acceptodds
Under review as a conference paper at ICLR 2027

Adversarial Scenario Attack: Query-Efficient Black-Box Generation of Natural Adversarial Examples

Abstract

Generating natural adversarial examples (NAEs) from a given image remains challenging under tight black-box query budgets, as many existing generative attacks rely on surrogate-model guidance or attack-specific training. We propose Adversarial Scenario Attack (ASA), a query-based black-box attack on image classifiers that searches over human-readable natural-language editing scenarios rather than pixel-level perturbations. ASA uses a multimodal large language model (MLLM) to propose image-dependent changes to backgrounds, weather, and surface appearance, and a generative editor to apply them to the original image. Winner–loser feedback guides subsequent proposals, while a Greedy Explorer composes complementary scenarios and retains only combinations that improve the attack score. Under a budget of 100 victim-model queries per image, ASA achieves the highest attack success rate on all ten evaluated ImageNet classifiers, including two vision–language models, without surrogate models or attack-specific training. On average, its success rate within 20 queries already exceeds that of every baseline at 100 queries, while it maintains competitive image-level transferability. Because each attack is expressed as a readable scenario, the discovered edits can also be reused: class-neutral scenarios that recur during search, such as light mist or dusk lighting, raise the average error rate to 2.0–3.3 times that of a control prompt when applied unchanged to images from all 1,000 ImageNet classes across 13 classifiers.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Natural Adversarial Examples, Black-box, Query-based, Natural Transformation Vulnerabilities

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 4 reviews, above 27% of the others.
Reviewer_WA3cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Adversarial Scenario Attack (ASA), a score-based black-box attack that generates natural adversarial examples by searching over natural-language editing scenarios rather than pixel-level perturbations. ASA uses an MLLM to propose edits in three categories—background, weather/atmosphere, and material/color—and applies them with a text-guided image editor. Victim-model feedback is used to iteratively refine the scenarios, while a Greedy Explorer combines complementary edits. The method achieves strong attack success under a 100-query budget.

Critical strengths
  • ASA achieves high ASR across diverse architectures, including CNNs, Transformers, SSMs, robust models, and VLMs, while using relatively few victim-model queries.

  • The attack outputs are interpretable and natural-looking.

  • Some class-neutral scenarios remain effective across many ImageNet classes and multiple model families, suggesting that these models may share common vulnerabilities to certain natural scene changes.

Critical weaknesses
  • The high-level idea of moving from pixel-level perturbations to natural-language-controlled semantic or scenario-level modifications does not seem fundamentally new. Prior work [a] has already explored generating natural adversarial examples by manipulating human-interpretable attributes through text prompts. ASA follows a similar formulation, with the main additions being victim-model feedback and MLLM-based iterative refinement of the editing scenarios. Therefore, I would view ASA more as an automated, closed-loop extension of existing semantic-prompt attacks rather than a fundamentally new attack paradigm.

  • A fairer comparison with query-based unrestricted attacks is needed to show that the gain really comes from ASA rather than its much larger perturbation space.

  • In the human study, only 75.4% of judgments considered the edited object to still be the same as the original one. This means a non-trivial fraction of “successful attacks” may simply change the image too much, which weakens the validity of the reported ASR as natural adversarial examples. The human study is also only conducted on ResNet-50 examples.

  • “Query-efficient” does not necessarily mean computationally efficient. ASA reduces victim-model queries, but each candidate still requires MLLM inference and large-scale image editing. The appendix shows that ASA has a higher per-image generation cost and a higher total runtime than AdvFlow, DifAttack, and CGAttack. Therefore, the efficiency claim should be understood mainly as query efficiency, not end-to-end efficiency.

  • ASA assumes access to the ground-truth label and confidence scores from the victim model. In many real black-box APIs, only the top-1 prediction may be available. For MLLM victims, the method even relies on self-reported confidence values, which the paper acknowledges are neither internal logits nor calibrated probabilities. This limits the practical generality of the attack setting.

[a] Intriguing Properties of Diffusion Models: An Empirical Study of the Natural Attack Capability in Text-to-Image Generative Models, CVPR 2024

Questions for authors

see Weaknesses

3 comments

We thank the reviewer for the constructive review. We address each point below. All table, figure, section, and appendix numbers refer to the revised version.


W1. Relation to [a].

ASA's contribution is a structured search over natural-language edits of a source image, proposed by an MLLM. It has two components:

  1. Scenario competition (S). Strategy-specific scenarios compete on victim scores, and winner–loser feedback from each competition guides the subsequent proposals.
  2. Greedy composition (G). A combination of edits is retained only when it improves the attack objective.

Victim score feedback and MLLM-based iterative refinement alone do not account for ASA's gains. ASA w/o S,G keeps both, with the same MLLM, image editor, and 100-query budget, and removes only S and G. It falls 37.7–53.6 pp below ASA across the four victims (largest on ConvNeXt: 33.6% vs. 87.2%; Fig. 4). Removing G alone (ASA w/o G) costs 7.7–16.6 pp.

ASA does not follow [a]'s formulation. The two differ in four respects:

Aspect[a]ASA (Ours)
Use of model outputsFixed text prompts, with no iterative optimization against a victimVictim scores guide an iterative scenario search
Starting pointClass-descriptive text promptExisting image
ManipulationRemoval or alteration of class-defining featuresEdits to background, weather, and surface appearance, intended to leave class-defining features unchanged
Intended human–model discrepancyHumans should not identify the class; the detector doesThe original class should be preserved for humans; the victim misclassifies the image

We have added [a] to the related work (Sec. 2.1).

[a] Intriguing Properties of Diffusion Models: An Empirical Study of the Natural Attack Capability in Text-to-Image Generative Models, CVPR 2024


W2. A comparison with query-based unrestricted attacks.

ASA is not strong merely because it is unbounded: it attacks along a different axis, images of scenes that could occur in the real world (our definition of natural, Sec. 3). Two comparisons in this unrestricted setting support this. Since, to our knowledge, no score-based query attack is unrestricted, we removed the bound from the Table 1 baselines, and ASA leads them by at least 42 pp (a). ASA's own ablation is a second unrestricted attack, and it falls at least 37 pp short of ASA (b).

