acceptodds
Withdrawn from ICLR 2027

Profiling Compression, Disrupting Distributions: Structured Adversarial Attacks on Token-Compressed Large Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual token compression offers a promising pathway to improve the efficiency of large vision-language models (LVLMs). However, it creates a fundamental mismatch for adversarial attacks: compression acts on token representations, whereas attack success is evaluated at the sequence level. Existing attacks typically identify tokens likely to survive compression and optimize perturbations over this predicted survivor set. Yet this strategy overlooks two challenges: token utility varies across compression schemes, and independent survival scores cannot capture the joint attack effect of retained tokens. To address these challenges, we propose Compression-Aware Profiling with Coordinated Distributional Disruption (CLOUD), a novel framework for conducting adversarial attacks that remain effective across distinct token-compression schemes. The key idea is to profile compression-dependent attack utility with a surrogate pool and then coordinate support construction with distributional perturbation. Specifically, we first introduce a utility-guided surrogate probing module, which applies Bayesian optimization over the surrogate configuration space to characterize compression-aware attack responses. Subsequently, we construct a dynamic cache to aggregate token-retention patterns and representation shifts across surrogates to calibrate the attack budget. Guided by these profiles, we leverage a determinantal point process (DPP) to construct a diverse support set by balancing compression resilience with complementary evidence coverage. Additionally, we maximize the discrepancy between clean evidence and its post-compression adversarial counterpart from the perspective of optimal transport (OT). This coordinated objective encourages the attack effect to persist after token reduction. Experiments across multiple LVLM backbones and compression schemes show that CLOUD consistently outperforms strong baselines. Code is available at https://anonymous.4open.science/r/CLOUD_For_Token_Compression_Robustness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large language model, token compression, adversarial attack

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_Ms9aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CLOUD, a compression-aware adversarial attack framework for large vision-language models (LVLMs). CLOUD leverages multiple compression surrogates to estimate adaptive attack budgets, followed by DPP-based token selection and optimal transport-based perturbation optimization to disrupt post-compression visual representations. Experiments across multiple LVLMs and token compression methods demonstrate improved attack effectiveness over existing baselines.

Critical strengths
  1. Clear and Well-Organized Presentation. The paper is well-written, with a clear structure and logical flow. The motivations, methodological designs, and technical details are presented clearly, making the proposed framework easy to follow.

  2. Useful Empirical Analyses. The paper provides detailed ablation studies and visualizations that help illustrate the roles of different components and offer insights into adversarial attacks under visual token compression.

Critical weaknesses

1. Limited Practicality of the Threat Model and Incremental Methodological Novelty.

The threat model raises concerns regarding the practical applicability of the proposed attack. Although the target compression strategy is assumed to be unknown, the attack still relies on access to a differentiable vision encoder for perturbation optimization. Such an assumption limits the practical significance of the proposed method, particularly for real-world black-box LVLM services. Furthermore, despite its seemingly sophisticated design, CLOUD appears to be largely an incremental extension of CAGE rather than a fundamentally new attack paradigm. The proposed framework combines several established techniques, including Bayesian optimization, Entmax, DPP, and optimal transport, to improve token selection and perturbation optimization. However, the underlying motivation remains closely aligned with CAGE: identifying and disrupting visual evidence that can influence the model after token compression. The authors should more clearly articulate the fundamental novelty beyond combining existing optimization techniques and justify whether the additional methodological complexity is necessary.

2. Limited Empirical Improvements, Insufficient Baseline Comparisons, and Unclear Generalization.

The empirical improvements over the strongest baseline, CAGE, are relatively modest or inconsistent across several experimental settings, raising questions about whether the additional complexity of CLOUD is sufficiently justified. For example, on LLaVA, the average improvements over CAGE on GQA are only 0.4–1.3 percentage points, while those on ScienceQA are similarly limited. Although larger improvements are observed in some Qwen3-VL settings, the overall advantage is not consistently substantial. Moreover, the adversarial comparisons are insufficiently comprehensive, including only VEAttack and CAGE, with CAGE being the only baseline specifically designed for token-compressed LVLMs. This makes it difficult to establish the broader competitiveness of the proposed approach. Another important concern is generalization to genuinely unseen compression methods. Although the paper evaluates several existing compression schemes, it remains unclear whether CLOUD would remain effective against a fundamentally new compression mechanism with substantially different token-selection or merging behaviors. A held-out compression-method evaluation would provide stronger evidence for the claimed transferability and practical generalization.

Questions for authors
  1. Given the reliance on access to a differentiable vision encoder, can CLOUD generalize to fully black-box LVLM settings where neither the model architecture nor its gradients are accessible?

