acceptodds
Under review as a conference paper at ICLR 2027

A Mirror Descent Framework for Adversarial Attacks over the Clipped Ball and Group Sparsity Constraints

Abstract

We introduce a Mirror Descent framework to solve optimization problems over a clipped ball, generalizing the -APGD algorithm to group-norm constraints. We resolve the expensive high-dimensional Bregman projection by reducing it to a continuous resource-allocation problem that is computed exactly via a efficient 1D root search. Furthermore, we propose a structurally relaxed projection that decouples within-group coordinate dependencies, reducing computational overhead while systematically enforcing group sparsity. Instantiating this framework as Group-Structured Mirror Descent (GSMD) for adversarial attack, our extensive evaluations on ImageNet reveal a clear trade-off: the exact projection accelerates objective convergence, whereas the relaxation yields strictly sparser modifications. GSMD generates localized, coherent perturbations and achieves state-of-the-art attack success rates on both standard and robust classifiers, requiring significantly fewer backward passes than existing structured attacks.

Primary area:
optimization
Keywords:
Optimisation, Sparsity, Adversarial Attack

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LxB6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GSMD, a mirror descent framework for generating structured sparse adversarial perturbations. It introduces an exact Bregman projection method and a relaxed projection strategy to improve computational efficiency, with evaluations on standard and robust image classifiers.

Critical strengths
  1. The paper is clearly written and well organized.

  2. The proposed projection method reduces computational complexity.

  3. Experiments show promising attack performance and structured sparsity compared with several baseline methods.

Critical weaknesses
  1. The method relies on predefined group structures, such as fixed grids and superpixels. A more systematic sensitivity analysis of grouping strategies and their parameters is needed. 
  2. Although several architectures and robust classifiers are included, the evaluation does not establish whether the method generalizes to other datasets. 
  3. The baseline comparisons are limited. Evaluations under matched perturbation budgets and computational costs would better establish the method's advantages over existing attacks. 
  4. Lower numbers of connected components do not necessarily imply that an adversarial perturbation is less perceptible. The paper would benefit from qualitative visualizations and additional perceptual analyses to substantiate claims about localized and coherent perturbations.
Questions for authors

see weakness

1 comment

We thank the reviewer for their feedback.

Two of the four weaknesses (W3, W4) ask for material that is already in the paper, with section and figure references below.

The other two (W1, W2) do not identify a specific deficiency.

None of the points concerns the correctness of the method or its analysis.

1. Already in the paper

W3 (compute and budgets): Runtime and the number of forward/backward passes are reported in Appendix C.1, Fig. 5. GSMD needs up to 93% fewer backward passes than StrAttack [1] and is the fastest method per sample (0.75s vs. 2.38–7.69s on ResNet50). The baselines use different threat models: an ℓ2,1\ell_{2,1} penalty (StrAttack [1]), a nuclear-norm constraint (FWnucl [2]) and non-convex regularizers (GSE [3]). Hence there is no common budget to match. All methods are therefore compared on ASRτ,ACP,ANCASR_{\tau}, ACP, ANC and d2,0d_{2,0} with the GSE hyperparameters [3] (Appendix B.2).

W4 (visualizations and perceptibility): Qualitative examples are in Appendix C.3, Fig. 6: 16 images comparing GSE [3] and GSMD masks next to Grad-CAM [6]. We do not claim imperceptibility; we claim structure, and ANCANC is the standard metric for that, taken from GSE [3]. The unstructured sparse setting is the singleton case of our framework (ℓ1\ell_1-APGD [4]); the group size controls the trade-off between pixel sparsity and structure.

2. No specific deficiency identified

W1 (grouping): The grouping is an input to the method, not part of it. The projection is exact for any partition (Prop. 3.1), and with singleton groups it recovers ℓ1\ell_1-APGD [4] exactly (Sec. 3.2). We evaluate two very different groupings, a fixed grid and SLIC superpixels [5], and GSMD outperforms all baselines with both. The dependence on the segmentation is discussed in Sec. 5.

W2 (datasets): We evaluate six models on ImageNet, three of them adversarially trained, following the GSE protocol [3]. We agree that broader evaluations are an interesting direction for future work.


References

[1] K. Xu, S. Liu, P. Zhao, P.-Y. Chen, H. Zhang, Q. Fan, D. Erdogmus, Y. Wang, X. Lin. Structured adversarial attack: Towards general implementation and better interpretability. ICLR, 2019.