(a) The Table 1 baselines with the bound removed (new). We re-ran all four baselines under the following setting:

  • Bound. $\epsilon = 255/255$, so the $\ell_\infty$ bound no longer excludes any image.
  • Budget and data. 100 queries, on the same images and 10 victims as Table 1.
  • Success criterion. A flip counts as a success only if ASA's Gemma-4-E4B verifier, with the same prompt, accepts the image as a natural result that preserves the labelled object. After a rejection, the attack continues with its remaining queries, and, as for ASA, verifier calls are not charged to the query budget.
  • DifAttack. $\sigma = 0.5$, its noise scale for the unbounded setting.
AttackRN-50ConvNeXtSwinDeiTViMMVAdv-ResAdv-IncCLIPSigLIP2
AdvFlow4.293.583.553.814.183.323.583.645.273.14
DifAttack30.4724.6425.9120.5520.2419.2923.9628.1225.3224.91
CGAttack2.361.942.402.652.462.282.681.942.442.44
MCGAttack1.291.021.041.171.181.351.231.091.161.28
ASA94.1087.2289.3488.0390.9077.9095.8797.6984.1967.17

Every baseline flips the label on 83.9–100% of the images, but the verifier rejects more than 99% of the candidates it checks. Without the bound, pixel-space attacks can therefore change the label, but not with images that could occur in the real world: enlarging $\epsilon$ does not reach this axis.

(b) An unrestricted attack built from ASA's components. ASA w/o S,G (Sec. 4.8, Fig. 4; see W1) differs from ASA only in lacking scenario competition (S) and greedy composition (G), so it is unrestricted in the same way. It falls 37.7–53.6 pp short of ASA across the four victims of Fig. 4.

We have added (a) to Tab. 1, Fig. 3(b), and Sec. 4.2.1, with full results in App. L.

W3. Human study.

We ran a new, larger human study with the majority-vote rule of [b]: a majority of seven raters judged 84.4% of all ASA successes on ResNet-50 (740/877), and 88% of a 100-image Swin-B sample, to be natural-looking and to show the original class.

Protocol.

  • Raters and images. Seven non-author participants each rated all 877 ASA images that fool ResNet-50 on NIPS 2017.
  • Presentation. One image at a time, without the original and without being told that the images were edited. Image order and the order of the two class labels in Q2 were randomized per participant.
  • Q1. "Does this image look like a natural photograph of a real scene?" (yes/no).
  • Q2. "Which label best describes the main object?" (original class, post-attack class, or "unsure").

The two questions are the two conditions of our definition of natural (Sec. 3): a scene that could occur in the real world, whose main object still belongs to the original class.

ResNet-50. By per-image majority (at least 4 of 7 raters), 97.3% were judged natural, 86.9% were assigned the original class, and 84.4% satisfied both.

Swin-B. The same participants rated 100 Swin-B adversarial images, with originals drawn at random from those on which ASA fooled both victims. 88% satisfied both criteria versus 81% for the ResNet-50 images of the same originals (paired difference +7.0 pp, 95% CI −2.0 to +16.0): on the same originals, the rate is not lower on a second architecture.

Effect on Tab. 1. We multiplied ASA's ASR on every victim by the ResNet-50 rate (740/877) and left every baseline undiscounted. ASA still ranks first:

  • vs. the unrestricted baselines ($\epsilon = 255/255$, W2): on all ten victims, by 49 pp on average (73.6% vs. 24.3%) and by at least 31 pp on every victim.
  • vs. the $\epsilon = 12/255$ baselines: on nine of ten victims, by 30 pp on average (73.6% vs. 43.4%).
RN-50ConvNeXtSwinDeiTViMMambaVisionAdv-ResAdv-IncCLIPSigLIP2Mean
Best baseline, $\epsilon = 12/255$ (Tab. 1)82.6249.3923.9326.5932.1220.8544.3052.7857.2043.8943.37
Best baseline, $\epsilon = 255/255$ (W2)30.4724.6425.9120.5520.2419.2923.9628.1225.3224.9124.34
ASA $\times$ ResNet-50 human-verified rate79.4073.5975.3874.2876.7065.7380.8982.4371.0456.6873.61

ASR (%). Last row: ASA's ASR × 740/877. Bold: highest in each column.

We have added the study to Sec. 4.6, the Swin-B results to App. O, and the table above to App. M.

[b] Constructing Unrestricted Adversarial Examples with Generative Models. NeurIPS 2018.


W4. Computational efficiency.

Our efficiency claim concerns queries, and the revision states this explicitly (App. D). The per-image time during the search (25.98 s, Tab. 14) is that of an off-the-shelf MLLM and image editor generating a scene-level change, and App. D lists it as a limitation for large-scale evaluation. In runtime, ASA also has advantages in three respects:

  • No attack-specific training. ASA requires none; the baselines need 651.6–69,343.2 s before generating (Tab. 14).
  • Lower total runtime at small scale. Because the baselines pay this cost up front, ASA's total runtime is below that of AdvFlow, DifAttack, CGAttack, and MCGAttack up to about 25, 190, 620, and 2,790 images, respectively (from Tab. 14).
  • Fast reuse. In the prompt-level transferability test (Sec. 4.3), scenarios found in the RN-50 search are applied to new images without search, at about 0.7 s per image, within the baselines' range. The three most frequent ones still reach 20.75–34.62% ASR over all 1,000 classes and 13 models (Tab. 3).

W5. Threat model

Ground-truth label. ASA makes the same label assumption as the baselines we compare with, and the assumption is removable. Wherever $y$ is used (Sec. 3.1–3.2), it can be replaced by the victim's prediction on the clean image, which equals $y$ on the clean-correct samples we evaluate and costs one extra query.

Confidence scores: classifiers. ASA is a score-based attack (Sec. 3.1), and score access is available in practice: commercial classifier APIs such as Google Cloud Vision return confidence scores, and score-based attacks have been demonstrated on them [c].

