GRADA: Graph-based Reranker against Adversarial Documents Attack

Retrieval-Augmented Generation (RAG) improves answer quality by pulling in external documents, but that retrieval step can be gamed. GRADA focuses on the weakest link: adversarial documents that look similar to a query but are misleading.

GRADA introduces a graph-based reranking layer that prioritizes documents that are coherent with the rest of the retrieved set, not just the query. This makes it harder for adversarial inserts to rise to the top while keeping relevant evidence intact.

In experiments across multiple LLMs, the approach sharply reduces attack success rates while maintaining minimal accuracy loss on downstream tasks. That makes it a practical, drop-in defense for production RAG systems.

Paper: https://arxiv.org/abs/2505.07546