Pith. sign in

REVIEW 4 major objections 9 minor 57 references

Towards Interpretable Adversarial Examples via Sparse Adversarial Attack

T0 review · 4 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that the NP-hard L0 sparse-attack problem can be approximated by learning a mask over a fixed I-FGSM perturbation, and that this yields faster, sparser, more transferable attacks whose perturbations reveal two…

desk verdict A fast, genuinely sparse attack worth knowing, buried under an unsupported near-optimality claim. read the letter →

arxiv 2506.17250 v1 pith:NPJVBC2S submitted 2025-06-08 cs.LG cs.AI

classification cs.LGcs.AI
keywords sparseadversarialattackL0norminterpretableexamplesdeepneuralnetworkrobustnessI-FGSMreparameterizationtransferabilityobscuringandleadingnoise
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to make sparse adversarial attacks practical and interpretable. It claims that the hard problem of finding the fewest pixels to change so a classifier mispredicts can be approximated by taking a dense I-FGSM perturbation and learning a mask over it, keeping only the pixels that matter. The authors introduce a differentiable surrogate for the $\ell^0$ norm via a Gaussian approximation of the Dirac delta, plus a box-constraint trick that guarantees valid images. On CIFAR-10 and ImageNet they report far fewer perturbed pixels than prior sparse attacks, orders-of-magnitude faster computation, higher transferability, and stronger attacks against robust models. The resulting minimal perturbations, they argue, expose two interpretable categories of noise: 'obscuring' and 'leading'.

What carries the argument

The central object is the mask $w$ applied element-wise to a one-time I-FGSM perturbation $\delta$, so the search space is the support of that fixed dense attack. To make the discrete $\ell^0$ penalty trainable, the Heaviside function $H(\pi(w))$ is approximated by the Gaussian $q_a(x)$ from Eq. (12), whose derivative provides a surrogate for the Dirac delta and lets gradient descent decide which pixels to keep. The tailored ReLU $\pi(w - \tau/\epsilon)$ shifts weights down before thresholding, which Proposition 1 shows can only reduce the $\ell^0$ norm, while Proposition 2's cap $\Omega = \min(x/\epsilon, (1-x)/\epsilon)$ guarantees the output stays a valid image. The algorithm alternates gradient steps on the adversarial loss plus $\lambda \sum_j H(\pi(w_j - \tau/\epsilon))$ and then hard-thresholds the final weights to produce the sparse perturbation $\delta^* = \Omega \odot H(\pi(w - \tau/\epsilon)) \odot \delta$.

What would settle it

On a small dataset, compute the exact minimal $\ell^0$ attack by exhaustive search over pixel subsets and compare it with the mask learned on the fixed I-FGSM $\delta$; finding any image where the exact solution uses a pixel with $\delta_j = 0$ or an opposite sign would refute the near-optimality claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the NP-hard sparse-attack problem can be approximated by a masked optimization: minimize $\|w\|_0$ subject to $f_\theta(x + w \odot \delta) = y_{\text{adv}}$, where $\delta$ is a fixed perturbation from I-FGSM and $w$ is a learned mask. Because the $\ell^0$ count of nonzero mask entries is non-differentiable, the authors replace the Heaviside step function with a zero-centered Gaussian surrogate $q_a(x) = \frac{1}{|a|\sqrt{\pi}} \exp(-(x/a)^2)$ that converges to the Dirac delta as $a \to 0$, turning the objective into a smooth loss plus a sparsity penalty. A shifted ReLU $\pi(w - \tau/\epsilon)$ provably reduces the number of nonzero pixels, and the box constraint is enforced by $\Omega = \min(x/\epsilon, (1-x)/\epsilon)$, which guarantees $x + w' \odot \delta \in [0,1]^d$. The authors report that this single optimization beats ten sparse-attack baselines in sparsity (an average of 57 ImageNet pixels for non-targeted attacks), computation time (seconds versus minutes), transferability across VGG and ResNet models, and fooling rate on adversarially trained models. They further claim that the resulting minimal perturbations reveal two interpretable noise types, 'obscuring noise' and 'leading noise', which respectively hide the true class's features and add the target class's features.

