acceptodds
Withdrawn from ICLR 2027

Query Efficient Black-box Adversarial Attack via Dataset Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Score-based black-box attacks seek effective adversarial perturbations under limited access to the victim model, making query efficiency a central challenge. Existing methods often improve query efficiency by incorporating prior information, such as structured search priors or surrogate models to guide the search toward promising perturbation directions. However, the effectiveness of such guidance depends strongly on the quality and availability of the underlying prior. To provide a richer and alternative source of guidance, we propose DDAttack, which leverages dataset distillation (DD) as prior information for gradient-free adversarial search. DD condenses class-discriminative information from large datasets into a compact set of synthetic examples, providing informative input-space cues that can guide adversarial optimization. DDAttack exploits this distilled information through two complementary stages: distillation-guided initialization, which constructs a promising starting perturbation from distilled examples, and distillation-guided search, which aggregates promising candidate perturbations to guide subsequent updates. Extensive experiments on adversarially trained models demonstrate that distilled examples provide effective prior information for score-based black-box attacks. The source code is included in the supplementary material and will be publicly available upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial robustness, Score-based black-box attack, Dataset distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tL1jOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents DDAttack - a score-based black box adversarial attack on image classifiers that uses dataset distillation to guide the initialization and the procedure of searching adversarial examples. The authors argue that the class-information available in distilled examples helps to determine better search detection which results in higher attack success and better query efficiency. They support their claim with experiments on ImageNet across several robust models and compare their results to existing score-based black box attacks. They also study the dependence on the distillation approach and the hyperparameters of their method.

Critical strengths
  1. Up to my knowledge usage of data distillation for black box adversarial attacks is a novel direction relevant for the community. The paper is clearly written.

  2. The paper supports its claims with a broad set of experiments on multiple robust models. The improvement appears to be consistent even if it is not big. The authors provide a reproducibility statement and share the code for experiments.

Critical weaknesses
  1. The main concern is that proposed method relies on having a distilled dataset which is a strong requirement. Besides that, DDAttack borrows significant parts of its update structure from the Square Attack (Section 3.3). Despite the factors of a significantly stronger requirement and an overlap with the Square Attack method, the success rate improvement on top of Square Attack which is demonstrated in the Tables 1 and 2 is quite moderate and is mostly below 2 percent points. Therefore, even though the improvement is rather consistent, the value for the black-box robustness evaluation field appears to be limited.

  2. The evaluation is limited. It only experiments with the transformer-based architecture (ViT, Swin). Experiments with CNNs would be valuable especially given that the square-form of the Square Attack update was shown to exploit the CNN filter structures and its performance might be better there. Also the paper is considering only untargeted attacks and only in the ℓ∞\ell_\infty setting. Targeted attacks and ℓ2\ell_2 setting are not considered.

  3. The method uses multiple hyperparameters in the initialization and search algorithms. Even though the Appendix presents multiple ablation studies experiments, the authors acknowledge that the number of competing classes LL has significant impact on the method efficiency (lines 787 - 795 in the Appendix B). Therefore, it is unclear whether the method generalizes beyond ImageNet with the same hyperparameters or additional hyperparameter tuning is required.

Questions for authors
  1. Can DDAttack be adapted to the targeted attack setting i. e. perturbing the class to the specific targeted class, not any other class than the predicted one? It is much more challenging for ImageNet which has many classes than the untargeted setting which was considered in the paper.
Reviewer_n4cXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DDAttack, a score-based black-box attack that uses precomputed distilled images as a search prior. It selects references from classes that the victim model considers competitive, constructs and evaluates candidate perturbations, then uses the best candidates to initialize and guide a Square-Attack-style search. Under a fixed query budget, DDAttack achieves lower robust accuracy than the reported Square Attack baseline across eight robust ImageNet models and two ℓ∞\ell_\infty perturbation budgets.

Critical strengths
  1. Using distilled data as a prior for query-based attacks is an interesting and practical idea. The method improves on the reported Square Attack baseline in every tested model and perturbation setting, with the initialization queries included in the total budget.
  2. The paper evaluates the two attack stages separately and compares different reference sources. These results offer initial evidence that the choice of prior matters, although they do not yet identify which property of distilled data drives the improvement.
Critical weaknesses
  1. The role of the distilled prior is not clear enough. SRe2L and RDED use pretrained ImageNet models to construct their reference images, but the paper does not identify the exact released variants, generating models, or number of images per class. It should also explain what information an attacker would need in a setting without a suitable public distilled dataset.
  2. Table 4 compares distilled images with randomly selected real images. Because RDED uses an observer model to select and combine real crops, this comparison does not show whether the benefit comes from distillation or from that selection process. A comparison using observer-selected real crops would address this gap. Using the model behind the prior as a surrogate would offer further context, provided the additional model access is made explicit.
  3. Several attack choices are not separated from the distilled prior. The margin-dependent patch size in Eq. 12 can be used without distilled images, yet has no separate ablation. The Square baseline uses an initial patch fraction of 0.8 without a tuning comparison. The paper also does not test random-class references or non-distilled candidates under the same best-of-100 selection. Even the random-initialization variant in Table 3 retains guidance from distilled candidates.
  4. Parameter sensitivity is studied on Swin-B MIMIR, which is also the test victim with the largest reported gain. Results from a separate tuning model would make it easier to judge whether the selected settings generalize.
  5. The gains over Square average 1.67 percentage points at 8/255 and 1.43 at 4/255. The attack is randomized, but the main results lack repeated-run uncertainty estimates. This matters particularly for ablation conclusions based on differences of 0.16 to 0.58 points.
  6. The query-efficiency evidence is incomplete. Figure 1 shows curves for two victims, and Table 5 averages queries over each method’s own successful images. Medians and results on images that both methods attack successfully would permit a direct comparison. Smaller query budgets are also important because initialization alone uses 101 queries.
  7. The evaluation is limited to untargeted ℓ∞ attacks on robust ImageNet ViT and Swin models. Its conclusions may not extend to CNNs, other datasets, different attack objectives, or settings where the distilled prior differs from the victim’s training distribution. Table 1 also describes 8/255 as a standard RobustBench setting, although the ImageNet ℓ∞\ell_\infty leaderboard uses 4/255. The surrogate comparison in Table 7 is less informative because Square already outperforms every reported surrogate configuration by 3.9 to 7.3 points.

