Membership inference
Also known as: membership inference attack
Membership inference is an attack where you probe a model to learn whether a particular person or record was in its training data.
- If you can tell “this person was in the dataset”, you’ve inferred something private.
- A privacy attack, not a correctness one.
- One of the motivations for differential privacy.
Why it matters: even if a model never *outputs* PII, it may *reveal* who was in the data. That's a privacy property to test, not assume.