  2. Given the limited improvements over CAGE in several settings, can the authors justify the additional methodological complexity and provide comparisons with more adversarial attack baselines?

  3. Can the authors evaluate CLOUD against entirely unseen token compression methods to further demonstrate its generalizability?

Reviewer_p4v2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CLOUD, which operates in two stages. The first stage is compression-aware profiling. The second stage is coordinated distributional disruption. DPP(Determinantal Point Process) selects a set of important yet non-redundant tokens as the attack support. The overall objective is to reliably degrade model accuracy under unknown compression schemes and retention budgets. This paper also provides conditional bounds on attack-support size and a greedy approximation guarantee for the log-determinant selection objective, with experiments on four LVLM backbones, five compression methods, and three benchmarks.

Critical strengths
  1. The method has a clear motivation, and its components follow from the observations in Section 3.1. Variation in effective attack budgets motivates sample-specific allocation; the removal or corruption of visual evidence motivates distributional disruption; and redundancy in attention-based Top-Kselection motivates DPP-based support construction. Building on earlier compression-aware attacks such as CAGE, CLOUD offers a concrete way to account for different compression behaviors and feature diversity within the attack support.

  2. The experimental coverage is fairly comprehensive, and the conclusions have some statistical reliability. The evaluation covers LLaVA-v1.5-7B, Qwen3-VL-8B, InternVL3.5, and GLM-4.6V-Flash; five compression methods (VisionZip, VisPruner, DivPrune, FlowCut, and PruMerge); and GQA, TextVQA, and SQA. It also varies token-retention budgets and the perturbation budgets. Some gains are substantial: on Qwen3-VL with GQA and 36 retained tokens, CLOUD reduces mean robust accuracy across compression methods from CAGE’s 57.4% to 48.1%, a decrease of 9.3 percentage points (Table 2).

  3. This paper includes useful analyses beyond aggregate accuracy. Table 8 compares fixed and adaptive support sizes with nearly identical mean token counts, reporting a reduction in mean robust accuracy from 16.20% to 14.63%. The appendix also includes a paired-bootstrap comparison and attack success on initially correct predictions (Figure 8), post-compression OT discrepancy (Figure 9), and runtime and GPU memory measurements (Table 10). These analyses provide evidence about both attack effectiveness and the proposed mechanism.

Critical weaknesses
  1. There remains a gap between the threat model and real-world deployment: The paper assumes the visual encoder is accessible while the target compressor C∗C^* and retention budget KmodelK_{model} are unknown, yet it does not provide a direct mismatch analysis of whether the four surrogates (attention, similarity, activation, merge) can cover the five real compression methods (VisionZip, VisPruner, DivPrune, FlowCut, PruMerge). It is also unclear whether the surrogates remain valid if the target model uses training-based or learnable compression, which directly affects the generalization boundary of the method.

  2. The contributions of adaptive budget and DPP are not fully disentangled. Table 8 compares fixed average and adaptive, but the fixed average only controls the mean KattackK_{attack} without further controlling the spatial distribution or token identity of the support. As a result, it is unclear how much of the improvement from adaptive allocation comes from the budget itself versus DPP selecting more complementary positions. If the fixed budget also used DPP for token selection, the gap might shrink.

  3. Several definitions and experimental details needed for reproduction are missing. The compression-sensitive disruption measure in Equation (3) lacks an explicit, implementable definition. The paper does not specify the source and size of XcalX_{cal}, whether it is disjoint from the first 1,000 evaluation instances of each benchmark, or the BO search space and number of evaluations. Algorithm 1 also leaves TprofileT_{profile} and TattackT_{attack} unspecified, making it difficult to assess whether CLOUD and the baselines receive comparable gradient-computation budgets. Although Table 10 reports 1.14 seconds per sample, the timing scope and treatment of offline BO calibration costs are unclear.

  4. The evidence for some components is inconclusive. In Table 7, Untuned + Median and BO + Median both report 14.6% mean robust accuracy. At KmodelK_{model}=32, BO increases accuracy from 13.8% to 15.3%, weakening the attack. In Table 3, the full surrogate pool improves on the best individual attention surrogate by only 0.3 percentage points: 14.9% for M1 versus 14.6% for M5. These small ablation gains are not accompanied by uncertainty estimates, leaving the benefit of the added complexity unclear.

