Methods
Membership inference attack
Also called: MIA, replica rate
A membership inference attack tries to tell whether a specific person's record was used to train a model. It is a standard privacy test for synthetic data.
In a membership inference attack (MIA), an attacker who has a real patient's record tries to determine whether that record was in the training data of a model or synthetic dataset. Success means the model leaks information about who was in the data.
For synthetic data, attacks often use distance: if synthetic records sit unusually close to training records compared with unseen records, membership can be inferred. An attacker AUC near 0.5 means the attacker cannot tell members from non-members.
A related check is the replica rate, the share of synthetic records that are exact or near-exact copies of real ones, which should be zero.
Tools that use this
Related terms
Educational reference. Not medical advice and not for clinical decision-making.