When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse
Poisoned RAG answers can look more confident, not less, so the paper looks inside the model’s attention instead.
The authors identify “Attention Collapse,” where generation attention concentrates on poisoned documents instead of staying dispersed across benign context. They propose D-SCAN, a lightweight detector that monitors document-level attention dynamics. The abstract says it works across multiple attack benchmarks and can flag attacks even when the final answer is unchanged. HF Daily Papers' note
The authors identify “Attention Collapse,” where generation attention concentrates on poisoned documents instead of staying dispersed across benign context. They propose D-SCAN, a lightweight detector that monitors document-level attention dynamics. The abstract says it works across multiple attack benchmarks and can flag attacks even when the final answer is unchanged. HF Daily Papers' note
score 5