[2] E. Kazemi, T. Kerdreux, L. Wang. Minimally distorted structured adversarial attacks. International Journal of Computer Vision, 131(1):160–176, 2023.

[3] S. Sadiku, M. Wagner, S. Pokutta. GSE: Group-wise sparse and explainable adversarial attacks. ICLR, 2025.

[4] F. Croce, M. Hein. Mind the box: ℓ1\ell_1-APGD for sparse adversarial attacks on image classifiers. ICML, PMLR 139:2201–2211, 2021.

[5] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE TPAMI, 34(11):2274–2282, 2012.

Reviewer_7RHWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of generating structured adversarial perturbations. Specifically, the authors propose Group-Structured Mirror Descent (GSMD), which operates over a clipped ℓ∞,1\ell_{\infty,1} ball and uses groups of pixels, such as patches or superpixels, to encourage spatially coherent perturbations. To this purpose, the authors introduce a relaxed projection to encourage group-level sparsity. The experimental results suggest that the method can generate localized perturbations against standard and robust classifiers while requiring fewer backward passes than the structured attacks considered.

Critical strengths

The paper addresses a relevant problem in adversarial robustness, i.e., generating perturbations that account for the spatial structure of images rather than treating pixels independently.

The results show an advantage in terms of backward passes (thus time complexity) compared to the structured attacks considered.

Critical weaknesses
  • Limited diversity of the evaluated defenses

The selection of robust models appears somewhat limited, with a primary focus on naive ℓ2\ell_2-adversarially trained models. However, the story around adversarial training is long and more and more advanced defenses around it have been developed. Specifically, the authors should consider additional robust models, for example from the RobustBench repository. I would find the conclusions more convincing with a more diverse set of defenses.

  • Comparison with patch-based attacks

The proposed method aims to generate spatially coherent, patch-like perturbations. However, the experimental analysis lacks a direct comparison with patch-based adversarial attacks. Of course, the comparison should account for differences in perturbation magnitude and patch location. But to the best of my understanding, the problem setup is the same as that of patch attacks. Furthermore, comparisons with ℓ∞\ell_\infty-bounded attacks and ℓ0\ell_0-constrained sparse attacks would help the reader understand what is gained through the proposed ℓ∞,1\ell_{\infty,1} formulation.

  • Lack of self-contained presentation

My main concern regarding the presentation is that several metrics used to support the central results are defined only in the Appendix. To be specific, Tables 1 and 2 rely on ASRτASR_\tau, ANCANC, d2,0d_{2,0}, and ACPACP, but their formal definitions, together with the explanation of the best, average, and worst-case settings, are not provided in the main paper.

In my opinion, the main paper should be self-contained. Because of that the are aspect that remain fundamental to keep in the main discussion, especially regarding the metrics needed to understand the experimental analysis and outcomes. A reader should not be required to repeatedly consult the Appendix to interpret the central results ^*. I acknowledge the additional page that would be granted in the camera-ready, but it is a responsibility of the authors to plan the structure of the paper so that the essential information is included in the main text.

^* according to the ICLR guidelines, the reviewer would not be forced to read the appendix.

  • Geometric intuition and visualization

A minor observation. While the qualitative examples in Figure 6 help illustrate the resulting perturbations, I would appreciate a more intuitive explanation of the Mirror Descent approach and the projection mechanism. In particular, a geometric visualization of the clipped ℓ∞,1\ell_{\infty,1} ball, possibly through a low-dimensional example, could help explain how it differs from standard ℓ1\ell_1 and ℓ∞\ell_\infty constraints. Similarly, provide also the visualization of some of the generated adversarial examples.

Questions for authors

Would the observed effectiveness and efficiency persist across a broader range of robust models and training threat models?

Would it be possible for the authors to provide a geometric illustration to explain the projection mechanism and its relationship to group-level sparsity?

Would it be possible for the authors to include visualizaiton of the output adversarial examples? Can we compare multiple of them with the ones created by other attacks?

3 comments

We thank the reviewer for their feedback. Below, we address the reviewer's concerns.

