SecureCollaRAG: Byzantine-Tolerant Framework for Secure Collaborative RAG
The paper "SecureCollaRAG" introduces a Byzantine-tolerant framework designed for secure collaborative retrieval-augmented generation (RAG). This innovative framework addresses knowledge corruption attacks in agent systems by implementing a Multi-source Knowledge Validation Mechanism. It also utilizes dynamic graph neural networks (GNN) for credibility scoring, which ensures document provenance verification. By effectively preventing stealthy attacks, SecureCollaRAG maintains the integrity of domain knowledge and showcases resilience against adversaries, even in non-IID data environments. The paper has been submitted to arXiv under the category of Computer Science > Cryptography and Security, with the ID 2608.04366.
Key facts
- Paper proposes SecureCollaRAG, a Byzantine-tolerant collaborative RAG framework.
- Addresses knowledge corruption attacks in agent systems.
- Uses Multi-source Knowledge Validation Mechanism.
- Employs dynamic GNN-based credibility scoring for document provenance verification.
- Prevents stealthy knowledge corruption attacks while preserving domain knowledge integrity.
- Demonstrates robustness against attackers under non-IID data distributions.
- Submitted to arXiv under Computer Science > Cryptography and Security.
- arXiv ID: 2608.04366.
Entities
Institutions
- arXiv
- arXivLabs