Confidence scores: MLLMs. On MLLM victims, ASA needs no provider-released scores. The victim is asked in text for its confidences, which are neither logits nor calibrated probabilities, and these suffice. We ran the protocol of App. J (100 clean-correct NIPS 2017 images, 100 queries) on the closed-source gpt-6-luna:

VictimASR (%)Avg. Q
Gemma-4-E4B76.011.14
Qwen3-VL-4B-Instruct83.08.71
gpt-6-luna (closed-source, new)78.012.53

Avg. Q: attack queries per successful image.

We have added this result to Sec. 4.7 (details in App. J).

[c] BruSLeAttack: A Query-Efficient Score-Based Black-Box Sparse Adversarial Attack. ICLR 2024.

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

The paper proposes Adversarial Scenario Attack (ASA), a query-based black-box attack that generates natural adversarial examples by editing a given image through natural-language scenarios. ASA uses an MLLM to propose image-dependent changes to backgrounds, weather, and surface appearance, and a generative editor to apply them to the original image. A Greedy Explorer combines or replaces scenario components and retains modifications that increase the attack score. Experimental results show that ASA achieves high attack success rates with low query cost and competitive image-level transferability.

Critical strengths
  1. The framework is interpretable, requires no attack-specific training, and achieves strong reported attack success across diverse classifiers.

  2. The prompt-transfer and noise-sensitivity results are interesting, and the component ablations provide evidence that the proposed search improves effectiveness.

Critical weaknesses
  1. Limited algorithmic novelty. Although empirically effective, the method appears to offer limited algorithmic novelty. Equations (1)–(3) describe a straightforward greedy search that retains promising candidates, sequentially combines or replaces components, and accepts modifications only when they improve the attack score. The contribution beyond this combination of standard search operations could be more clearly established.

  2. Insufficient validation of visual naturalness. The evaluation does not directly substantiate the claimed visual naturalness. In Figure 2, the red-painted car is transformed into a highly reflective chrome-like vehicle. Although this transformation is not inherently physically implausible, its visibly edited appearance highlights the need for a clearer operational definition of naturalness. The verifier in Appendix B primarily assesses object preservation and semantic plausibility while explicitly tolerating minor artifacts and lighting inconsistencies. Likewise, the metrics in Section 4.5 and Appendix H assess similarity, identity, distributional fidelity, or perceptual quality, rather than directly establishing photographic realism. The human study in Section 4.6 asks whether paired images depict the same object; its 75.4% positive-response rate therefore measures perceived identity preservation, not visual realism or retention of the original fine-grained category. Consequently, it remains unclear how much of the reported attack success corresponds to genuinely category-preserving, natural-looking images.

  3. High per-image generation cost. Despite its query efficiency, ASA incurs substantially higher generation latency. Appendix I reports an average generation time of 25.98 seconds per image, compared with 0.14–1.13 seconds for the evaluated baselines. Although ASA avoids attack-specific training, this per-image overhead may limit its practicality for large-scale robustness evaluation and highlights the distinction between query efficiency and computational efficiency.

Questions for authors
  1. How does ASA perform on nighttime or low-light images? Although scenario generation is image-conditioned, recurring edits such as dusk lighting may be less effective when the source image is already dark. Limited visibility may also make it difficult to determine whether generative editing preserves the original object or introduces details that cannot be verified from the source image. Could the authors provide a small-scale analysis of nighttime or low-light images that are correctly classified before editing, reporting attack success rate, query cost, and preservation of the original category?

  2. How well does ASA generalize to closed-source MLLMs? Appendix J evaluates two small open-source MLLMs, but its effectiveness against closed-source models remains unclear. Could the authors discuss its applicability in this setting, particularly when internal classification scores are unavailable, and provide a limited evaluation if feasible?

  3. What motivates the choice of background shifts, weather and atmosphere changes, and material and color modifications as the three editing strategies? Were these categories selected empirically, or according to specific criteria for naturalness and category preservation? Could ASA accommodate additional strategies, such as viewpoint changes, mild geometric transformations, or partial occlusion?

2 comments

We thank the reviewer for the constructive review. We address each point below. All table, figure, section, and appendix numbers refer to the revised version.


W1. Limited algorithmic novelty.

ASA's contribution is to search for a victim's weakness in a form that can be read and reused: an editing scenario that an image editor executes. The search has two components:

  1. Scenario competition (S). In each round, strategy-specific scenarios compete on the same image. The victim's scores show the MLLM which scenario the victim is weak to and which it is not (winner–loser feedback); the next proposals refine the former and avoid the traits of the latter.
  2. Greedy composition (G). Because each scenario is an instruction, two that were evaluated separately can be joined and rendered as one edit (Eqs. (1)–(3)).

Difference from diffusion-based attacks. DiffAttack, ACA, and NatADiff (App. E, Tab. 11) optimize diffusion latents or sampling trajectories with surrogate or classifier gradients. ASA needs only the victim's confidence scores and runs directly against a black-box victim.

Two results in the paper support this contribution.

Where the gain comes from. Both components add to ASR (Sec. 4.8, Fig. 4):

ASR gain (pp)RN-50SwinViMConvNeXt
S (ASA w/o S,G → ASA w/o G)+30.0+34.9+30.2+40.1
G (ASA w/o G → ASA)+7.7+16.6+15.2+13.5

What the search returns. A weakness shared across images and models: the three scenarios that recur most often in the search against RN-50 alone, applied unchanged to images of all 1,000 classes, reach an average ASR of 20.75–34.62% over 13 models, whereas the most frequent scenario that never succeeded in the same search stays at 10.44% (Sec. 4.3, Tab. 3).

We have revised Sec. 1 and Sec. 3.3 to state the contribution this way.


W2. Validation of visual naturalness.

We conducted a new human study with the majority-vote rule of [R1]: a majority of seven raters judged 84.4% of all ASA successes on ResNet-50 (740/877), and 88% of a 100-image Swin-B sample, to look like a photograph of a real scene and to show the original class.

Operational definition (Sec. 3.1). An adversarial example is natural if (i) it looks like a photograph of a scene that could occur in the real world and (ii) its main object still belongs to the original class. Both conditions are judged by humans who see the image on its own.

Protocol (Sec. 4.6).

  • Raters and images. Seven non-author participants each rated all 877 ASA images that fool ResNet-50 on NIPS 2017.
  • Presentation. One image at a time, without the original and without being told that the images were edited. Image order and the order of the two class labels in Q2 were randomized per participant.
  • Q1, condition (i). "Does this image look like a natural photograph of a real scene?" (yes/no).
  • Q2, condition (ii). "Which label best describes the main object?" (original class, post-attack class, or "unsure").

ResNet-50. By per-image majority (at least 4 of 7 raters), 97.3% met condition (i), 86.9% met condition (ii), and 84.4% met both.

Swin-B (App. O). The same raters rated 100 Swin-B examples: 88% met both conditions, versus 81% for the ResNet-50 examples of the same originals, so the rate is not specific to ResNet-50.

Effect on Tab. 1 (sensitivity analysis, App. M). We multiplied ASA's ASR on every victim by the ResNet-50 rate (740/877) and left every baseline uncorrected. ASA stays above:

  • the best $\epsilon = 12/255$ baseline on nine of the ten victims;
  • the best baseline without the bound ($\epsilon = 255/255$, Sec. 4.2.1) on all ten, by at least 31 pp.
ASR (%)RN-50ConvNeXtSwinDeiTViMMambaVisionAdv-ResAdv-IncCLIPSigLIP2Mean
Best baseline, $\epsilon = 12/255$82.6249.3923.9326.5932.1220.8544.3052.7857.2043.8943.37
Best baseline, $\epsilon = 255/255$30.4724.6425.9120.5520.2419.2923.9628.1225.3224.9124.34
ASA, projected with the ResNet-50 rate79.4073.5975.3874.2876.7065.7380.8982.4371.0456.6873.61

Bold: highest in each column.

[R1] Constructing Unrestricted Adversarial Examples with Generative Models. NeurIPS 2018.

W3. Computational efficiency.

Our efficiency claim concerns queries, and the revision states this explicitly (App. D). The per-image time during the search (25.98 s, Tab. 14) is that of an off-the-shelf MLLM and image editor generating a scene-level change, and App. D lists it as a limitation for large-scale evaluation. In runtime, ASA also has advantages in three respects:

  • No attack-specific training. ASA requires none; the baselines need 651.6–69,343.2 s before generating (Tab. 14).
  • Lower total runtime at small scale. Because the baselines pay this cost up front, ASA's total runtime is below that of AdvFlow, DifAttack, CGAttack, and MCGAttack up to about 25, 190, 620, and 2,790 images, respectively (from Tab. 14).
  • Fast reuse. In the prompt-level transferability test (Sec. 4.3), scenarios found in the RN-50 search are applied to new images without search, at about 0.7 s per image, within the baselines' range. The three most frequent ones still reach 20.75–34.62% ASR over all 1,000 classes and 13 models (Tab. 3).

Q1. Nighttime and low-light images.

We ran ASA on 100 low-light ImageNet images per victim. It fools ResNet-50 on 89.0% and Swin on 73.0% of them.

Setup (new). ASA runs with the paper's configuration and a budget of 100 queries. The images were selected as follows:

  1. We ranked all 50,000 ImageNet validation images by mean luminance.
  2. We screened the 336 darkest, without access to model predictions, to keep genuine low-light scenes (night, dim interiors, candle, fire, or stage lighting).
  3. We kept those the victim classifies correctly.
  4. We took the 100 darkest, with at most 5 per class (mean luminance below 0.10 on a 0–1 scale).
VictimASR (%)Avg. Q
ResNet-5089.05.99
Swin73.08.40

Low-light ImageNet, 100 images per victim. Avg. Q is averaged over successful attacks.

Preservation of the original category. The seven raters of W2 judged the 89 ResNet-50 successes under the same protocol and majority rule: 70 of the 89 (78.7%) met both conditions.

Limitation. ASA is weaker on low-light sources in two respects:

  • ASR is below that on NIPS 2017 (94.10% and 89.34%, Tab. 1). Lighting edits such as dusk are less effective on dark sources: a single Weather & Atmosphere scenario flips ResNet-50 in 10.1% of trials, against 16.1% on NIPS 2017.
  • The human-validated rate is below that on NIPS 2017 (84.4%, W2). With little of the object visible in the source, the edited image is less often recognized as the original class.

We have added the results to App. K and the limitation to App. D.


Q2. Generalization to closed-source MLLMs.

We ran ASA against the closed-source gpt-6-luna, and it succeeds there as it does on the two open-source MLLMs of Sec. 4.7, reaching 78.0% ASR.

Setup. We sampled 100 images with random seed 0 from gpt-6-luna's clean-correct NIPS 2017 images and allowed 100 queries per image. As for the open-source MLLMs, the only victim output used is its text answer: the top-two classes with self-reported confidences, not internal scores.

VictimASR (%)Avg. Q
Gemma-4-E4B76.011.14
Qwen3-VL-4B-Instruct83.08.71
gpt-6-luna (closed-source, new)78.012.53

Avg. Q: attack queries per successful image.

We have added this result to Sec. 4.7 (details in App. J).


Q3. Motivation and scalability.

Motivation. We chose the three strategies so that their edits can meet both conditions of our definition of natural (Sec. 3.1, see W2): (i) they are variations that occur, and vary widely, in real scenes, and (ii) they can be specified without changing the object's geometry, which is intended to preserve its class. Background Shift, Weather & Atmosphere, and Material & Color Reskinning satisfy both.

Scalability. ASA is not restricted to these three: the search is defined over a strategy set $\mathcal{R}$ (Sec. 3.2), and a strategy is added by adding its instruction to the proposal prompt. The image editor is likewise interchangeable (Tab. 7, three editors). Viewpoint changes, mild geometric transformations, and partial occlusion would enter as follows:

  • Viewpoint and mild geometric changes require re-rendering the object while keeping its identity. They could be supported by replacing FLUX.2 with a generative model built for viewpoint control, or with a 3D scene model.
  • Partial occlusion can be added as a fourth strategy with the current editor, as long as the result still meets both conditions.
Reviewer_42vVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Adversarial Scenario Attack (ASA) to generate natural adversarial examples in black-box settings. ASA uses a multimodal large language model to generate editing prompts and combines different scenarios through greedy search. Experiments show high attack success rates and strong query efficiency across different image classifiers.

Critical strengths
  1. The paper presents a new approach to black-box attacks through natural language editing prompts.
  2. The method achieves high attack success rates across ten image classifiers with a limited query budget.
Critical weaknesses
  1. The comparison between ASA and existing black-box attacks is not entirely fair. AdvFlow, DifAttack, CGAttack, and MCGAttack use an LpL_p perturbation bound of ϵ=12/255\epsilon=12/255, while ASA does not have this constraint. This gives ASA a much larger search space and allows it to make more substantial changes to the original images. Therefore, the improvement in attack success rate may come from the different attack settings rather than the proposed method itself. A comparison with unrestricted black-box attacks would make the experimental results more convincing.

  2. The Greedy Explorer combines different editing scenarios and keeps a combination only when it improves the attack score. This is a relatively simple search strategy, and its advantage over other search methods remains unclear. The ablation study in Section 4.7 only compares ASA with and without the Greedy Explorer. Although the results show the benefit of scenario combination, they do not demonstrate the advantage of the proposed greedy strategy over other alternatives. Comparisons with random combination, beam search, or direct scenario generation by the MLLM under the same query budget would better support the method design.

Questions for authors

Please see the Weaknesses.

1 comment

We thank the reviewer for the constructive review. We address each point below. All table, figure, section, and appendix numbers refer to the revised version.


W1. A comparison with query-based unrestricted attacks.

ASA is not strong merely because it is unbounded: it attacks along a different axis, images of scenes that could occur in the real world (our definition of natural, Sec. 3). Two comparisons in this unrestricted setting support this. Since, to our knowledge, no score-based query attack is unrestricted, we removed the bound from the Table 1 baselines, and ASA leads them by at least 42 pp (a). ASA's own ablation is a second unrestricted attack, and it falls at least 37 pp short of ASA (b).

(a) The Table 1 baselines with the bound removed (new). We re-ran all four baselines under the following setting:

  • Bound. $\epsilon = 255/255$, so the $\ell_\infty$ bound no longer excludes any image.
  • Budget and data. 100 queries, on the same images and 10 victims as Table 1.
  • Success criterion. A flip counts as a success only if ASA's Gemma-4-E4B verifier, with the same prompt, accepts the image as a natural result that preserves the labelled object. After a rejection, the attack continues with its remaining queries, and, as for ASA, verifier calls are not charged to the query budget.
  • DifAttack. $\sigma = 0.5$, its noise scale for the unbounded setting.
AttackRN-50ConvNeXtSwinDeiTViMMVAdv-ResAdv-IncCLIPSigLIP2
AdvFlow4.293.583.553.814.183.323.583.645.273.14
DifAttack30.4724.6425.9120.5520.2419.2923.9628.1225.3224.91
CGAttack2.361.942.402.652.462.282.681.942.442.44
MCGAttack1.291.021.041.171.181.351.231.091.161.28
ASA94.1087.2289.3488.0390.9077.9095.8797.6984.1967.17

Every baseline flips the label on 83.9–100% of the images, but the verifier rejects more than 99% of the candidates it checks. Without the bound, pixel-space attacks can therefore change the label, but not with images that could occur in the real world: enlarging $\epsilon$ does not reach this axis.

(b) An unrestricted attack built from ASA's components. ASA w/o S,G (Sec. 4.8, Fig. 4; see W1) differs from ASA only in lacking scenario competition (S) and greedy composition (G), so it is unrestricted in the same way. It falls 37.7–53.6 pp short of ASA across the four victims of Fig. 4.

We have added (a) to Tab. 1, Fig. 3(b), and Sec. 4.2.1, with full results in App. L.


W2. Advantage of the greedy strategy over other search methods under the same query budget.

Under the same 100-query budget, greedy is the most query-efficient of the tested rules: no alternative improves on greedy at 20 queries, and it is the best rule in three of the four settings.

Setup. All rows use ASA's MLLM, editor, strategies, and feedback, and every rendered image, including each composition trial, costs one query.

  • MLLM direct generation (ASA w/o G, Fig. 4). No combination; all queries go to scenarios the MLLM writes from the feedback.
  • Random combination. Eqs. (1)–(3), but candidates are visited in random order instead of descending attack score.
  • Beam search, $B = 2$. The two best combinations at each level are kept, and each is extended with every unused candidate.
ASR (%)RN-50, 20 queriesRN-50, 100 queriesSwin, 20 queriesSwin, 100 queries
MLLM direct generation73.8286.3754.3472.73
Random combination87.5593.8872.0085.27
Beam ($B = 2$)85.0995.9268.3487.25
Greedy (ASA)88.3094.1074.1989.34

Clean-correct NIPS 2017 images (932 for RN-50, 957 for Swin). The 20-query columns count attacks of the same runs that succeeded within 20 queries.

Analysis.

  • Composition is the main source of the gain. Every combination rule, not only greedy, is 11.3–19.9 pp above direct generation at 20 queries and 7.5–16.6 pp above it at 100 queries.
  • Ordering candidates by attack score adds to it. Random combination is below greedy in all four columns, by 2.2–4.1 pp on Swin.
  • A wider search pays off only with enough queries. Beam evaluates more combinations per iteration, which costs 3.2–5.9 pp at 20 queries; with all 100 queries it overtakes greedy on RN-50 (+1.8 pp) but not on Swin (−2.1 pp).

We have added this table and analysis to the revision (App. P, Table 18) and refer to it in the ablation study (Sec. 4.8).

