New attack exploits vision-language model failures from compression

Researchers have developed a new adversarial attack method called Feature-Aware Token Attack (FATA) designed to exploit vulnerabilities in large vision-language models (VLMs) that arise from visual-token compression. This attack aims to cause stealthy failures where the model performs correctly with full tokens but errs after compression, even if both inference paths initially succeed on the clean image. FATA achieves this by suppressing attention while preserving cosine-based features of salient tokens, demonstrating high accuracy retention and conditional blinding on the LLaVA-1.5-7B model across various compressors and tasks. AI

IMPACT This research highlights potential security vulnerabilities in compressed VLMs, necessitating further investigation into robust evaluation methods.

RANK_REASON The cluster contains a research paper detailing a novel attack method for large vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Creative Artists Agency
• Feature-Aware Token Attack
• LLaVA-1.5-7B

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →