EntropyScan detects LVLM backdoors using visual attention anomalies

Researchers have developed EntropyScan, a new method for detecting backdoors in Large Vision-Language Models (LVLMs). This approach is model-level and does not require knowledge of the training data or specific attack triggers. EntropyScan identifies backdoors by analyzing anomalies in the visual attention allocation of LVLMs when processing benign samples, indicating a disruption in cross-modal alignment. The method utilizes Tsallis entropy to quantify these distortions, achieving high accuracy in experiments. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel method for detecting security vulnerabilities in vision-language models, crucial for safe deployment.

RANK_REASON Academic paper introducing a new method for backdoor detection in LVLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• EntropyScan
• LVLMs
• Tsallis entropy

• paper
• safety