acceptodds
Withdrawn from ICLR 2027

SCAMP: Black-Box Knowledge Poisoning Attack on Multimodal Retrieval-Augmented Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large vision–language models benefit from multimodal retrieval-augmented generation, which grounds responses in external image–text evidence to mitigate hallucination. However, reliance on external knowledge bases introduces a critical vulnerability: knowledge poisoning. Existing poisoning methods are often impractical, as they rely on white-box access, target limited sample-wise scenarios, and neglect reranking components. Additionally, they underutilize the visual modality's potential and suffer from cross-modal inconsistencies. We propose SCAMP, a black-box framework that iteratively crafts aligned poisoned pairs through a self-competitive agent loop. SCAMP integrates a Visual Attacker and a Text Attacker to jointly optimize, Competing Agents to simulate benign samples and improve attack generalization, and three Evaluators to guide optimization via retrieval, generation, and consistency feedback. For class-wise attacks, we introduce agent-driven prototype synthesis to construct effective visual anchors without gradients. Extensive experiments show SCAMP consistently outperforms baselines in black-box settings and remains robust against diverse defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Poisoning Attack, Retrieval-Augmented Generation, Multimodal Retrieval-Augmented Generation

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

The paper proposes SCAMP, a black-box knowledge-poisoning attack against multimodal RAG systems. The authors argue that existing MRAG poisoning attacks share five limitations: white-box access, sample-wise-only scope, no reranker awareness, cross-modal inconsistency, and an underused visual channel. SCAMP is a multi-agent self-competitive loop, all driven by GPT-4o, comprising a Visual Attacker, a Text Attacker, a Competing Agent that synthesizes benign documents to limit overfitting, and three Evaluators that score each candidate and feed back suggestions. For class-wise attacks without the target image, an agent-driven prototype is synthesized via a T2I model under leave-one-out retrieval validation. On an OVEN-Wiki 2M corpus with a SigLIP retriever and LLaVA-1.6 as reranker/generator, injecting a single poisoned pair per target, SCAMP-Typo reports the highest reranked and end-to-end ASR in the sample-wise setting and the highest ASR, though in single digits, in the class-wise setting. Appendices add transfer across retrievers and VLM rerankers/generators, perceptibility metrics, defense evaluations, and the full agent prompts.

Critical strengths

S1. A practically motivated threat model and a broad defense sweep. The black-box, single-injection, class-wise formulation is more realistic than the white-box, sample-wise setting assumed by most prior work, and the explicit treatment of a reranking stage is appropriate for real pipelines. The OCR-masking result in Table 13 is a genuinely informative negative result that the authors report honestly.

S2. Ablations that localize where the effect comes from. The modality ablation and the component ablation consistently point to the same mechanism: text alone retrieves well but fails reranking, and the visual overlay is what survives the reranker. I recomputed the cross-references and they are internally consistent; for example, the full-pipeline ASR of 80.4 in Table 11 matches the last ablation row in Table 3. This transparency makes the paper’s actual contribution easy to identify.

Critical weaknesses

W1. The success metric is not independent of the attack. SCAMP’s Generator and Retrieval Evaluators are GPT-4o agents prompted to help the attack succeed, so they are an optimization signal, not a neutral judge. The problem peaks in Appendix H: when GPT-4o is itself the victim, SCAMP-Typo’s ASR jumps to 73.0%, far above the 48.4% on the default LLaVA victim. When attacker, optimizer, and victim share a model, “success” is largely self-consistency, not black-box transfer. The paper reports no independent, non-GPT-4o ASR judge, so the headline numbers cannot be trusted.

W2. Claimed superiority is unsupported, and the real mechanism is a fragile, known shortcut. SCAMP only wins in the sample-wise setting. In the realistic class-wise setting its ASR is 9.0% / 7.3%, retrieval is below PoisonedRAG, and prototype synthesis adds under 1 point. The working variant, SCAMP-Typo, just renders the answer as text on the image, and OCR masking alone cuts it from 80.0% to 42.0%. Baselines are also handicapped by a surrogate–victim mismatch, and the consistency defense SCAMP survives optimizes the same SigLIP objective SCAMP trains on.

W3. Numbers and models that do not reconcile. SCAMP uses no surrogate, yet its Table 7 (CLIP-victim) and Table 2 (SigLIP-victim) rows are nearly identical, which is unexplained. Appendix H lists “Gemini-3,” an unfamiliar model needing specification.

W4. Overstated threat model. The method calls GPT-4o repeatedly as attacker and judge, and its best results are where that model is also the victim, not the cross-provider black-box setting promised. The single-injection budget is quietly walked back in Appendix N (“attackers would inject multiple poisons per class”), and the agent-plus-diffusion cost is high versus the text baselines.