Minor:

  1. The role of the cosine term in Eq. 6 needs clarification because its gradient is zero at initialization. The effects of its weight, the step count, and the step size are not examined.
  2. Appendix A leaves some conditions untested. Figure 2 varies candidate count and remaining query budget together, while Table 6 measures final perturbations rather than the proposed guidance direction.
  3. The low NES and ZOsign-SGD success rates need fuller implementation details. The stated ResNet-151 surrogate should be checked.
  4. Appendix A announces three questions but lists two. The variable NN in Eq. 33 is undefined, and the reported 1.67 percent improvement should be stated in percentage points.
  5. The related-work discussion should consider attacks that initialize query search with local-model candidates [1], learned proposal schedules for Square Attack [2], and analyses of how auxiliary data and attacker resources affect black-box comparisons [3].

References:

[1] Suya, Fnu, Jianfeng Chi, David Evans, and Yuan Tian. Hybrid batch attacks: Finding black-box adversarial examples with limited queries. In USENIX Security , 2020.

[2] Yatsura, Maksym, Jan Metzen, and Matthias Hein. Meta-learning the search distribution of black-box random search based adversarial attacks. In NeurIPS, 2021.

[3] Suya, Fnu, Anshuman Suri, Tingwei Zhang, Jingtao Hong, Yuan Tian, and David Evans. Sok: Pitfalls in evaluating black-box attacks. In SaTML, 2024.

Questions for authors
  1. Which released RDED and SRe2L variants were used, including their IPC and generating models? What data, model access, and offline cost would be required in a setting without a suitable public distilled set?
  2. Under the same candidate construction and query budget, how does RDED compare with real crops selected by its observer model? If the generating model is available, a surrogate-based result would provide additional context, provided its extra model access is stated.
  3. How much of the gain remains when references come from random classes, non-distilled candidates undergo the same best-of-100 selection, and Square Attack uses the adaptive patch-size rule alone? How sensitive is the Square baseline to its initial patch size?
  4. Can the authors report paired uncertainty estimates for the robust-accuracy gains, query counts on images both methods attack successfully, and performance at smaller budgets above the 101-query initialization cost?
Reviewer_sACUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DDAttack. It is a score-based black-box adversarial attack that leverages dataset distillation as a source of prior information to improve query efficiency. The method uses distilled samples to construct and select promising initial perturbations, and then guides subsequent adversarial search through weighted direction aggregation and adaptive patch updates. Experiments on ImageNet against adversarially robust models show improved attack performance over existing black-box attacks under fixed query budgets.

Critical strengths

The proposed method consistently outperforms representative black-box attack baselines under fixed query budgets.

Critical weaknesses
  1. Lack of fundamental novelty. The idea of leveraging distilled knowledge for adversarial attacks is not new, and simply applying dataset distillation to score-based black-box attacks does not automatically constitute a novel contribution. The proposed method mainly combines reference-guided perturbation generation with existing search strategies, without identifying any new adversarial structure, principle, or mechanism specific to dataset distillation.

  2. The paper does not adequately clarify the scope and significance of its contribution, especially given its reliance on additional distilled data and score-based feedback. The related work lacks a comprehensive discussion of closely related ideas across different attack settings, while the experiments mainly demonstrate performance improvements without directly validating the proposed mechanism or isolating the actual source of improvement.

Questions for authors
  1. What is the threat model? particularly the assumption that attackers have access to precomputed class-labeled distilled datasets? How realistic is this assumption in practical black-box attack scenarios?

  2. How much offline computation and auxiliary data are required to obtain the distilled samples? How would the claimed efficiency improvements change if these costs were taken into account?

  3. Why is the proposed method specifically designed for score-based attacks? Could the distilled priors also be useful in decision-based or other adversarial attack settings?

  4. Could you provide more fine-grained ablation studies to separate the contributions of margin-weighted aggregation, adaptive patch sizing, and guidance probability decay?

  5. Are the average query counts computed over the same successfully attacked samples across different methods? Could you provide a comparison on matched samples to ensure a fairer evaluation of query efficiency?

  6. How well does the proposed method generalize beyond ImageNet and the evaluated robust Transformer models? Could the authors provide evidence across different datasets or model families?

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

Related papers

Discussion (0)

Sign in to comment.