Load-bearing premise

The load-bearing premise is that one fixed I-FGSM perturbation $\delta$ already contains every pixel and sign direction the optimal sparse attack will ever need; any pixel where $\delta$ is zero or has the wrong sign is permanently excluded from the search.

Editorial extensions

If this is right

  • Sparse attacks become fast enough (seconds on ImageNet) to serve as a routine robustness benchmark for CNNs.
  • Near-100% non-targeted fooling can be reached with about 57 ImageNet pixels on standard classifiers, and the examples transfer across VGG and ResNet families better than existing sparse attacks.
  • The attack keeps meaningful fooling rates on adversarially trained models, suggesting sparse perturbations remain a threat that robustness training does not fully close.
  • The discovered 'obscuring noise' and 'leading noise' give a concrete, visual vocabulary for how minimal pixel changes redirect a classifier's attention.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not explore is to run the mask optimization over several I-FGSM initializations (different $\epsilon$, random restarts, or targeted variants) and check whether sparsity improves, since the single fixed $\delta$ bounds the search space.
  • If the two-noise taxonomy is right, it makes a testable causal prediction: deleting only the 'obscuring' pixels should restore the true class's saliency, and deleting only the 'leading' pixels should remove the targeted misprediction; the released code makes this ablation straightforward.
  • Because the method is so cheap, it could plausibly be inserted into adversarial training as a regularizer that forces models to be robust to sparse changes, although the paper does not claim this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. The paper proposes a sparse adversarial attack for image classifiers. The method first computes a dense initial perturbation δ with I-FGSM (Eqs. (3)-(4)), then optimizes a continuous weight w, masked through ReLU and a Heaviside threshold, such that the perturbed image x + π(w − τ/ε)⊙δ misclassifies to the adversarial label y_adv while the number of active pixels is minimized (Eqs. (6)-(10)). Differentiability is addressed by approximating the Heaviside derivative with a Gaussian q_a (Eqs. (11)-(18)), sparsity is increased by a shifted ReLU threshold τ/ε (Eq. (19)), and box-validity is enforced by the per-pixel bound Ω = min(x/ε, (1−x)/ε) (Eq. (22)). The authors state that solving the masked problem Eq. (6) yields near-optimal solutions to the original L0 problem Eq. (5), and they prove two propositions: sparsity monotonicity of the thresholded mask (Proposition 1) and box-validity of the output (Proposition 2). Experiments on MNIST, CIFAR-10, and ImageNet compare against ten sparse-attack baselines and report superior sparsity, computational cost, transferability, and fooling rates on robust models, together with a Grad-CAM-based interpretation of perturbations as 'obscuring noise' and 'leading noise'. Code is released.

Significance. The empirical core of the paper is strong and appears genuine. The authors release their code, evaluate against ten sparse-attack baselines across three datasets and several architectures (VGG-16/19, ResNet-50/101/152), report ablations over the three hyperparameters a, λ, τ, and test against adversarially trained models (PGD-AT, Fast-AT). The main empirical claims — that masking a dense I-FGSM perturbation with a learned thresholded weight yields very sparse, fast, and transferable attacks — are falsifiable and supported by the tables (e.g., Table 1: 57 ImageNet pixels at 100% fooling rate; Table 4: 4.6 s versus 87.2 s for BruSLeAttack on VGG-16). The interpretability analysis in Section 5.3 is qualitative but suggestive. If the claims survive rescoping, the paper would provide a practically useful sparse-attack baseline and a starting point for interpreting the spatial structure of minimal perturbations. The theoretical contribution is limited: Propositions 1 and 2 are elementary, and the near-optimality assertion relative to Eq. (5) is unsupported, so the paper's value lies primarily in its empirical findings rather than in new optimization theory.

