New 'Eyes-on-Me' method enables scalable RAG system poisoning

Researchers have developed a new method called "Eyes-on-Me" to more effectively poison retrieval-augmented generation (RAG) systems. This technique decomposes adversarial documents into reusable "Attention Attractors" and "Focus Regions," allowing attackers to adapt to new targets with minimal cost. The method significantly increases attack success rates across various RAG configurations and demonstrates the potential for scalable RAG data poisoning. AI

IMPACT Introduces a scalable threat to RAG systems, highlighting vulnerabilities in attention mechanisms and the need for improved defenses.

RANK_REASON Academic paper detailing a new method for RAG system poisoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Attention Attractors
• CORE Recommender
• DagsHub
• Eyes-on-Me
• Focus Regions
• Hugging Face
• IArxiv Recommender
• retrieval-augmented generation
• Yen-Shan Chen

• paper
• safety

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