM, are also very fast. Also, for Table 2, do all the algorithms stop when they successfully attack all the images? 2. In line 165, the authors mention another distributional attack, why this method is not compared to in the paper? 3. In line 39, the paper mentions the approach "is better suited for gradient-based optimization methods than duality approach adopted in most of the works to date". Can you clarify on this point?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors discuss some of the limitations of the work, including only valid on small-scale attacks. More limitations could include the stability of the attack and how it compares to more state-of-the-art attack methods. The method also should be compared with more attack methods, like DRO-based attacks. From the current experiments, it is not very convincing that the method outperforms many other adversarial attack methods.

Reviewer fHjF6/10 · confidence 4/52023-07-07

Summary

This paper links more general attacks with question of out-of-sample performance and Knightian uncertainty. To evaluate the distributional robustness of neural networks, this paper proposes a first-order AA algorithm and its multistep version. The proposed attack algorithms include Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) as special cases. Furthermore, the authors 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. This paper also conducts numerical experiments on the CIFAR-10 dataset to illustrate the theoretical results.

Strengths

1. This paper proposes a unified approach to adversarial attacks and training based on sensitivity analysis for Wasserstein DRO. This approach leveraging results from Bartl et al. (2021) is better suited for gradient-based optimization methods than duality approach adopted in most of the works to date. This paper further links the adversarial accuracy to the adversarial loss, and investigate the out-of-sample performance. 2. This paper derives a general adversarial attack method. As a special case, this recovers the classical FGSM attack lending it a further theoretical underpinning. However, the proposed method also allows to carry out attacks under a distributional threat model which has not been done before. This paper develops certified bounds on adversarial accuracy including the classical pointwise perturbations. The bounds are first-order accurate and much faster to compute than the AutoAttack benchmark.

Weaknesses

1. The Wasserstein distance for the adversarial robustness was usually discussed in the theoretical analysis. This paper utilizes the first order approximation of the adversarial loss on the CIFAR-10 for the verification. This approach contributes on the efficiency of the Wasserstein distributional robustness for the practical use. Yet, it is still questionable on the robustness on the large-scale dataset. Due to the computational complexity of the data dimension for the Wasserstein distance, it would be interesting to see the performance and the efficiency of the proposed approach on the image data such as tiny-ImageNet with dimension 64 and ImageNet with dimension 224. 2. The Wasserstein distance is about the overall loss rather than the single norm such as L_2 and L_\infty. Thus, the Wasserstein robustness may contribute to other types of robustness such as common corruptions, natural adversarial examples and the perturbations beyond the adversarial robustness. It would be more sufficient to see the potential improvement of the proposed approach on more general perturbed images.

Questions

The proposed results are valid for small attacks and it is still questionable for the proposed approach for the practical use. Additionally, it would be more general for the proposed method to further discuss on the generalization ability of the Wasserstein-robust models.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This paper uses Wasserstein distributionally robust optimization (DRO) and obtain novel contributions leveraging recent insights from DRO sensitivity analysis. Yet, it would be better to clarify with more sufficient evidences. The detailed comments can be seen in the weaknesses.

Reviewer 1Lzk2023-08-17

Re: Rebuttal

Thanks for the rebuttal. I will keep my positive rating.

Reviewer Q9Xf2023-08-18

Thank you for addressing my query. I stand by my score and recommend accepting the paper.

Reviewer fHjF2023-08-19

Thanks for the detailed response. The authors have provided the convincing explainations for the weaknesses of the proposed approach such as the general perturbed images. I agree with the mentioned limitations to some degree. The authors also provided the empirical evidences in higher dimensions. Thus, I would raise my score.

Authorsrebuttal2023-08-19

Thank you for getting back to us. We tried our best to address all the questions and queries comprehensively. We are therefore very pleased to hear you found our rebuttals detailed and convincing. If you have any remaining questions or doubts please let us know and we would be very glad to engage in a discussion.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC