acceptodds
Withdrawn from ICLR 2027

SGA: Stamp Guided Black-box Adversarial Attack against Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language models (VLMs) have demonstrated remarkable multimodal capabilities, yet remain vulnerable to jailbreak attacks that circumvent safety alignment and elicit harmful answers. Existing visual attacks typically either optimize pixel-level adversarial perturbations with limited explicit semantics and white-box model access, or construct semantically meaningful multimodal inputs through specific content transformation and external model assistance. In this work, we investigate whether a simple visual cue with widely recognized semantics can itself serve as an effective attack medium. To this end, we propose Stamp Guided Attack (SGA), a straightforward black-box adversarial attack framework that overlays an authorization-themed visual stamp onto the original image and augments the textual query with a fixed, scenario-agnostic priming suffix to induce a harmful response. SGA introduces the reusable checkmark as an explicit semantic hint and represents it with a compact, interpretable parameterization over its shape, size, position, color, and opacity. This formulation reduces attack generation to a low-dimensional black-box optimization problem, which is efficiently solved using a mixed-integer Covariance Matrix Adaptation Evolution Strategy (CMA-ES) through only input-output queries to the victim model. Extensive experiments on seven representative VLMs demonstrate the effectiveness and generalizability of SGA, achieving average Attack Success Rates (ASRs) of 78.78% across five open-source models and 54.25% across two closed-source models. The results reveal that VLMs are susceptible to high-level visual semantics introduced beyond the original image content, highlighting the importance of accounting for such cues in multimodal safety evaluation and alignment. Code is available at https://anonymous.4open.science/r/Stamp_Guided_Attack-7EFA/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Jailbreak Attack, Vision-Language Models, Black-box Adversarial Attack, Multimodal Safety, Evolution Strategy

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_o1VBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a straightforward black-box adversarial attack framework that overlays an authorization-themed visual stamp onto the original image and appends a fixed, scenario-agnostic priming suffix to the textual query to induce harmful responses. Experimental results show that their method outperforms current baselines.

Critical strengths
  1. Jailbreak attacks against large models are a very cutting-edge and practically meaningful topic;

  2. The attacked models selected are representative;

  3. The paper provides the code.

Critical weaknesses
  1. The baseline methods do not include jailbreak attacks from 2026;

  2. Figure 1 should tell the authors' work in the most concise and refined way, but the authors fail to do so;

  3. The ablation experiments mainly explore various parameter ablations of the evolutionary algorithm, and the significance of these experiments is limited. Instead, the authors should focus on comparing different optimization algorithms;

  4. The experimental data samples are too few to fully support the experimental conclusions.

Questions for authors

See Weaknesses.

Reviewer_PDCZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SGA, a black-box jailbreak attack against vision-language models (VLMs). The method overlays a checkmark-shaped stamp onto the original image and combines it with a fixed textual suffix to induce unsafe responses from the target model. The authors employ CMA-ES to optimize the stamp's shape, size, position, color, and opacity, achieving average attack success rates of 78.78% and 54.25% across five open-source and two closed-source VLMs, respectively.

Critical strengths
  1. The paper investigates the impact of visual symbols on the safety behavior of VLMs. This is an interesting research direction, particularly regarding whether multimodal models may mistakenly interpret certain visual symbols as trustworthy authorization signals.

  2. The proposed method is relatively simple, involving only a small number of interpretable parameters and requiring no access to the target model's gradients or internal architecture.

  3. The experiments cover multiple open-source and closed-source VLMs, and the reported attack success rates are relatively high, demonstrating that the proposed attack is effective against the evaluated models.

Critical weaknesses
  1. Limited novelty. SGA is closely related to existing adversarial patch attacks, mainly replacing the patch with a checkmark and optimizing its visual attributes using CMA-ES. The technical novelty beyond existing methods is not sufficiently clear.

  2. Insufficient ablation studies. The paper lacks comparisons with other visual symbols, such as circles, crosses, or random patterns. It is therefore unclear whether the improvement comes from authorization semantics or simply from visual perturbations.

  3. Unfair baseline comparisons. SGA allows large and highly visible image modifications, whereas pixel-level attacks typically operate under stricter perturbation constraints. Moreover, VAE and UMK are evaluated through cross-model transfer, while SGA is optimized directly against each target model.

  4. Lack of optimization baselines. The paper does not compare CMA-ES with random search under the same query budget, making it difficult to assess the actual contribution of the optimization strategy.

Questions for authors
  1. What are the main methodological differences between SGA and existing adversarial patch attacks? Beyond restricting the patch to a checkmark shape, what are the substantive technical contributions of this work?

  2. If the checkmark is replaced with simple geometric shapes or symbols without authorization semantics, while using the same optimization procedure, how would the attack success rate change? How do the authors verify that the improvement comes from authorization semantics rather than general visual perturbations?

  3. How much image distortion do the optimized stamps introduce? Does SGA still maintain a clear advantage when different methods are evaluated under the same or comparable visual perturbation constraints?

  4. Why are VAE and UMK evaluated in a cross-model transfer setting, while SGA is directly optimized against each target model? How do the authors justify the fairness of these comparisons?

  5. How much does CMA-ES outperform random search under the same query budget?

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

The paper proposes Stamp Guided Attack (SGA), which tests the safety robustness of vision-language models by overlaying a checkmark-style stamp on the original image and appending a fixed textual suffix. It parameterizes the stamp’s shape, size, position, color, and opacity, and uses CMA-ES to search under black-box query access. Evaluated on 130 samples from MM-SafetyBench across seven models, SGA achieves reported average attack success rates of 78.78% on open-source models and 54.25% on closed-source models.

Critical strengths

The research premise is intuitive. Compared with difficult-to-interpret pixel-level noise, a stamp with recognizable visual meaning provides an understandable and inspectable test object.

The evaluation covers multiple models. Experiments include five open-source and two closed-source models, showing that the observed phenomenon is not restricted to a single architecture.

Basic component ablations are provided. Table 3 separately removes the stamp, optimization, and textual suffix, providing preliminary evidence about visual and textual contributions.

Optimization and final evaluation use different judges. Search uses HarmBench, while final evaluation uses GPT-4o, reducing the risk of directly optimizing and reporting results against the same judge.

Critical weaknesses
  1. The comparison settings are not aligned, and the benefits of optimization are not sufficiently isolated. SGA directly queries the target model, whereas VAE and UMK in Appendix B are optimized on MiniGPT-4v and then transferred to other models. These are different evaluation settings. Moreover, w.o. Optimization uses a randomly configured stamp without a random-search baseline under the same query budget. The results therefore do not isolate the contribution of CMA-ES updates from that of multiple attempts. Budget-matched comparisons should be added, and conclusions about white-box methods should be qualified.

  2. Direct controls supporting the “authorization semantics” explanation are missing. The experiments compare the presence or absence of stamps and different visual attributes, but do not include neutral shapes with matched visual properties. Consequently, explanations involving occlusion or visual salience cannot be ruled out, and the gains cannot be specifically attributed to authorization semantics. Neutral-shape controls matched in size, color, opacity, and position should be added.

  3. The relationship between the two judges and the definition used for query-cost statistics are unclear. Optimization stops early when the HarmBench score exceeds 0.5, whereas final success requires a GPT-4o score of 5. The paper does not report agreement between these judgments, and Table 5 does not explicitly state which judge defines successful cases. Cross-tabulated results and a clear success definition for query-cost reporting should be provided.

Questions for authors

Under the same query budget, how much additional benefit does CMA-ES provide over random search? What proportion of successful cases occurs before the first distribution update?

Does the checkmark stamp retain a significant advantage over neutral shapes with matched visual properties? How can authorization semantics be distinguished from occlusion and visual salience?

How consistent are HarmBench-triggered early stopping and GPT-4o’s final success judgments? Which success definition is used in Table 5, and were the final results validated through fresh response generation or human review?

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

Related papers

Discussion (0)

Sign in to comment.