major comments (4)
  1. [§4.4, Eq. (6)] The sentence following Eq. (6) claims that solving the masked problem 'yields near-optimal solutions for Eq. (5), i.e., with the fewest perturbed pixels,' but no proof or argument is given, and the claim is false in general. The feasible outputs of Eq. (6) are exactly masks of the fixed I-FGSM perturbation δ: since δ is computed once and never updated (Algorithm 1), any optimal sparse perturbation of Eq. (5) that requires a pixel with δ_j = 0, a sign opposite to δ_j, or a magnitude larger than ε is unreachable. Propositions 1 and 2 only establish that the final output is sparser than the initial mask and is box-valid; they say nothing about proximity to the minimum of Eq. (5), and no proposition guarantees that the thresholded output δ* = Ω⊙H(π(w−τ/ε))⊙δ even satisfies f(x+δ*) = y_adv. Since the abstract and Section 1 invoke 'theoretical performance guarantees,' the paper must either prove a formal guarantee under explicit assumptions on the quality of the I-FGSM initialization or rescope the claim to 'sparse successful masks of one fixed dense perturbation,' ideally adding a sanity check on MNIST where near-exhaustive L0 search can estimate the true optimum for a few examples.
  2. [§4.4, Eq. (18) and Algorithm 1 (Appendix B)] The reparameterization claimed to make the L0 term tractable is not reflected in Algorithm 1. After Eq. (18), the paper states dH/dx ≈ q_a(x), but the pseudocode computes J_adv = J(f(x+w⊙δ), y_adv) + λ Σ_j H(w_j) and then updates w with ∇_w J_adv without any appearance of q_a or a. Since H is a step function, the term λ Σ_j H(w_j) has zero gradient almost everywhere; as written, the sparsity penalty cannot influence the optimization, and the actual sparsification in the algorithm comes from the per-iteration thresholding w := π(w − τ/ε). Either the gradient computation in the implementation uses the surrogate q_a (in which case the pseudocode and the text around Eq. (18) should state this explicitly, including how a is chosen or annealed), or the method does not use the surrogate (in which case the claims that the technique 'approximates the NP-hard l0 optimization problem' and makes it 'computationally tractable' should be revised to describe a projected or thresholded gradient method).
  3. [Appendix A.2, Eq. (35)] The proof of Proposition 2 begins by assuming δ_j ∈ {ε, 0, −ε}, which is the output of a single FGSM step but not of the I-FGSM initialization used in Algorithm 1: after T iterations of Eq. (3) or Eq. (4), δ_j is an integer multiple of α bounded in absolute value by ε and is generically not equal to ±ε. The box-validity statement is still true for any δ with |δ_j| ≤ ε and Ω_j = min(x_j/ε, (1−x_j)/ε), but the proof needs to treat the general case x_j + Ω_j δ_j ∈ [x_j − Ω_j ε, x_j + Ω_j ε] instead of the two extreme sign cases; Eq. (35) should be removed or corrected.
  4. [§5.1 and Algorithm 1] The manuscript reports non-targeted attack results in Tables 1 and 2, but Algorithm 1 is written only for the targeted formulation: it minimizes the cross-entropy J(f(x + w⊙δ), y_adv), and Section 5.1 describes the targeted choice of y_adv as the least-likely class. For non-targeted attacks the paper does not state which label is used in J, what sign appears in front of the loss term, or whether the I-FGSM initialization follows Eq. (3) or Eq. (4). Because the sign of the gradient and the choice of y_adv drastically change the behavior of the algorithm, this specification gap should be closed for the non-targeted experiments to be reproducible.
minor comments (9)
  1. [§2, §5.2, Appendix C.2] The text contains repeated typos, including 'constriant' (twice in Section 2), 'SpareFool' (Section 5.2), 're-writed', 'Obivisly', and 'the internet' (Appendix C.2); these should be corrected.
  2. [Appendix C.1 vs. §2] Appendix C.1 cites 'SparseFool [7]', but the main text introduces SparseFool as reference [35]; the citation numbering should be made consistent.
  3. [§4.4, Eq. (12)] The function q_a is introduced as a 'zero-centered normal distribution' without stating its variance; writing q_a(x) = N(x; 0, a²/2) would make the normalizing constant 1/(|a|√π) self-evident and the connection to the Dirac delta clearer.
  4. [Appendix A.1] The first step of the proof of Proposition 1 ('δ appears in all three terms, so we only need to prove...') is not a valid reduction in general; although the claimed inequalities are true, the product with δ requires a short case analysis on the support of δ.
  5. [Algorithm 1] The variable w is overwritten by π(w − τ/ε) at the start of every iteration and then updated by a gradient step with respect to ∇_w J_adv, so it is unclear whether gradients are taken with respect to the pre- or post-threshold value; distinct symbols (e.g., w̃ for the thresholded value) would remove the ambiguity.
  6. [§5.5, Table 3] The reported five-trial means have no error bars or ranges; several adjacent entries (e.g., 86.3 vs. 84.2 for ResNet-101-generated attacks on VGG-19) are close, so standard errors or per-trial ranges should be reported.
  7. [§5.3] The 'obscuring noise' and 'leading noise' categories are assigned by visual inspection of one or two images using hand-drawn boxes; the paper should state explicitly that these are post-hoc interpretations and could quantify the spatial overlap of perturbed pixels with the relevant Grad-CAM regions to support the proposed mechanism.
  8. [§4.4, Eq. (7)] The term 'Lagrangian relaxation' is imprecise: Eq. (7) is a regularized objective that informally trades off the adversarial loss against L0 sparsity, not the Lagrangian of the constrained problem Eq. (6), whose constraint is not differentiable.
  9. [Theorem 1, §4.4] The proof of Theorem 1 establishes pointwise convergence of q_a to the Dirac delta; since the use in Eq. (18) is as a surrogate for the distributional derivative of H, a remark that q_a has unit integral and converges in the distributional sense, together with a comment on the bias of the surrogate gradient for finite a, would make the argument complete.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical evaluations are independent benchmarks and the theoretical results are elementary statements about the method's own quantities, not disguised fits.

