Membership Inference Attack

An attack that determines whether a particular person's recording was part of the dataset used to train a shared decoder or neural model, typically by exploiting the model's overconfidence on data it has already seen. For neural data the mere fact of membership can be sensitive — it may reveal that someone is a patient in a specific clinical study or carries a particular diagnosis.

Membership inference is a standard benchmark for whether a training pipeline leaks information about its participants, and it is precisely the failure mode that differential privacy is designed to bound. Overparameterized deep decoders trained on small neural cohorts are especially exposed, because they can memorize individual sessions.