acceptodds
Withdrawn from ICLR 2027

Feature-Aware Token Attack for Compression-Triggered Stealthy Failures in Large Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual-token compression improves the efficiency of large vision-language models, but can expose failures that full-token evaluation misses. We study adversarial images that preserve full-token correctness yet induce errors after compression, even when both inference paths succeed on the clean image. Creating such failures is challenging because perturbing token importance can also damage the visual content needed for full-token inference. We propose Feature-Aware Token Attack (FATA), which couples attention suppression with cosine-based feature preservation on a fixed set of salient clean-image tokens. In the primary LLaVA-1.5-7B setting, FATA uses only vision-encoder gradients, without access to the deployed compressor, token budget, or downstream task. Across four visually dependent task subsets and four compressors under a controlled reconstruction protocol, FATA achieves SR = 96.3% full-token accuracy retention and CBR = 22.1% conditional blinding, compared with 89.8% and 15.7% for CAA. Ablations support the role of both objectives in balancing compressed-path failure against full-token preservation. FATA also has the lowest measured detection rate among four attacks across three evaluated detectors at a 5% false-positive rate. These findings motivate assessing adversarial robustness jointly across full-token and compressed inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large vision-language models, visual token compression, adversarial robustness, stealthy adversarial attacks, compression-triggered failures

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

This paper proposes a PGD-like method to modify input images, which makes the victim model produce wrong results when visual token compression is applied.

Critical strengths
  1. This paper has a comprehensive evaluation under different conditions, and although the method (PGD-like) is naive, it seems to work.
  2. The loss design is well validated.
Critical weaknesses
  1. The core contribution seems not enough. This paper seems to be an incremental work compared with the existing work CAA, in that it only adds an explicit semantic preservation loss.
  2. The background of the problem is unclear; adding a problem formulation section would be better for readers.
  3. The fragility of this attack method is unclear (no defense methods are used in this paper), such as (i) input purification (Gaussian denoising / median filter / bit-depth reduction), and (ii) a token-level mitigation such as randomized selection or ensembling over compression budgets.
Questions for authors
  1. Could you please clarify the runtimes of this method and the baselines?
  2. Could you please provide the performance under different perturbation budgets?
Reviewer_ybHdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adversarial images that a vision-language model can answer correctly with all tokens but gets wrong after compression. FATA reduces attention to selected clean-image tokens while keeping their feature directions close to those of the clean image, using only vision-encoder gradients in the main LLaVA setting. The paper reports 96.3% full-token accuracy retention and 22.1% conditional blinding across four task subsets and four compression paths, along with model-specific extensions and detection experiments.

Critical strengths
  1. The two-term FATA objective is simple and only needs vision-encoder gradients, making it useful when the downstream model is unavailable.

  2. Choosing budgets from clean results and using the paired metric in Eq. 8 helps separate attack-induced failures from ordinary compression errors.

Critical weaknesses
  1. The CBR results need to be checked. Table 2 reports 22.1% for FATA and 15.7% for CAA, but the supplementary results/paper/CBR_RECOMPUTATION.md marks these values as pending recomputation and leaves the corresponding counts unavailable. This also affects the ablation and cross-model results in Tables 3, 5, and 7. Average accuracies cannot recover the paired counts required by Eq. 8, so I cannot tell whether the reported advantage over CAA holds under this definition.

  2. It is unclear whether FATA works equally well with ordinary token pruning. Section 3.1 and Fig. 2D first select K representatives, then reconstruct all 576 slots before passing them to LLaVA; the supplied model extensions also reconstruct full-length sequences. This changes how often tokens are repeated and where they appear, and the reconstruction itself could contribute to the failures. The setup is explained in the paper, but a comparison that actually passes only K tokens downstream is still needed. The Qwen and InternVL extensions also use extra task or compressor information, so they do not show that the original encoder-only attack transfers unchanged.

Questions for authors
  1. Why does the supplement mark CBR as pending when Tables 2, 3, 5, and 7 give final values? Please provide the paired correctness records or verified numerator/denominator counts for each setting so that the CAA comparison and loss ablations can be checked against Eq. 8.

  2. Using the same LLaVA adversarial images and practical budgets, does FATA still improve on CAA when only K tokens are passed downstream, without the reconstruction in Fig. 2D? This would help show whether the result comes from pruning or from the added reconstruction step.

  3. The supplied results give a visual-stage TPR@5FPR of 97.4% for FATA and 93.03% for CAA, while averaging FATA's three stages gives 49.72%. Given this strong visual-stage detector, how much support is there for the low-detectability claim? Please show the stage-level results and whether the same pattern holds for attacks that actually cause a compression-triggered failure.

Reviewer_QK3tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose a novel method for attacking vision-language models that utilize attention-based compression approaches, dubbed Feature-Aware Token Attack (FATA).

In a nutshell, the proposed method attempts to simultaneously optimize two losses in order to keep the original (full-token “resolution”) inference intact, while also achieving a successful adversarial attack on the compressed inputs. Namely, they pick the top-N high-attention tokens in the visual encoder and minimize their average attention score so they are less likely to be selected during the compression stage of inference. On the other hand, for these top-N tokens, they encourage the representations generated after the attack to be similar to those of the original image, what helps limiting performance degradation during full-token inference.

The proposed method showed promising results on a number of benchmarks, outperforming other approaches in terms of keeping the model’s original performance in the full-token inference regime, while also showing successful attacks in the compressed case.

Critical strengths
  • The paper is clearly written, the method is logical, and the authors properly address the motivation behind introducing a new type of attack for compression-based VLMs.
  • The empirical validation features proper evaluation, demonstrating an improvement over existing approaches
  • the paper tackles an important and timely problem of VLM adversarial attacks, which has critical use cases in real-world applications.
Critical weaknesses
  • The method requires access to the visual encoder, which is a less strict requirement compared to full model access, though it still might not be realistic in many practical scenarios.
  • Following the previous point, as the authors also acknowledged, there is no clear evidence that attack success transfers to closed-source models. This means it is not clear if the proposed attack can indeed be used in widespread real-world settings. Adding this kind of evaluation might significantly increase the impact of the paper.
  • In terms of benchmarking, the authors use only moderately small models (<10B), which limits generalization of the conclusions to larger models (especially MoE-based models, for which attention suppression and feature-preserving methods might not work in a straightforward manner).
  • Conceptually, the proposed method could similarly work for video / multi-frame inputs, though its effectiveness might be limited. Additional evaluation could be very valuable.
Questions for authors
  • Would it be correct to say that you select the important tokens by averaging the last-layer CLS-to-patch attention scores across all heads? Why this strategy? Did you test alternative aggregation strategies, such as maximum attention across heads, weighted averaging, etc? The motivation for choosing simple averaging is not sufficiently clear.
  • Do adversarial images generated against one model transfer to another without reoptimization? Also, what happens when models use different encoders?
  • Could you evaluate transfer to a closed-source model?

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

Related papers

Discussion (0)

Sign in to comment.