Robust models. Our robust models already span three different training threat models: ℓ2\ell_2 adversarial training [1], ℓ2\ell_2 training fine-tuned against ℓ∞\ell_\infty and ℓ1\ell_1 [2], and an ℓ∞\ell_\infty-robust ConvNeXt with ConvStem [3], an architecture that yields the most ℓ∞\ell_\infty-robust ImageNet models in its parameter range [3]. GSMD achieves the highest ASR0.5\mathrm{ASR}_{0.5} against all three (Table 2).

Patch attacks. The setting is different. Patch attacks [5, 6] place one contiguous patch of fixed shape, with unbounded magnitude inside it, and optimize its location. GSMD bounds the total perturbation magnitude, does not fix the number, shape or location of active regions, and lets the optimizer select them. The two threat models are not comparable without changing one of them.

ℓ1\ell_1 and ℓ0\ell_0 attacks. ℓ1\ell_1-APGD [7], the strongest ℓ1\ell_1-constrained sparse attack, is a baseline in every table. It is also the special case of our framework with singleton groups and Euclidean geometry. The comparison shows the gain from the group formulation: ℓ1\ell_1-APGD changes fewer pixels (lower ACP), but they are scattered (ANC 434 vs. 2.70 on ResNet50, average case).

Self-contained presentation. All metrics (ASRτ,ACP,ANCASR_{\tau}, ACP, ANC, d2,0d_{2,0}) and the best/average/worst-case protocol are standard in the structured sparse attack literature and taken from [8]. Sec. 4 names each metric, and their formal definitions are given in Appendix B.1. In the revised version, we will extend Appendix B.1 with a more detailed explanation of each metric and of the best/average/worst-case protocol, and add an explicit pointer to it in Sec. 4.

Geometric illustration. We will add a low-dimensional figure comparing the clipped ℓ∞,1\ell_{\infty,1} ball with the ℓ1\ell_1 and ℓ∞\ell_\infty balls and showing how the projection sets whole groups to zero.

Adversarial examples. Fig. 6 (Appendix C.3) compares perturbation masks of GSE [8] and GSMD on 16 images. We will add the corresponding adversarial images for all methods.


References

[1] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, A. Vladu. Towards deep learning models resistant to adversarial attacks. ICLR, 2018.

[2] F. Croce, M. Hein. Adversarial robustness against multiple and single ℓp\ell_p-threat models via quick fine-tuning of robust classifiers. ICML, PMLR 162:4436–4454, 2022.

[3] N. D. Singh, F. Croce, M. Hein. Revisiting adversarial training for ImageNet: Architectures, training and generalization across threat models. NeurIPS, 36:13931–13955, 2023.

[4] F. Croce, M. Andriushchenko, V. Sehwag, E. Debenedetti, N. Flammarion, M. Chiang, P. Mittal, M. Hein. RobustBench: A standardized adversarial robustness benchmark. NeurIPS Datasets and Benchmarks Track, 2021.

[5] T. B. Brown, D. Mané, A. Roy, M. Abadi, J. Gilmer. Adversarial patch. arXiv:1712.09665, 2017.

[6] D. Karmon, D. Zoran, Y. Goldberg. LaVAN: Localized and visible adversarial noise. ICML, PMLR 80:2507–2515, 2018.

[7] F. Croce, M. Hein. Mind the box: ℓ1\ell_1-APGD for sparse adversarial attacks on image classifiers. ICML, PMLR 139:2201–2211, 2021.

[8] S. Sadiku, M. Wagner, S. Pokutta. GSE: Group-wise sparse and explainable adversarial attacks. ICLR, 2025.

Dear authors,

Thank you for the response and the planned additions. I would like to clarify a few points that remain open.

Robust models. I acknowledge that the evaluation includes three training threat models and a robust ConvNeXt model (considering that the RobustBench repo I indicated contains more robust models). My concern was about the breadth of the evaluation across defenses and architectures, including within each threat model (my review: “…with a primary focus on naive”). In my opinion, covering several norms is fundamental, and at the same time testing a small number of models does not establish whether the effectiveness and efficiency gains consistently generalize.

Patch attacks and practical motivation. Fair point from the authors regarding the different threat model of their attack and patch attacks. I agree with them on this distinction. However, a controlled comparison with a contiguous-region baseline could still be informative. For example, one could impose the same perturbation-magnitude constraint and compare performance at matched support sizes. It is a baseline that would establish what is gained by allowing several freely selected regions instead of one contiguous region. Again, a baseline that is also cheap to have. Furthermore, the contiguous perturbed regions visible in Figure 6 further motivate a comparison with patch-based attacks, as the resulting noise seems to be located often in contiguous regions.

