A membership inference attack determines whether a particular data point was included in a model’s training data. It turns a model into a way of testing for possible exposure of individual records, making it useful for evaluating privacy risks and data leakage. section 3.4
The technique addresses unintended memorization in large language models (LLMs), but it does not by itself establish why information was exposed or guarantee that a suspected record was memorized. It belongs alongside defenses such as differential privacy and training changes designed to reduce verbatim reproduction. section 3.4
Testing a training-set record
Given a specific data point and access to a trained model, an evaluator can assess whether that point was likely part of the model’s training set, treating the result as evidence of a possible privacy risk. section 3.4
