Membership Inference Attack

Definition: A membership inference attack occurs when an adversary attempts to determine whether a specific data point was included in a machine learning model’s training dataset.

Context/Example: An attacker observes a model’s confidence scores and notices that it responds more confidently to data it has seen during training. By comparing the model’s output on a given input, the attacker can infer whether that individual’s data was part of the training set.

Related Concepts: Overfitting, Privacy Leakage, Differential Privacy, Model Generalization