RADAR: Defending RAG Dynamically against Retrieval Corruption

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 By incorporating a Bayesian memory node, RADAR recursively updates a belief state instead of archiving raw historical documents, effectively balancing stability against attacks with adaptability to genuine knowledge shifts.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.

AI systems that retrieve live information from the web to answer questions are vulnerable to clever tricks, especially when the information changes rapidly. Existing defenses either can’t keep up with evolving attacks or require storing huge amounts of old data, which is impractical. We created RADAR, a new approach that turns the problem of finding trustworthy information into a kind of mathematical balancing act, solved efficiently without storing everything. Instead of keeping all past web pages, RADAR maintains a compact “belief state” that updates as new information arrives, helping it resist attacks while still adapting to real changes in the world. In tests with a realistic, time‑varying dataset, RADAR proved far more robust and delivered better answers than current methods, all while using minimal storage. This makes it possible to deploy safer, more efficient AI in fast‑moving environments like news feeds or social media.