PRA-RAG Defends Retrieval-Augmented Generation Against Poisoning Attacks

PRA-RAG utilizes geometric structures in embedding spaces to identify robust subsets of retrieved text, providing theoretical bounds on the impact of poisoned data. The algorithm mitigates RAG vulnerabilities by sampling multiple text combinations to derive a stable aggregated representation.