Furthermore, the response from the authors raises another question in my mind related to their threat model. I would also appreciate a clearer explanation of the intended application. In some practical settings, the attacker can modify only a particular region, such as a visible surface on an object. (e.g., a t-shirt or sunglasses). How does the proposed formulation relate to these restrictions, and when is the freedom to select regions throughout the image useful?

Sparse attack baselines. Thank you for your response on ℓ0\ell_0 and ℓ1\ell_1 attacks. I wish to be clear, ℓ1\ell_1-APGD addresses the ℓ1\ell_1 comparison, but leaves the question of ℓ0\ell_0 baselines open. Relevant examples from the literature include PGD-ℓ0\ell_0, sPGD, or σ\sigma-zero as white-box references and Sparse-RS as a black-box one.

Self-contained presentation. My concern, weak compared to the others, is specifically about the information available in the main paper. Extending Appendix B.1 and adding a pointer to it does not fully address this. Even if the metrics and evaluation protocol follow previous work, readers should be able to interpret the central tables without already knowing that work or repeatedly consulting the appendix.

I hope the authors will find these comments constructive and helpful for their paper.

References

  • [PGD-L0] Croce et al. Sparse and Imperceivable Adversarial Attacks. ICCV 2019

  • [sPGD] Zhong et al, Towards Efficient Training and Evaluation of Robust Models against ℓ0\ell_0 Bounded Adversarial Perturbations. ICML 2024

  • [sigma-zero] Cinà et al, sigma-zero: Gradient-based Optimization of ℓ0\ell_0-norm Adversarial Examples. ICLR 2025.

  • [Sparse-RS] Croce et al. Sparse-RS: A Versatile Framework for Query-Efficient Sparse Black-Box Adversarial Attacks. AAAI 2022.

Robust models. Our evaluation covers three training threat models and architectures from convolutional to modern ConvNeXt, including a multi-norm fine-tuned model and a ConvNeXt with ConvStem, an architecture that yields the most ℓ∞\ell_\infty-robust ImageNet models in its parameter range; neither is a naive defense. The gains are consistent and large across all six models. The comment does not point to a defense or architecture for which we should expect a different outcome, and our claims concern the optimization method, whose behavior does not depend on the specific defense.

Contiguous-region baseline. A single region is a special case: run GSMD on that region only, with any segmentation intersected with it; the method and its guarantees are unchanged. The restricted attack optimizes over a lower-dimensional subset.

Region-restricted threat models. If the attacker can only modify a region (e.g., a t-shirt), run GSMD on that region only, with its own grid or superpixel segmentation. Selecting regions over the whole image is useful to find the most vulnerable regions and gives the worst case over all such restrictions. The two can be combined: a second GSMD run can refine the attack within the selected regions, with a finer or different segmentation.

Sparse attack baselines. ℓ1\ell_1-APGD is included because it is the same constraint without groups (the singleton case), so the comparison shows what the group structure adds; it is not meant as a sparsity benchmark. Our claims concern structure, and our baselines for that are StrAttack, FWnucl and GSE. We are not aware of structured extensions of Sparse-RS, PGD-ℓ0\ell_0, sPGD or σ\sigma-zero.

Self-contained presentation. Correction to our previous reply: ASR$_\tau$ is our own metric, not from prior work. We will add its one-sentence definition to Sec. 4; the remaining metrics follow GSE [2] and are defined in Appendix B.1.

Reviewer_p3bzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a mirror descent framework for adversarial optimization under a clipped ℓ∞,1\ell_{\infty,1}-ball constraint, generating localized and coherent perturbations that achieve stronger attack performance. The main contributions are twofold. First, the authors propose an exact Bregman projection that reformulates the high-dimensional projection problem as a continuous resource allocation problem, which is then solved exactly via sorting and one-dimensional root finding. Second, they introduce an efficient relaxed structured projection that controls sparsity while reducing computational overhead without compromising attack performance. Experiments show that GSMD achieves higher attack success rates than existing structured sparse attacks.

Critical strengths
  1. The authors extend the exact geometric projection to group-structured sparsity via the Bregman projection, reformulating the projection as a resource allocation problem that can be solved efficiently.

