Be Kind, Rewrite: Benign Projections via Rewriting Defend Against LLM Data Poisoning Attacks
Researchers have developed a new defense strategy called Open-Book Benign Rewriting (OBBR) to protect Large Language Models (LLMs) from data poisoning attacks. This method involves rewriting training data to align with benign prompts, effectively neutralizing harmful content. OBBR has demonstrated significant improvements in safety performance, outperforming existing defenses by an average of 51% across various LLMs and known attack patterns. AI
IMPACT Introduces a novel defense mechanism that significantly enhances LLM security against data poisoning, potentially improving trust and safety in LLM deployments.