GuardRAG: a post-retrieval defense against RAG poisoning attacks
Start here
An arXiv paper proposes GuardRAG, a training-free, plug-and-play post-retrieval defense that identifies mutually corroborating suspicious document groups injected into retrieval results by RAG poisoning attacks: after expanding the retrieval scope, it clusters them using hierarchical agglomerative clustering with a query-adaptive threshold, and prunes suspicious clusters based on relevance and other criteria. Experiments on three question-answering datasets, multiple attacks, and six LLMs show that it substantially reduces the attack success rate while maintaining answer utility under clean retrieval.
Generated by AI from reports · Updated 16 days ago
Attention and developments
Attention, hourly
Now 24, 0% in 24 hours·Peak 24 (Nov 6 04:00)
Related research
This event's paper
- GuardRAG: A Post-Retrieval Defense against Poisoning Attacks in Retrieval-Augmented Generation
arXiv · 2026-10-26 · Deep reading
Progress timeline
- arXivGuardRAG: a post-retrieval defense against retrieval-augmented generation (RAG) poisoning attacks
GuardRAG is a training-free, plug-and-play post-retrieval defense that identifies mutually corroborating suspicious document groups injected into retrieval results by RAG corpus poisoning attacks. It first expands the retrieval scope and removes near-duplicates, then clusters using hierarchical agglomerative clustering (HAC) with query-adaptive thresholds, and prunes suspicious clusters by relevance, cluster size and semantic compactness. Experiments on three question-answering datasets, multiple poisoning attacks and six LLMs show that it substantially reduces attack success rates while maintaining answer utility under clean retrieval.