Reviewer_hrB6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper present Adversarial Scenario Attack (ASA) which is query-based black-box attack on image classification models. Unlike prior work that performs optimize and search for the adversarial perturbation noise in the image or noise space, ASA performs this search in a text scenario space and uses an ImageEdit generative model to edit the source image with the edit operations that ASA discovers to produce an "edited" adversarial image. The framework supports three groups of editing scenarios (Background shift, Weather & Atmosphere, Material & Color Reskining). The output from image editing is verified for naturalness by an multi-modal LLM which compares the generated image to the reference input image. ASA uses a greedy explorer , and utilize the history of previous attempted edits to make the attack more effective and query efficient.

Critical strengths

Authors have used a large variety of vision models (ranging from CNNs to the more modern SSMs and VLMs) to demonstrate that their technique is not strictly tied to a certain model architecture.

Critical weaknesses

In section 2, Page 3 ... "Because these two methods generate a new image of the target concept they do not edit a specific source image". while this applies to AdvDiff and NatADiff , it is not applicable to ACA which similar to the authors work edits a source image.

  • The figure qualitative examples, shows that the ASA images are way more corrupted and have visible noise even when compared to ACA which is also unrestricted in the pixel space .

  • Figure 3: Results show significant query efficiency and attack success rate compared to other blackbox methods, but that result is miss-leading because the other methods included in the same figure are all "restricted L_P" bound methods while ASA is unrestricted attack. Would the authors be able to compare query efficiency against other black-box unrestricted attacks ? (or relax the L_P bound of these restricted methods to match the L_p bound observed from the ASA results in order to ensure we have a fair comparison ?!

  • Section 3 (Method) is hard to follow, and sometimes abbreviations and methods are used without being defined first. e.g.what is COMP (section 3.3) ?

  • In section 4.3, the control prompt is chosen as "The most frequent prompt that never succeeded during the search". This means that the control prompt is a very weak baseline.

  • The human evaluation in section 4.6, was carried out for ASA but not for the other unrestricted attack methods (would suggest doing it at least for ACA) which is the most similar approach.

Questions for authors
  • Figure 3: Results show significant query efficiency and attack success rate compared to other blackbox methods, but that result is miss-leading because the other methods included in the same figure are all "restricted L_P" bound methods while ASA is unrestricted attack. Would the authors be able to compare query efficiency against other black-box unrestricted attacks ? (or relax the L_P bound of these restricted methods to match the L_p bound observed from the ASA results in order to ensure we have a fair comparison ?!

  • In section 4.2.1: "To Avoid ASR inflation from prompt leakage" .. what is the definition of prompt leakage here?

  • For the prompt transferability attack, the control prompt is very weak baseline (it is chosen as the most frequent prompt that NEVER succeeded during the search). what can be a better choice for a control prompt to verify your hypothesis ?

  • The human evaluation in section 4.6, was carried out for ASA but not for the other unrestricted attack methods (would suggest doing it at least for ACA) which is the most similar approach.

3 comments

We thank the reviewer for the constructive review. We address each point below. All table, figure, section, and appendix numbers refer to the revised version.


W1: ASA images are way more corrupted and have visible noise

ASA adds no noise to the image. It never optimizes or perturbs pixels or latents: each ASA image is the unmodified output of the image-editing model for a short natural-language scenario. What looks like noise in Fig. 1 is the texture these scenarios ask for:

Fig. 1 exampleScenario found by ASA
Castle (first row)"place in arid plains and introduce light drizzle"
Castle (third row)"covered in weathered stone"

Drizzle and weathered stone are fine-grained textures over the scene or the object, as they would appear in a photograph taken under that condition. ASA attacks through changes of this kind: background, weather, and surface material. In our human study, a majority of seven raters judged 97.3% of the 877 ASA successes on ResNet-50 to look like a photograph of a scene that could occur in the real world, the first condition of our definition of natural (Sec. 3.1, Sec. 4.6).

We have added the scenario of each ASA example to the caption of Fig. 1.


W2 & Q1. A comparison with query-based unrestricted attacks.

ASA is not strong merely because it is unbounded: it attacks along a different axis, images of scenes that could occur in the real world (our definition of natural, Sec. 3). Two comparisons in this unrestricted setting support this. Since, to our knowledge, no score-based query attack is unrestricted, we removed the bound from the Table 1 baselines, and ASA leads them by at least 42 pp (a). ASA's own ablation is a second unrestricted attack, and it falls at least 37 pp short of ASA (b).

(a) The Table 1 baselines with the bound removed (new). We re-ran all four baselines under the following setting:

  • Bound. $\epsilon = 255/255$, so the $\ell_\infty$ bound no longer excludes any image.
  • Budget and data. 100 queries, on the same images and 10 victims as Table 1.
  • Success criterion. A flip counts as a success only if ASA's Gemma-4-E4B verifier, with the same prompt, accepts the image as a natural result that preserves the labelled object. After a rejection, the attack continues with its remaining queries, and, as for ASA, verifier calls are not charged to the query budget.
  • DifAttack. $\sigma = 0.5$, its noise scale for the unbounded setting.
AttackRN-50ConvNeXtSwinDeiTViMMVAdv-ResAdv-IncCLIPSigLIP2
AdvFlow4.293.583.553.814.183.323.583.645.273.14
DifAttack30.4724.6425.9120.5520.2419.2923.9628.1225.3224.91
CGAttack2.361.942.402.652.462.282.681.942.442.44
MCGAttack1.291.021.041.171.181.351.231.091.161.28
ASA94.1087.2289.3488.0390.9077.9095.8797.6984.1967.17

Every baseline flips the label on 83.9–100% of the images, but the verifier rejects more than 99% of the candidates it checks. Without the bound, pixel-space attacks can therefore change the label, but not with images that could occur in the real world: enlarging $\epsilon$ does not reach this axis.

(b) An unrestricted attack built from ASA's components. ASA w/o S,G (Sec. 4.8, Fig. 4; see W1) differs from ASA only in lacking scenario competition (S) and greedy composition (G), so it is unrestricted in the same way. It falls 37.7–53.6 pp short of ASA across the four victims of Fig. 4.

We have added (a) to Tab. 1, Fig. 3(b), and Sec. 4.2.1, with full results in App. L.

W4 & Q3. The control prompt.

