New RACER framework repairs backdoors in Multimodal LLMs

Researchers have developed RACER, a novel framework designed to repair backdoors in Multimodal Large Language Models (MLLMs). Unlike previous methods that focus on inference-time filtering, RACER operates at the model level to eliminate latent backdoors. It identifies and addresses modality-dependent anomalies in the model's internal representations, specifically targeting regions that encode trigger features. Through adversarial fine-tuning, RACER effectively suppresses backdoor behaviors while preserving the model's utility on clean tasks. AI

IMPACT Introduces a new method for enhancing the security and reliability of deployed MLLMs by removing latent backdoor vulnerabilities.

RANK_REASON Academic paper detailing a new method for repairing backdoors in MLLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• MLLMs

• paper
• safety

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