Reproduce “Benchmarking Poisoning Attacks against Retrieval-Augmented Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Baolei Zhang, Haoran Xin, Jiatong Li, Dongzhe Zhang, Minghong Fang, Zhuqing Liu, Lihai Nie, Zheli Liu
RAG Security Bench evaluates 13 poisoning attacks and 7 defenses across five QA datasets and expanded variants. It measures accuracy, attack success, and poisoned-text retrieval. On NQ, BPI’s ASR falls from 0.94 to 0.16 on NQ-EX-M and 0.08 on NQ-EX-L, while CRAG-AK remains at 0.50 and 0.20. These benchmark results support an evidence-density interpretation, but do not establish a universal defense.
Key insights
- Enriching knowledge bases with abundant, diverse correct-answer texts can serve as a passive defense, significantly reducing poisoning attack effectiveness.
- State-of-the-art LLMs, despite alignment training, remain highly vulnerable to poisoned *contextual inputs*, indicating a gap in current LLM security against embedded harms.
- Adaptive retrieval strategies in RAG architectures (e.g., FLARE) show promise for defense by selectively avoiding unnecessary retrieval, thereby reducing exposure to poisoned content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.