Laith Zumot

This paper demonstrates that retrieval-augmented in-context learning systems for document QA are vulnerable to membership inference attacks, proposing novel black-box methods that exploit query prefixes to distinguish training data members.