EL-MIA: Quantifying Membership Inference Risks of Sensitive Entities in LLMs

Ali Satvaty, Suzan Verberne, Fatih Turkmen

Membership inference attacks (MIA) aim to infer whether a particular data point is part of the training dataset of a model. In this paper, we propose a new task in the context of L…