RAG system architectures show varied robustness to knowledge base poisoning

Researchers have investigated the vulnerability of Retrieval-Augmented Generation (RAG) systems to knowledge base poisoning, finding that system architecture significantly impacts adversarial robustness. Evaluations on the Natural Questions dataset revealed that architectures designed to handle conflicting information, such as Recursive Language Models (RLM), were substantially more resistant to poisoning attacks compared to vanilla RAG systems. The study indicated that adversarial framing, rather than retrieval optimization, was the primary driver of attack success for most architectures, highlighting the content-reasoning stage as a key vulnerability. AI

IMPACT Highlights architectural choices as critical for RAG system security against adversarial attacks, influencing future system design.

RANK_REASON Academic paper detailing a new evaluation of RAG system architectures against knowledge base poisoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CorruptRAG-AK
• MADAM-RAG
• Natural Questions
• Recursive Language Models

• paper
• safety

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