full rationale

The paper's central derivation replaces the L0 attack (Eq. 5) by a masked optimization over a fixed I-FGSM perturbation (Eq. 6), and then asserts without proof that Eq. (6) is near-optimal for Eq. (5). This is a genuine completeness and soundness gap: the output is by construction a sparse mask of one fixed dense perturbation, so pixels outside the support of that perturbation, or with sign opposite to it, are never available, and the paper's Propositions 1 and 2 establish only monotone sparsity reduction and box validity, not proximity to the optimum of Eq. (5). However, that gap is not circularity. Theorem 1 is a standard mollifier convergence fact with a proof given in the paper; Proposition 1 is an elementary inequality comparing three L0 counts; Proposition 2 is an elementary box-validity bound using the assumed sign structure of I-FGSM. None of these results is assumed from or equivalent to the empirical claims. The experimental comparisons in Tables 1 through 4 are run against external, independently published sparse attacks such as C&W, SparseFool, GreedyFool, Sparse-RS, and BruSLeAttack, with hyperparameters tuned by the authors' own ablations rather than fitted to the benchmark outcomes. No fitted parameter is renamed as a prediction, and no load-bearing uniqueness or ansatz is imported from the authors' prior work. The interpretability finding of obscuring noise versus leading noise is a post-hoc visualization category, not a derivation masquerading as a first-principles result. The unsupported near-optimality statement is a correctness risk that deserves a formal approximation bound, but it does not make the paper's derivation circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The central method rests on two unproved modeling choices: the sufficiency of the fixed I-FGSM perturbation and the validity of the surrogate-gradient approximation. Three hyperparameters are tuned on the benchmark data. The theoretical results in the paper are standard distributional facts or simple inequalities, not guarantees that the final hard-thresholded perturbation is adversarial.

