Membership Inference — Detecting Training Data From Model Behavior

Why it matters

Membership inference threatens training data privacy. Understanding shapes what data can be safely used for training.

The architecture

Compare model behavior: training data typically has higher confidence/lower loss than non-training.

Statistical tests distinguish based on scores.

How it works end to end

Mitigations: differential privacy training (adds noise, provably bounds inference). Deduplicate training data (reduces memorization). Regularization.

Evaluation: measure inference attack success rate on your model.