A membership inference attack is a privacy attack that tries to determine whether a specific record was used to train a machine learning model. The attacker compares the model’s outputs, confidence scores, or behavior on similar inputs and looks for signals that suggest the target example was “seen” during training.
This matters because training data can include personal, sensitive, or proprietary information. If a model reveals that a record was part of its dataset, it can expose private facts or confirm the presence of a person’s data in an application. Defenders use membership inference testing to measure privacy risk, especially after retraining, fine-tuning, or machine unlearning. Common mitigations include limiting output detail, reducing overfitting, using differential privacy, and checking whether model behavior still leaks training-set membership.