free parameters (3)
  • lambda (λ) = 1e-2 for MNIST and CIFAR-10, 1e-3 for ImageNet
    Balances the classification loss and the L0 penalty. Set empirically per dataset and affects both sparsity and fooling rate.
  • a = 0.1
    Width of the Gaussian approximating the Dirac delta. Grid-searched for sparsity and convergence speed, then fixed for all datasets.
  • tau (τ) = 0.30
    Shift in the tailored ReLU that increases sparsity. Grid-searched between 0.2 and 0.4 based on ImageNet performance.
assumptions (5)
  • standard math The zero-centered Gaussian q_a(x) converges to the Dirac delta in the distributional sense as a approaches 0.
    Used in Theorem 1 to justify approximating the derivative of the Heaviside step function.
  • domain assumption Different DNNs learn similar decision boundaries, so dense attack perturbations transfer to other models.
    Section 4.3 cites prior work to motivate starting from I-FGSM perturbations for transferability.
  • ad hoc to paper The fixed initial perturbation δ from I-FGSM contains all directions needed for the optimal sparse attack.
    Equation (6) restricts the search to masking δ. The paper asserts near-optimality without proof.
  • ad hoc to paper Replacing the derivative of the Heaviside function with the Gaussian q_a(x) is a valid optimization surrogate.
    Equation (18) states dH/dx ≈ q_a(x), but Algorithm 1 does not specify how this surrogate is used in the gradient computation.
  • domain assumption Hard-thresholding the optimized weights with H(π(w - τ/eps)) preserves the adversarial property of the perturbation.
    Algorithm 1 binarizes the mask after optimization; there is no guarantee that this final discrete mask still fools the classifier.
invented entities (2)
  • Obscuring noise
    purpose: A qualitative category of adversarial perturbations that hide the true class's discriminative features, making the classifier unable to see the correct label.
    Defined through Grad-CAM visual inspection of two examples; no quantitative or falsifiable criterion is provided to distinguish it from leading noise.
  • Leading noise
    purpose: A qualitative category of adversarial perturbations that add features of the target class, steering the classifier toward the incorrect prediction.
    Illustrated with Grad-CAM on two examples only; the category is descriptive rather than measured or independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Interpretable Adversarial Examples via Sparse Adversarial Attack." pith.science (2026). https://pith.science/paper/NPJVBC2S

@misc{pith2026250617250,
  author       = {Pith},
  title        = {Pith review of: Towards Interpretable Adversarial Examples via Sparse Adversarial Attack},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPJVBC2S}},
  note         = {Machine review of arXiv:2506.17250}
}
read the original abstract

Sparse attacks are to optimize the magnitude of adversarial perturbations for fooling deep neural networks (DNNs) involving only a few perturbed pixels (i.e., under the l0 constraint), suitable for interpreting the vulnerability of DNNs. However, existing solutions fail to yield interpretable adversarial examples due to their poor sparsity. Worse still, they often struggle with heavy computational overhead, poor transferability, and weak attack strength. In this paper, we aim to develop a sparse attack for understanding the vulnerability of CNNs by minimizing the magnitude of initial perturbations under the l0 constraint, to overcome the existing drawbacks while achieving a fast, transferable, and strong attack to DNNs. In particular, a novel and theoretical sound parameterization technique is introduced to approximate the NP-hard l0 optimization problem, making directly optimizing sparse perturbations computationally feasible. Besides, a novel loss function is designed to augment initial perturbations by maximizing the adversary property and minimizing the number of perturbed pixels simultaneously. Extensive experiments are conducted to demonstrate that our approach, with theoretical performance guarantees, outperforms state-of-the-art sparse attacks in terms of computational overhead, transferability, and attack strength, expecting to serve as a benchmark for evaluating the robustness of DNNs. In addition, theoretical and empirical results validate that our approach yields sparser adversarial examples, empowering us to discover two categories of noises, i.e., "obscuring noise" and "leading noise", which will help interpret how adversarial perturbation misleads the classifiers into incorrect predictions. Our code is available at https://github.com/fudong03/SparseAttack.

Figures

Figures reproduced from arXiv: 2506.17250 by the authors.

Figure 1
Figure 1. Illustration of adversarial examples (AEs) computed by different attack algo [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Distribution of qa(x) under different values of a. As a approaches to 0, qa(x) increasingly resembles the Dirac delta function. Second, when x ̸= 0, we need to consider both the positive and negative directions of a, i.e., lima→0+ and lima→0− . Starting with the positive direction, let t = 1 a , we have lim a→0+ qa(x) = lima→0 1 a √ π exp−(x/a) 2 = limt→∞ √ πt · exp−x 2 t 2 = limt→∞ √ πt expx2t 2 = limt→∞ √ π 2x 2t … view at source ↗
Figure 3
Figure 3. Illustration of targeted attacks by sparse attack counterparts (Top) and our ap [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Illustration of how adversarial perturbations computed by our approach mislead [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Illustration of adversarial examples computed by our approach and other sparse [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Illustration of how adversarial perturbation computed by our approach misleads [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]
Figure 7
Figure 7. Figure 7: Illustration of how the hyperparameters a, λ and τ affect the computational com￾plexity (or fooling rate) and the sparsity. C.3 Ablation Studies We conduct experiments to show the impact of hyperparameters a, λ, and τ on perfor￾mance. We use VGG-16 to perform white-box…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 50 canonical work pages