  2. The authors propose an efficient relaxed structured projection whose main computational cost scales with the number of groups rather than the number of coordinates.

  3. Experiments on ImageNet validation images cover both standard classifiers (ResNet-50, EfficientNet-B3, and ViT-B/16) and adversarially robust ones. Across different sparsity levels, GSMD achieves higher ASR than the baseline methods.

Critical weaknesses
  1. The main text focuses on the implementation details of the algorithm but does not clearly describe the overall attack pipeline of GSMD.

  2. The paper does not adequately explain how the mirror descent update in Eq. 3 leads to the two-step form in Eqs. 4 and 5, and the connection between them is not made clear.

  3. The paper claims that GSMD produces localized and coherent perturbations, but the method section only describes the projection operations themselves and does not explain how these operations lead to coherence.

  4. The paper does not adequately explain why the exponentiated gradient update induced by the smoothed entropy mirror map outperforms other gradient update schemes.

  5. The experimental evaluation is limited. For example, the range of surrogate models is not sufficiently comprehensive, and the transferability of GSMD has not been evaluated.

Questions for authors
  1. Could the authors provide more details on the GSMD attack pipeline, clarifying how the mirror descent update and the structured projection are integrated into the overall attack process?

  2. Could the authors explain in detail why the mirror descent update in Eq. 3 is equivalent to the two-step form in Eqs. 4 and 5?

  3. The paper provides the procedure for computing the projection by examining breakpoints and critical points, but the conditions under which this procedure is guaranteed to yield the exact projection are not stated. Could the authors specify these conditions and provide the corresponding theoretical justification?

  4. The paper states that the exponentiated gradient update with the smoothed entropic mirror map outperforms standard PGD, but this claim is not supported by theoretical analysis or direct experimental comparison. Could the authors provide either to substantiate it?

  5. The baseline GSE was evaluated on VGG-19 under both targeted and untargeted settings. Could the authors report the performance of GSMD on VGG-19, particularly in the untargeted setting, to enable a direct comparison with GSE?

  6. Transferability is an important metric for evaluating adversarial attacks. Could the authors report the transfer performance of GSMD across different models and compare it with existing sparse attacks such as GSE?

  7. It remains unclear whether the performance gains stem from the proposed optimization method, the group partitioning strategy, or their combination. Could the authors provide ablation studies that isolate the contribution of each component?

4 comments

We thank the reviewer for their feedback.

The reviewer's questions fall into three groups:

(i) Q1–Q3 and W3 concern content that is stated in the paper (Sec. 3.3, Assumptions 1–2, Prop. 3.1, Theorem D.1) or is a textbook property of Mirror Descent (Q2);

(ii) Q5 and Q6 lie outside the scope of the paper: we make no transferability claim, and transferability depends on the loss and surrogate, not on the optimization method, which is our main contribution;

(iii) Q4 and Q7 are answered below.

None of the questions identifies an error in the method or its analysis.

Short answers to Group (i)

Q1 (pipeline). GSMD follows the standard projected-gradient attack template [1, 2], with the Euclidean step replaced by a mirror step. Each iteration is one mirror step (Eq. 4) followed by the projection Π~ϕ\widetilde{\Pi}_{\phi} (Alg. 2), as written in Sec. 3.3. The hyperparameters are in Appendix B.2.

Q2 (Eq. 3 vs. Eqs. 4–5). This is the textbook two-step form of Mirror Descent [3, 4]. Both problems have the same optimality condition, ∇Φ(δt+1)−∇Φ(δt)+ηtgt∈−NC(δt+1)\nabla\Phi(\delta_{t+1})-\nabla\Phi(\delta_{t})+\eta_t g_t \in -N_C(\delta_{t+1}), and a unique solution, since Φ\Phi is strictly convex.

Q3 (conditions for exactness). The conditions are stated in Assumption 1 (ϕ\phi separable, symmetric, strictly convex, ϕ(0)=0\phi(0)=0) and Assumption 2. Exactness is proven in Prop. 3.1, Lemma 3.2 and Theorem D.1 (Appendices D and E).

