On the Adversarial Robustness of Large Vision-Language Models under Visual Token Compression

In three linesStudy of adversarial robustness in compressed vision-language models. Authors propose CAGE attack that exploits the mismatch between perturbation optimization (full tokens) and inference (via compression). CAGE combines expected feature disruption and rank distortion alignment to expose hidden vulnerabilities in compressed LVLMs.

Summary generated by Claude — human-verified