  5. The relationship between the surrogate objective and the deployed compressor needs stronger support. The four heuristic surrogates are plausible, but the paper provides limited analysis of how well they approximate the retention and aggregation behavior of the evaluated methods, particularly a method such as FlowCut that uses information across layers. Equations (8)–(10) operate on DPP-selected adversarial tokens rather than the representations ultimately retained or merged by the target compressor. The theoretical results do not resolve this gap. Theorem 3.1 provides conditional support-size bounds under the assumption that a majority of surrogate memories approximate an ideal utility, while Theorem 3.2 gives a (1−1/e1-1/e) approximation guarantee for the log-determinant objective. Neither establishes a link to downstream attack success. Figure 9 measures post-compression feature discrepancy, but a sample-level analysis connecting surrogate OT loss, post-compression distortion, and prediction flips is still missing. Applicability to learned token-reduction modules also remains an open question beyond the evaluated pruning and merging methods.

Questions for authors
  1. Please specify the full definition of MsM_s, the calibration split, the BO search configuration, and the values of TprofileT_{profile} and TattackT_{attack} . Do CLOUD’s gains over CAGE and Base persist under matched gradient-evaluation budgets? Does Table 10 include BO calibration, and if not, how is that cost amortized across evaluation samples?

  2. What does BO contribute when the reported mean in Table 7 is unchanged and performance deteriorates at Kmodel=32K_{model}=32? Are the gains from BO and the full surrogate pool stable across datasets and random initializations? In Table 8, does the fixed-budget variant use the same DPP selection, profiling initialization, and perturbation objective, with only KattackK_{attack} fixed to 13? If so, please make this explicit so that the gain from adapting the support size can be assessed directly.

  3. How closely do the surrogate responses track the behavior of the evaluated compressors? Is CLOUD intended to extend to learned token-reduction modules, such as TokenPacker or Q-Former-based resamplers, and would the surrogate design need to change? Within the evaluated setting, how strongly does the surrogate OT loss correlate with post-compression distortion and attack success, especially for merging methods such as PruMerge and across retention budgets?

  4. The surrogate pool contains attention, similarity, activation, and merge surrogates. Can you provide a more direct mapping between each surrogate and the five compression methods evaluated? Would the surrogate pool need to be redesigned for training-based or learnable compression?

Reviewer_yj9COpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on attacking LVLMs under visual token compression. Because perturbations are added before token compression while the model output only depends on the compressed sequence, standard attacks often lose effectiveness. The authors introduce CLOUD to address this gap. Results across multiple LVLMs and compression methods show that CLOUD lowers model accuracy more effectively than prior baselines.

Critical strengths
  1. Strongly motivated perspective and clear intuition. Formulating visual token compression as an ``input-dependent evidence bottleneck'' is intuitive and well-founded. The preliminary empirical study in Section 3.1 effectively illustrates why relying solely on token-wise survival probabilities is suboptimal.

  2. Broad empirical evaluation. The paper evaluates the proposed method across four LVLM architectures, five token compression algorithms, and three VQA benchmarks with distinct visual reasoning characteristics.

Critical weaknesses
  1. Over-engineering with marginal gains over simpler alternatives: Surrogate probing, Bayesian Optimization (BO), alpha-Entmax, DPP, and Partial Optimal Transport (POT) are all established techniques; the main contribution lies in their integration. However, the empirical results do not sufficiently justify why this complex pipeline is necessary compared to simpler diversity regularizers, basic surrogate ensembles, or fixed-budget strategies. For instance, in Appendix Table 7, Untuned + Median (without BO calibration) achieves the exact same average robust accuracy (14.6%) as BO + Median, and even outperforms the BO-tuned version by 1.5% at K_model = 32 (13.8% vs. 15.3%).

Despite introducing computationally heavy modules (such as leave-one-surrogate-out BO calibration, greedy DPP selection, and per-step Sinkhorn POT solving within the PGD loop with O(N^2) complexity), the main text completely omits any runtime or complexity analysis. Furthermore, Appendix Table 10 falls short: it is restricted to a single fixed-resolution setting on LLaVA-v1.5, leaves out the offline BO overhead and exact step counts (T_profile and T_attack), and lacks scalability tests on models with larger token counts like Qwen3-VL.

  1. Lack of evaluation on decoder-layer token pruning: All five baselines evaluated in the paper compress tokens on the vision side before the projector (Appendix A.1). This leaves out another major family of methods discussed in Section 2, such as FastV and SparseVLM, which prune tokens inside the language decoder layers using cross-modal attention. Evaluating against at least one text-guided decoder pruning method would make the results much more convincing.
Questions for authors

How does the total runtime (including calibration) compare to CAGE and VEAttack, and what are the exact step counts for T_profile, T_attack, and the baselines? Please also break down the time spent on profiling, DPP, and Sinkhorn OT, and show the runtime/memory cost on models with longer token sequences (e.g., Qwen3-VL).

Could you give more details on the BO calibration (size of X_cal, number of queries, and sensitivity to domain shift)? Since Untuned + Median matches the full model on average in Table 7, is BO actually needed?

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

Related papers

Discussion (0)

Sign in to comment.