W3 (coherence). Coherence comes from the constraint, as explained in Sec. 3.3. Each group is charged by its largest entry, so once a group is active, its other coordinates can be perturbed at no extra cost. The relaxed projection (Sec. 3.2) then applies a hard cutoff to the group maxima, which sets weak groups to zero. The result is a few active groups, each filled out, and this is what gives the low ANC and d2,0d_{2,0} in Tables 1–3.

Short answers to Group (ii)

Q5 (VGG-19). We evaluate six models: ResNet50, EfficientNet-B3, ViT-B/16, two adversarially trained ResNet50s and a robust ConvNeXt. Together they cover convolutional, transformer and robust architectures. VGG-19 is an older convolutional model and would add no new architecture type.

Q6 (transferability). This question does not relate to any claim in the paper, as we make no transferability claim. All results are white-box. Our contribution is an optimization method, namely Mirror Descent with an exact (and a relaxed) Bregman projection onto the clipped ℓ∞,1\ell_{\infty,1} ball, and the attack is an application where the solver determines the outcome. Transferability is governed by the surrogate model and the loss, not by the solver, so transfer results would neither support nor weaken any of our claims. GSMD can be combined with any transfer-oriented loss or surrogate.

Short answers to Group (iii)

Q4 (entropic vs. Euclidean). Mirror Descent with an entropic potential is the standard choice for ℓ1\ell_1-type geometry. Its convergence depends only logarithmically on the dimension, whereas Euclidean projected gradient depends polynomially [5, 4, 6, 7]. The Euclidean case is part of our framework: with singleton groups it is exactly ℓ1\ell_1-APGD [2], which is a baseline in every table.

Q7 (ablation). Sec. 4.1 isolates the projection: exact vs. relaxed, with the same groups and the same mirror map. ℓ1\ell_1-APGD [2] corresponds to Euclidean geometry without grouping.


References

[1] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, A. Vladu. Towards deep learning models resistant to adversarial attacks. ICLR, 2018.

[2] F. Croce, M. Hein. Mind the box: ℓ1\ell_1-APGD for sparse adversarial attacks on image classifiers. ICML, PMLR 139:2201–2211, 2021.

[3] A. Beck, M. Teboulle. Mirror descent and nonlinear projected subgradient methods for convex optimization. Operations Research Letters, 31(3):167–175, 2003.

[4] G. Lan. First-order and Stochastic Optimization Methods for Machine Learning. Springer, 2020.

[5] A. S. Nemirovskij, D. B. Yudin. Problem Complexity and Method Efficiency in Optimization. Wiley, 1983.

[6] U. Ghai, E. Hazan, Y. Singer. Exponentiated gradient meets gradient descent. ALT, PMLR 117:386–407, 2020.

[7] W. Shao, F. Sivrikaya, S. Albayrak. Optimistic optimisation of composite objective with exponentiated update. Machine Learning, 2022.

AuthorsOpenReview ↗

Follow up Q5 (VGG-19 Evaluation)

Dear Reviewer, following your request, we additionally evaluated both GSMD variants on VGG-19 under targeted and untargeted settings for direct comparison with GSE. The results below demonstrate consistently stronger targeted performance. For untargeted attacks, both variants improve ASR0.8,ANCASR_{0.8}, ANC, and d2,0d_{2,0}, with higher ACP as the trade-off.


Targeted Attacks

Note: The targeted VGG-19 results for GSE in [1] were reported on 1,000 ImageNet samples. We extend this evaluation to the full set of 10,000 samples.

Best case

ModelAttackASR$_{0.0}$ASR$_{0.8}$ACPANCd2,0d_{2,0}
GSE100%87%1,5976.453,526
VGG19GSMD100%100%1,0701.091,682
GSMD-s100%100%8941.031,588

Average case

ModelAttackASR$_{0.0}$ASR$_{0.8}$ACPANCd2,0d_{2,0}
GSE100%52%6,54515.110,939
VGG19GSMD100%91%2,6732.523,948
GSMD-s100%98%1,9671.993,251

Worst case

ModelAttackASR$_{0.0}$ASR$_{0.8}$ACPANCd2,0d_{2,0}
GSE98%14%17,04325.822,595
VGG19GSMD100%57%5,4505.497,758
GSMD-s100%89%3,8764.026,066

Untargeted Attacks

ModelAttackASR$_{0.0}$ASR$_{0.8}$ACPANCd2,0d_{2,0}
GSE100%98%7985.441,885
VGG19GSMD100%99%9361.731,571
GSMD-s100%100%8641.571,620

References:

[1] S. Sadiku, M. Wagner, S. Pokutta. GSE: Group-wise sparse and explainable adversarial attacks. ICLR, 2025.

