RADAR: Defending RAG Dynamically against Retrieval Corruption

2026-05-21

AI Summary

This work addresses the vulnerability of dynamic Retrieval-Augmented Generation (RAG) systems to time-varying retrieval poisoning attacks in open-web environments, a challenge inadequately mitigated by existing static defenses that struggle to balance evolutionary adaptability with storage efficiency. The authors formulate reliable context selection as an energy minimization problem on a graph, solved exactly via Max-Flow Min-Cut optimization, and integrate Bayesian memory nodes that recursively update belief states to jointly enhance adversarial robustness and knowledge evolution. Evaluated on a novel dynamic dataset, the proposed method significantly outperforms baseline approaches, achieving simultaneous improvements in system robustness and response quality while incurring minimal storage overhead.

Abstract

While RAG systems are increasingly deployed in dynamic web search, temporal volatility amplifies their vulnerability to adversarial attacks. Existing static-oriented defenses struggle to handle evolving threats and incur prohibitive storage costs in dynamic settings. We propose RADAR, a framework that models reliable context selection as a graph-based energy minimization problem, solved exactly via Max-Flow Min-Cut. By incorporating a Bayesian memory node, RADAR recursively updates a belief state instead of archiving raw historical documents, effectively balancing stability against attacks with adaptability to genuine knowledge shifts. Experiments on a novel dynamic dataset show that RADAR achieves superior robustness and response quality with minimal storage overhead compared to the baselines.

Problem

• RAG
• adversarial attacks
• dynamic environments
• retrieval corruption
• storage overhead

Innovation

• RAG
• retrieval corruption
• graph-based energy minimization
• Max-Flow Min-Cut
• Bayesian memory