The scenarios ASA discovers expose weaknesses that models share, and this holds against two new comparison prompts chosen without reference to any attack outcome. Found against ResNet-50 alone, the scenarios are more harmful on 13 models than a null edit, a benchmark edit drawn at random from MagicBrush [R1], and a scenario that ASA rejected: significantly so in 113 of the 117 model-level tests, and in 466 of 468 tests under Gaussian noise.

What each comparison shows.

  • Null edit (2.45%). Passing an image through the editor without asking for any change rarely flips a prediction, so the ASR comes from the content of the scenarios, not from the small changes the editor introduces by itself.
  • Benchmark edit (15.50%). Models are sensitive to ordinary scene edits, yet the discovered scenarios exceed this by 5.25–19.12 pp on average, so their effect is not that of any natural edit.
  • Rejected scenario (10.44%). A scenario that ASA proposed but never succeeded with stays below even the benchmark edit, so the search feedback separates harmful scenarios from harmless ones.
  • Gaussian noise. Even where an edit does not flip the prediction, the discovered scenarios leave images more fragile than all three comparison prompts do.

Setup.

  • Null edit. "keep the image unchanged".
  • Benchmark edit. "make it sunny", drawn from the MagicBrush test split: of its 1,053 instructions, eight pass fixed keyword rules for global scene edits (time of day, weather, season, lighting, sky, or background; App. N), and one of them is drawn uniformly at random. The rules and the seed were fixed before rendering.
  • Common settings. Both use the same editor, settings, 10,000 images, and 13 models as Table 3.

Transfer (Table 3)

PromptRoleMean ASR (%)Range across 13 models (%)Significantly higher than null / rejected / benchmark (of 13 models)
"keep the image unchanged"Null edit (new)2.451.35–4.64—
"coated in fine grit"Rejected by ASA10.446.97–15.87—
"make it sunny"Benchmark edit (new)15.509.42–27.20—
"change lighting to dusk"Discovered by ASA20.759.59–51.1613 / 13 / 9
"coated in fine dust"Discovered by ASA27.1418.31–36.2713 / 13 / 13
"introduce light mist"Discovered by ASA34.6211.78–68.2913 / 13 / 13

Paired two-sided exact McNemar, Holm–Bonferroni over the 117 scenario $\times$ comparison $\times$ model tests.

The discovered scenarios are significantly lower in none of the 117 tests; the four exceptions are Dusk versus the benchmark edit, where the difference is not significant. Light mist and Fine dust are significantly higher than all three comparison prompts on every one of the 13 models, which span CNNs, transformers, Mamba models, adversarially trained models, and vision-language models.

Gaussian noise (Table 4)

PromptRole$\sigma=0.005$$\sigma=0.010$$\sigma=0.015$$\sigma=0.020$Significantly higher than null / rejected / benchmark (of 52 tests)
"keep the image unchanged"Null edit (new)−0.06−0.10−0.11−0.17—
"coated in fine grit"Rejected by ASA+0.22+0.60+0.78+1.02—
"make it sunny"Benchmark edit (new)+0.44+0.84+1.24+1.54—
"change lighting to dusk"Discovered by ASA+2.23+4.24+6.08+7.8652 / 52 / 50
"coated in fine dust"Discovered by ASA+1.88+3.94+5.67+6.9152 / 52 / 52
"introduce light mist"Discovered by ASA+7.11+13.02+18.23+22.9052 / 52 / 52

Entries are the errors added by the edit under the same noise (pp, mean over 13 models as in Table 4). The last column counts the tests (13 models $\times$ 4 noise levels) in which the scenario fails significantly more often than each comparison prompt under identical noise, on images that both edits leave correctly classified before noise (same test, Holm–Bonferroni over the 468 tests).

The discovered scenarios are significantly higher in 466 of the 468 tests and significantly lower in none; the two exceptions are Dusk against the benchmark edit at the smallest noise level.

We have added these results to the main text: the two new prompts as rows of Table 3 and Table 4, and the comparisons against the three prompts in Sec. 4.3–4.4.

[R1] MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. NeurIPS 2023.

W3. Clarity of Sec. 3 and the definition of Comp.

Definition. $\mathrm{Comp}(p)$ returns the strategy-keyed components of a scenario $p$: the single-strategy instructions it is made of, each paired with its strategy. For example,

$$\mathrm{Comp}(\text{“introduce light mist and coat the surface in fine dust”}) = \lbrace (\mathrm{WA}, \text{“introduce light mist”}),\ (\mathrm{MCR}, \text{“coat the surface in fine dust”}) \rbrace,$$

where WA and MCR are the Weather & Atmosphere and Material & Color Reskinning strategies.

Use in Sec. 3.3. When the previous greedy scenario is among the remaining candidates, i.e., it was not selected as the initial greedy state, the Greedy Explorer replaces it in the visiting list with these components rather than attaching it as a whole. Keying by strategy lets a new component replace the existing one of the same strategy (Eq. (2)).

Revision. The following are now defined at first use:

  • $\mathrm{Comp}(\cdot)$ and strategy-keyed components: right after Eq. (1) (Sec. 3.3).
  • The verifier's base description: the template “a photo of ⟨class name of $y$⟩” (Sec. 3.2).
  • $M$: the number of elements visited in iteration $t$ (Sec. 3.3).

Q2. Definition of prompt leakage.

Prompt leakage occurs when a scenario explicitly names an ImageNet category; e.g., "coated in soft velvet" names the class velvet. Since the editor may then add that content to the image, a prediction of that category reflects the edit rather than an error of the classifier. We therefore count predictions of the leaked category as failures. Leakage occurs in only 0.9–7.2% of ASA-generated scenarios (App. D). We have added this definition to Sec. 4.2.1.


Q4. Human evaluation.

The human study tests one thing: whether ASA's successes are natural as we define the term (Sec. 3.1). ASA counts an attack as successful only if its verifier accepts the image as natural, so the study checks ASA's own success criterion with humans. It is not a comparison between attacks, and it is therefore run on ASA's examples: 740 of 877 (84.4%) meet the definition.

Criteria. Both conditions of the definition are absolute: each image meets them or not on its own, which is also how prior work on unrestricted adversarial examples evaluates them [R2].