I thank the authors for their detailed response. Although the paper proposes a different type of optimization algorithm to address the sparse attack problem, the research itself belongs to the field of adversarial attacks rather than optimization. The authors should therefore focus on the existing limitations and open problems in adversarial attacks in general and sparse attacks in particular. Given the current emphasis on evaluating and comparing attack transferability, as exemplified by the experimental setup in the GSE paper, the authors should extend their evaluation beyond white-box attacks to include such comparisons. Based on these considerations, I maintain my original score.

We agree that transferability is an interesting property of adversarial attacks. However, we believe it is not within the scope of our contribution. Our paper proposes an optimization method, Mirror Descent with exact and relaxed Bregman projections onto the clipped ℓ∞,1\ell_{\infty,1} ball, and white-box evaluation is the standard way to assess the strength of such solvers, as in ℓ1\ell_1-APGD [1].

This is consistent with the closest baseline, GSE [2], whose main evaluation is also white-box. GSE considers transferability only in a small-scale experiment in its appendix not as part of its main claims. Within this white-box setting, we extend GSE's evaluation in several ways:

  • added EfficientNet_B_3 as additional base classifier;
  • added two adversarially trained models in the targeted setting;
  • the full set of 10,000 ImageNet samples for targeted attacks, instead of the 1,000 used in GSE;
  • following your request, VGG-19 in both the targeted and untargeted settings, for direct comparison with GSE.

In the revised version, we will add transferability to the discussion as a limitation and a direction for future work.

We hope this clarifies the scope of the paper, and we would be grateful if you would take the extended evaluation into account in your final assessment.


References

[1] F. Croce, M. Hein. Mind the box: ℓ1\ell_1-APGD for sparse adversarial attacks on image classifiers. ICML, 2021.

[2] S. Sadiku, M. Wagner, S. Pokutta. GSE: Group-wise sparse and explainable adversarial attacks. ICLR, 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers for their feedback. We have updated the paper with the following changes:

  1. Adversarial examples for all methods (Reviewer 7RHW, Reviewer LxB6). We added qualitative examples showing the adversarial images and perturbations of GSMD next to all baseline attacks, so the structure of the perturbations can be compared directly. [Fig. 7,8, App. C.3]

  2. Geometric intuition (Reviewer 7RHW). A new low-dimensional figure compares the clipped ℓ∞,1\ell_{\infty,1} ball with the ℓ1\ell_1 and ℓ∞\ell_\infty balls and shows how the projection sets entire groups to zero. [Fig. 4, App. A]

  3. VGG-19 results for direct comparison with GSE (Reviewer p3bz). We evaluate both GSMD variants on VGG-19 in the targeted (best/average/worst case) and untargeted settings, using all 10,000 samples. In the targeted average case, ASR0.8ASR_{0.8} rises from 52% (GSE) to 91% (GSMD) and 98% (GSMD-s), while ANC falls from 15.1 to 2.52 and 1.99. In the untargeted setting, GSMD also improves ASR0.8ASR_{0.8}, ANC and d2,0d_{2,0}. [Sec. 4.2, Table 1, App. C.2, Table 4]

  4. Self-contained evaluation (Reviewer 7RHW). All metrics (ASRτASR_\tau, ACPACP, ANCANC, d2,0d_{2,0}) and the best/average/worst-case protocol are now defined in the main paper (Sec. 4).

The remaining questions concerned material already in the paper, and we answered them in the individual responses:

  • exactness conditions of the projection (Assumptions 1–2, Prop. 3.1, Thm. D.1)
  • the two-step Mirror Descent form (Sec. 3)
  • the attack pipeline (Sec. 3.3)
  • the ablation of exact vs. relaxed projection (Sec. 4.1)
  • runtime and backward passes (App. C.1, Fig. 6)
  • the two grouping strategies, grid and SLIC (App. B.1 Fig. 5)
  • the robust models covering three training threat models (Sec. 4.3 Table 2)

We would like to thank the reviewers again for their help in improving the paper. We would sincerely appreciate it if you could consider updating your initial scores in light of our responses and the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.