Membership Inference Attack

A method for testing whether a specific example was part of a model's training set, typically by exploiting a model's tendency to behave with higher confidence on data it was trained on. The primary available tool for probing whether unlearning worked, though passing a specific test does not prove zero remaining influence.

An AI-specific security threat in which an adversary attempts to determine whether a specific individual's data was included in a model's training set, a risk category a PIPIA's threat analysis for an AI system should address alongside conventional data-security threats.

Determining whether a specific individual's data was in a model's training set, potentially exposing sensitive associations.

An attack in which an adversary, given access to a trained model or its outputs and a candidate record, attempts to determine whether that record was part of the model's training data, with better-than-chance accuracy. A key test for whether a generator's output can leak information about specific individuals.

An attack that determines whether a specific record was used to train a model, without necessarily recovering the record's content, typically by exploiting the model's higher confidence or lower loss on data it has seen during training.