Protocol. Seven non-author participants each rated all 877 ASA successes on ResNet-50 (NIPS 2017), one at a time, without seeing the originals or being told that the images were edited. For each image, they judged whether it looks like a natural photograph of a real scene, and chose the main object's class among the original class, the victim's post-attack prediction, and "unsure".

Result. By per-image majority (at least four of seven), 97.3% met the first condition and 86.9% the second; 740 met both.

Comparison with other attacks. Our comparisons with other attacks do not depend on the study:

  • Query-based baselines (Tab. 1). With ASA's ASR multiplied by this rate and every baseline left uncorrected, ASA stays above the best baseline without the bound ($\epsilon = 255/255$) on all ten victims, by 31.77–56.93 pp (App. M).
  • Generative unrestricted attacks (Tab. 5). The comparison with ACA and DiffAttack is made with paired metrics against the source image, and ASA is best on all of them: DINOv2 similarity 0.787 against 0.520 (ACA) and 0.713 (DiffAttack), and ID-Sim distance 0.313 against 0.584 and 0.415.

We have revised Sec. 4.6 to state this purpose and added the comparison to App. M.

[R2] Constructing Unrestricted Adversarial Examples with Generative Models. NeurIPS 2018.

Comments on the paper

1 comment

Summary of the revision

We thank all reviewers for their time and constructive reviews, which have strengthened the paper. ASA is a query-based black-box attack that finds a victim's vulnerabilities to natural scene changes, in a form that can be read, composed, and reused: a natural-language editing scenario. This comment summarizes the evidence for this contribution and what the revision adds on the points shared across reviews. All numbers refer to the revised version.


1. Contribution (XnKW, WA3c)

  • Scenario competition (S) and greedy composition (G) drive the gain. ASA w/o S,G keeps the same MLLM, victim-score feedback, image editor, and budget, and falls 37.7–53.6 pp below ASA (Fig. 4). S adds 30.0–40.1 pp and G a further 7.7–16.6 pp.
  • The discovered scenarios are reusable. The three scenarios that recur most often in the search against ResNet-50 alone, applied unchanged to images of all 1,000 classes, reach 20.75–34.62% average ASR over 13 models, against 2.45% for a null edit, 15.50% for a benchmark edit drawn at random from MagicBrush [R1] (both new, chosen without reference to attack outcomes), and 10.44% for a scenario the search rejected (Tab. 3). The search returns a weakness shared across images and models.
  • No gradients, surrogate, or attack-specific training. DiffAttack, ACA, and NatADiff optimize diffusion latents or sampling trajectories with surrogate or classifier gradients; ASA uses only the victim's confidence scores.
  • Highest ASR under the same budget, and query-efficient. With 100 queries, ASA is above every Table 1 baseline on all ten victims, including the baselines without the ℓ∞\ell_\infty bound. Its mean ASR within 20 queries (75.79%) exceeds what every baseline reaches at 100 in either setting (at most 43.36%; Fig. 3).
  • It applies to MLLM victims, including a closed-source one. Using only the victim's text answer, ASA reaches 76.0–83.0% ASR on three MLLMs.

Revision: Sec. 1, Sec. 2.1, Sec. 3.3, Sec. 4.3.

[R1] MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. NeurIPS 2023.


2. Comparison in the unrestricted setting (hrB6, 42vV, WA3c)

ASA's lead does not come from the absence of an ℓ∞\ell_\infty bound. We re-ran all four Table 1 baselines with ϵ=255/255\epsilon = 255/255 on the same images, ten victims, and 100-query budget, counting a flip as a success only if the naturalness verifier accepts the image, as for ASA.

  • The best baseline without the bound reaches 19.29–30.47%, against 67.17–97.69% for ASA: a gap of at least 42 pp on every victim.
  • The baselines flip the label on 83.9–100% of the images, but the verifier rejects more than 99% of the candidates it checks: removing the bound yields label flips, but rarely images accepted as natural and label-preserving.

Among the query-based attacks compared, ASA achieves substantially higher ASR under the same verification criteria in the unrestricted setting.

Revision: Tab. 1, Fig. 3(b), Sec. 4.2.1; full results in App. L.


3. Naturalness of ASA's successes (XnKW, WA3c, hrB6)

We ran a new human study with the majority-vote rule of [R2].

  • Result. A majority of seven non-author raters judged 84.4% of all ASA successes on ResNet-50 (740/877) to look like a photograph of a real scene and to show the original class. On a 100-image Swin-B sample, the rate is 88%, against 81% for ResNet-50 on the same originals.
  • What it means. The large majority of ASA's successes are natural as defined in Sec. 3.1, and this is not specific to ResNet-50.
  • Sensitivity of Table 1. With ASA's ASR on every victim multiplied by the ResNet-50 rate and every baseline left uncorrected, ASA stays above the best baseline without the bound (ϵ=255/255\epsilon = 255/255) on all ten victims, by 49 pp on average and at least 31 pp on every victim. It also stays above the best ϵ=12/255\epsilon = 12/255 baseline on nine of ten.

Revision: Sec. 3.1, Sec. 4.6, App. M, App. O.

[R2] Constructing Unrestricted Adversarial Examples with Generative Models. NeurIPS 2018.


4. Practical application to MLLM victims (XnKW, WA3c)

To examine how practical ASA is when a victim exposes no internal scores, we applied it to MLLM victims through their text interface, including the closed-source gpt-6-luna (new).

  • Result. ASA reaches 76.0–83.0% ASR on three MLLMs. On gpt-6-luna, it reaches 78.0% at 12.53 queries per success (100 clean-correct NIPS 2017 images, 100 queries).
  • What it means. The only victim output used is its text answer: the top-two classes with self-reported confidences, which are neither logits nor calibrated probabilities. This suggests that ASA remains practical for deployed models that return only text.

Revision: Sec. 4.7, App. J.


Therefore, ASA finds a victim's vulnerabilities in a form that can be read, composed, and reused. We hope our responses address the reviewers' concerns, and we thank them again for the constructive reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.