  1. [1]

    Enabling scalable and adaptive machine learning training via serverless computing on public cloud.Performance Evaluation, 167:102451, 2025

    Ahsan Ali, Xiaolong Ma, Syed Zawad, Paarijaat Aditya, Istemi Ekin Akkus, Ruichuan Chen, Lei Yang, and Feng Yan. Enabling scalable and adaptive machine learning training via serverless computing on public cloud.Performance Evaluation, 167:102451, 2025

  2. [2]

    Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural net- works. InIEEE Symposium on Security and Privacy, 2017

  3. [3]

    Jordan, and Martin J

    Jianbo Chen, Michael I. Jordan, and Martin J. Wainwright. Hopskipjumpattack: A query- efficient decision-based attack. InIEEE Symposium on Security and Privacy, 2020, 2020

  4. [4]

    More data can expand the general- ization gap between adversarially robust and standard models

    Lin Chen, Yifei Min, Mingrui Zhang, and Amin Karbasi. More data can expand the general- ization gap between adversarially robust and standard models. InInternational Conference on Machine Learning (ICML), 2020

  5. [5]

    Bonemet: An open large-scale multi-modal murine dataset for breast cancer bone metastasis diagnosis and prognosis

    Tiankuo Chu, Fudong Lin, Shubo Wang, Jason Jiang, Wiley Jia-Wei Gong, Xu Yuan, and Liyun Wang. Bonemet: An open large-scale multi-modal murine dataset for breast cancer bone metastasis diagnosis and prognosis. InThe Thirteenth International Conference on Learning Representations (ICLR), 2025. 16 F. Lin, J. Lou, et al

  6. [6]

    Singh, Nicolas Flammarion, and Matthias Hein

    Francesco Croce, Maksym Andriushchenko, Naman D. Singh, Nicolas Flammarion, and Matthias Hein. Sparse-rs: A versatile framework for query-efficient sparse black-box ad- versarial attacks. InAAAI, 2022

  7. [7]

    Sparse and imperceivable adversarial attacks

    Francesco Croce and Matthias Hein. Sparse and imperceivable adversarial attacks. InICCV, 2019

  8. [8]

    BERT: pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. InConference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (NAACL-HLT), 2019

Show all 57 references
  1. [9]

    Greedyfool: Distortion-aware sparse adversarial attack

    Xiaoyi Dong, Dongdong Chen, Jianmin Bao, Chuan Qin, Lu Yuan, Weiming Zhang, Nenghai Yu, and Dong Chen. Greedyfool: Distortion-aware sparse adversarial attack. InNeurIPS, 2020

  2. [10]

    Boosting adversarial attacks with momentum

    Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. InConference on Computer Vision and Pattern Recognition (CVPR), 2018

  3. [11]

    Evading defenses to transferable ad- versarial examples by translation-invariant attacks

    Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable ad- versarial examples by translation-invariant attacks. InConference on Computer Vision and Pattern Recognition (CVPR), 2019

  4. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  5. [13]

    Robust physical-world attacks on deep learning visual classification

    Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. InCVPR, 2018

  6. [14]

    Sparse adversarial attack via perturbation factorization

    Yanbo Fan, Baoyuan Wu, Tuanhui Li, Yong Zhang, Mingyang Li, Zhifeng Li, and Yujiu Yang. Sparse adversarial attack via perturbation factorization. InECCV, 2020

  7. [15]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing ad- versarial examples. InInternational Conference on Learning Representations (ICLR), 2015

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

  9. [17]

    Interpretable minority synthesis for imbalanced classification

    Yi He, Fudong Lin, Xu Yuan, and Nian-Feng Tzeng. Interpretable minority synthesis for imbalanced classification. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 2542–2548, 2021

  10. [18]

    Highly accurate protein structure prediction with alphafold.Nature, 596(7873):583–589, 2021

