RAG-leaks: Difficulty-Calibrated Membership Inference Attacks on Retrieval-Augmented Generation

RAG-leaks demonstrates that retrieval difficulty can leak information about whether a sample was part of the training corpus.

Recommended citation: G Wang, J He, H Li*, M Zhang, D Feng. (2025). "RAG-leaks: Difficulty-Calibrated Membership Inference Attacks on Retrieval-Augmented Generation." Science China Information Sciences.