SecureCollaRAG: A Byzantine-Fault-Tolerant Secure Collaborative RAG Framework for Resisting Knowledge Corruption in Agent Systems
Combating Knowledge Corruption in Agent Systems: A Byzantine-Tolerant Secure Collaborative RAG Framework
AI summary
SecureCollaRAG is a Byzantine fault-tolerant secure collaborative RAG framework for agent systems, used to defend against knowledge corruption attacks caused by RAG document poisoning. It verifies document sources through a multi-source knowledge verification mechanism and a trust scoring core based on a dynamic GNN, defending against covert knowledge corruption while preserving the integrity of domain knowledge. Evaluation and formal analysis show that the framework remains robust against attackers under non-IID data distributions.