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ron- neberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold.Nature, 596(7873):583–589, 2021

  11. [19]

    Physgan: Generating physical-world- resilient adversarial examples for autonomous driving

    Zelun Kong, Junfeng Guo, Ang Li, and Cong Liu. Physgan: Generating physical-world- resilient adversarial examples for autonomous driving. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

  12. [20]

    Learning multiple layers of features from tiny images.Technical Report, University of Toronto, 2009

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.Technical Report, University of Toronto, 2009

  13. [21]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. InAdvances in Neural Information Processing Systems (NIPS), 2012

  14. [22]

    Goodfellow, and Samy Bengio

    Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. InInternational Conference on Learning Representations (ICLR), 2017

  15. [23]

    Goodfellow, and Samy Bengio

    Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial machine learning at scale. In5th International Conference on Learning Representations (ICLR), 2017. Towards Interpretable Adversarial Examples via Sparse Adversarial Attack 17

  16. [24]

    Inan, Sahar Abdelnabi, Janardhan Kulkarni, Lukas Wutschitz, Reza Shokri, Christopher G

    Guangchen Lan, Huseyin A. Inan, Sahar Abdelnabi, Janardhan Kulkarni, Lukas Wutschitz, Reza Shokri, Christopher G. Brinton, and Robert Sim. Contextual integrity in LLMs via reasoning and reinforcement learning.arXiv preprint arXiv:2506.04245, 2025

  17. [25]

    MNIST handwritten digit database

    Yann LeCun and Corinna Cortes. MNIST handwritten digit database. 2010

  18. [26]

    Mmst-vit: Climate change-aware crop yield prediction via multi-modal spatial-temporal vi- sion transformer

    Fudong Lin, Summer Crawford, Kaleb Guillot, Yihe Zhang, Yan Chen, Xu Yuan, et al. Mmst-vit: Climate change-aware crop yield prediction via multi-modal spatial-temporal vi- sion transformer. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 5751–5761, 2023

  19. [27]

    Towards robust vision trans- former via masked adaptive ensemble

    Fudong Lin, Jiadong Lou, Xu Yuan, and Nian-Feng Tzeng. Towards robust vision trans- former via masked adaptive ensemble. InProceedings of the 33rd ACM International Con- ference on Information and Knowledge Management (CIKM), pages 1389–1399, 2024

  20. [28]

    Comprehensive transformer-based model architecture for real-world storm predic- tion

    Fudong Lin, Xu Yuan, Yihe Zhang, Purushottam Sigdel, Li Chen, Lu Peng, and Nian-Feng Tzeng. Comprehensive transformer-based model architecture for real-world storm predic- tion. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML-PKDD), p...

  21. [29]

    Delving into transferable adversarial examples and black-box attacks

    Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. InICLR, 2017

  22. [30]

    Learning sparse neural networks through l_0 regularization

    Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through l_0 regularization. InICLR, 2018

  23. [31]

    Malletrain: Deep neural networks training on unfillable supercomputer nodes

    Xiaolong Ma, Feng Yan, Lei Yang, Ian Foster, Michael E Papka, Zhengchun Liu, and Ra- jkumar Kettimuthu. Malletrain: Deep neural networks training on unfillable supercomputer nodes. InProceedings of the 15th ACM/SPEC International Conference on Performance Engineering, pages 19...

  24. [32]

    Towards deep learning models resistant to adversarial attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. InInternational Con- ference on Learning Representations (ICLR), 2018

  25. [33]

    The curious case of adversarially robust models: More data can help, double descend, or hurt generalization

    Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help, double descend, or hurt generalization. InUncertainty in Artificial Intelligence, 2021

  26. [34]

    Transferable structural sparse adver- sarial attack via exact group sparsity training

    Di Ming, Peng Ren, Yunlong Wang, and Xin Feng. Transferable structural sparse adver- sarial attack via exact group sparsity training. InComputer Vision and Pattern Recognition (CVPR), pages 24696–24705, 2024

  27. [35]

    Sparsefool: A few pixels make a big difference

    Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: A few pixels make a big difference. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

  28. [36]