W5. Self-serving defense evaluation. SCAMP survives defenses aligned with its own objective but is halved by the one neutral defense (OCR masking), which the paper dismisses as impractical without quantifying any benign-traffic cost. There is no adaptive OCR-plus-consistency defender and no corpus-scale test of how often the poison is even selected among realistic benign neighbors.

Questions for authors

Please refer to the weaknesses.

Reviewer_Kk7COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates knowledge poisoning attacks against multimodal retrieval-augmented generation (RAG) systems, addressing the limitations of existing methods in black-box settings, attack generalization, and cross-modal consistency. The authors propose SCAMP, a self-competitive multi-agent framework that jointly optimizes poisoned image–text pairs through visual and text attackers, competing agents, and multiple evaluators, along with agent-driven prototype synthesis for class-wise attacks. Extensive experiments demonstrate that SCAMP consistently outperforms existing baselines and remains effective against various defense mechanisms.

Critical strengths
  1. The writing of this paper is well-structured and easy to follow.
  2. The identified limitations of existing works are valid and reasonable.
  3. The proposed solution generally makes sense, and experiments are extensive.
Critical weaknesses
  1. The target limitations or problems are general and broad, while the technical challenges or in-depth difficulties are not sufficiently presented or discussed. Most of the limitations are paradigm or formulation-level deficiency, even though every individual limitation does make sense, simply and superficially combining them is not enough.
  2. Given the broad and general limitations, the proposed method is also not technically novel, it reads like using an agentic system to cater to new problem setup.
  3. The overall performance of the proposed method does not show significant improvement over baselines.
Questions for authors

please see the weaknesses above.

Reviewer_ABztOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates knowledge-poisoning attacks against multimodal retrieval-augmented generation systems, focusing on the challenges of black-box access, reranking, and cross-modal consistency. It proposes SCAMP, a black-box framework that iteratively constructs poisoned image-text pairs through a self-competitive multi-agent process. The proposed framework combines visual and textual attackers with competing agents and three evaluators for retrieval, generation, and consistency. Experiments across several benchmarks demonstrate improvements over existing baselines, particularly in sample-wise settings, and assess robustness against various defenses.

Critical strengths
  • The paper is generally well written, with clear motivation and a well-organized presentation.
  • The experimental evaluation is relatively comprehensive.
Critical weaknesses
  • Insufficient validation of the black-box optimization. SCAMP relies on GPT-4o-based evaluators to simulate the target retrieval, reranking, and generation processes. However, it remains unclear how accurately these simulated evaluations reflect the actual MRAG pipeline. I suggest the authors quantitatively assess the agreement between evaluator predictions and actual retrieval, reranking, and generation outcomes.

  • Limited evidence for the necessity of the multi-agent design. Several building blocks, including typography-based visual attacks and multimodal knowledge poisoning, have been explored in prior work [1*, 2*]. Although the paper provides component-wise ablations, it remains unclear whether the improvements arise from the proposed self-competitive mechanism or simply from additional optimization attempts. I suggest comparing SCAMP with simpler alternatives, such as single-agent optimization or best-of-N candidate selection, under comparable computational budgets.

  • Limited effectiveness of class-wise attacks. Despite being an important contribution, SCAMP achieves only 9.0% and 7.3% ASR on ImageNet-1K and Places-365, respectively. Moreover, the proposed prototype synthesis improves ASR by only 0.6 percentage points in Appendix L. These results raise concerns about the practical significance of the class-wise attack. I suggest including comparisons with simpler prototype selection strategies and further analyzing why the improvements in retrieval success do not translate into end-to-end attack success.

  • Insufficient support for the robustness claims. Although the paper evaluates several defenses, some substantially reduce attack effectiveness. For example, SCAMP-Typo's ASR decreases from 80% to 38% under the evaluated RoCLIP-style defense and to 42% under OCR masking. Furthermore, the evaluation lacks a systematic analysis of defense false-positive rates and their impact on benign MRAG performance. I suggest reporting the trade-off between attack success reduction and benign task performance to better support the robustness claims.

References

[1*] Yichen Gong et al. FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts. AAAI, 2025.

[2*] Chenyang Zhang et al. PoisonedEye: Knowledge Poisoning Attack on Retrieval-Augmented Generation based Large Vision-Language Models. ICML, 2025.

Questions for authors
  • How accurately do the GPT-4o-based evaluators approximate the actual retrieval and reranking results? Was their prediction accuracy validated against the target MRAG pipeline?
  • In the class-wise setting, how is the query image used by the Competing Image Module (Eq. 5) obtained? Are the evaluation images strictly excluded from the attack construction process?
  • How were the defense thresholds selected, and what is their impact on benign MRAG performance? This information is important for assessing the practical significance of the reported robustness results.

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

Related papers

Discussion (0)

Sign in to comment.