Less Is More -- Until It Breaks: Security Pitfalls of Vision Token Compression in Large Vision-Language Models

TL;DR

This paper uncovers that visual token compression in large vision-language models, while improving efficiency, significantly reduces robustness and introduces security vulnerabilities, especially under small perturbations, revealing a critical efficiency-security trade-off.

Contribution

The study reveals the security risks of visual token compression in LVLMs, identifies instability in token importance ranking as the cause, and proposes a targeted attack method to exploit this vulnerability.

Findings

Token compression degrades model robustness.

Small perturbations can cause significant token ranking changes.

Compression introduces hidden security vulnerabilities.

Abstract

Visual token compression is widely adopted to improve the inference efficiency of Large Vision-Language Models (LVLMs), enabling their deployment in latency-sensitive and resource-constrained scenarios. However, existing work has mainly focused on efficiency and performance, while the security implications of visual token compression remain largely unexplored. In this work, we first reveal that visual token compression substantially degrades the robustness of LVLMs: models that are robust under uncompressed inference become highly vulnerable once compression is enabled. These vulnerabilities are state-specific; failure modes emerge only in the compressed setting and completely disappear when compression is disabled, making them particularly hidden and difficult to diagnose. By analyzing the key stages of the compression process, we identify instability in token importance ranking as the primary cause of this robustness degradation. Small and imperceptible perturbations can significantly alter token rankings, leading the compression mechanism to mistakenly discard task-critical information and ultimately causing model failure. Motivated by this observation, we propose a Compression-Aware Attack to systematically study and exploit this vulnerability. CAA directly targets the token selection mechanism and induces failures exclusively under compressed inference. We further extend this approach to more realistic black-box settings and introduce Transfer CAA, where neither the target model nor the compression configuration is accessible. We further evaluate potential defenses and find that they provide only limited protection. Extensive experiments across models, datasets, and compression methods show that visual token compression significantly undermines robustness, revealing a previously overlooked efficiency-security trade-off.