    Deepfool: A sim- ple and accurate method to fool deep neural networks

    Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A sim- ple and accurate method to fool deep neural networks. InConference on Computer Vision and Pattern Recognition (CVPR), 2016

  29. [37]

    McDaniel, Somesh Jha, Matt Fredrikson, Z

    Nicolas Papernot, Patrick D. McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. InEuropean Symposium on Security and Privacy (EuroS&P), 2016

  30. [38]

    Fast minimum-norm adver- sarial attacks through adaptive norm constraints

    Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adver- sarial attacks through adaptive norm constraints. InNeural Information Processing Systems (NeurIPS), pages 20052–20062, 2021

  31. [39]

    Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences.Proceedings of the National Academy of Sciences, 118(15), 2021

    Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C Lawrence Zitnick, Jerry Ma, et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences.Proceedings of the National A...

  32. [40]

    Berg, and 18 F

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhi- heng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and 18 F. Lin, J. Lou, et al. Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge.International Jour...

  33. [41]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient- based localization. InICCV, 2017

  34. [42]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. InInternational Conference on Learning Representations (ICLR), 2015

  35. [43]

    One pixel attack for fooling deep neural networks.IEEE Trans

    Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks.IEEE Trans. Evol. Comput., 2019

  36. [44]

    Hybrid batch attacks: Finding black- box adversarial examples with limited queries

    Fnu Suya, Jianfeng Chi, David Evans, and Yuan Tian. Hybrid batch attacks: Finding black- box adversarial examples with limited queries. InUSENIX, 2020

  37. [45]

    Goodfellow, and Rob Fergus

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. InInternational Conference on Learning Representations (ICLR), 2014

  38. [46]

    Fooling automated surveillance cam- eras: Adversarial patches to attack person detection

    Simen Thys, Wiebe Van Ranst, and Toon Goedemé. Fooling automated surveillance cam- eras: Adversarial patches to attack person detection. InIEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

  39. [47]

    Goodfellow, Dan Boneh, and Patrick D

    Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian J. Goodfellow, Dan Boneh, and Patrick D. McDaniel. Ensemble adversarial training: Attacks and defenses. InInternational Conference on Learning Representations (ICLR), 2018

  40. [48]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems (NIPS), 2017

  41. [49]

    Brusleattack: a query-efficient score- based black-box sparse adversarial attack

    Viet Quoc V o, Ehsan Abbasnejad, and Damith Ranasinghe. Brusleattack: a query-efficient score- based black-box sparse adversarial attack. InInternational Conference on Learning Representations(ICLR), 2024

  42. [50]

    Heaviside step function, 2024

    Wikipedia. Heaviside step function, 2024. [Online; accessed 01-January-2024]

  43. [51]

    Black-box sparse adversarial attack via multi-objective optimisation

    Phoenix Neale Williams and Ke Li. Black-box sparse adversarial attack via multi-objective optimisation. InComputer Vision and Pattern Recognition (CVPR), 2023

  44. [52]

    Zico Kolter

    Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. InInternational Conference on Learning Representations (ICLR), 2020

  45. [53]

    Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

  46. [54]

    Structured adversarial attack: Towards general implementation and better interpretability

    Kaidi Xu, Sijia Liu, Pu Zhao, Pin-Yu Chen, Huan Zhang, Quanfu Fan, Deniz Erdogmus, Yanzhi Wang, and Xue Lin. Structured adversarial attack: Towards general implementation and better interpretability. InICLR, 2019

  47. [55]

    Fedcust: Offloading hyperparameter customization for federated learning.Perfor- mance Evaluation, 167:102450, 2025

    Syed Zawad, Xiaolong Ma, Jun Yi, Cheng Li, Minjia Zhang, Lei Yang, Feng Yan, and Yux- iong He. Fedcust: Offloading hyperparameter customization for federated learning.Perfor- mance Evaluation, 167:102450, 2025

  48. [56]

    Zeiler and Rob Fergus

    Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. InEuropean Conference on Computer Vision, 2014

  49. [57]

    pelican" ∥δ∥0 = 248 (b) GreedyFool. “goldfish

    Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. InICML, 2021. Towards Interpretable Adversarial Examples via Sparse Adversarial Attack 19 Outline This document supplements the main paper in three aspects. F...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.