Model Inversion Attack

Definition: A model inversion attack is when an attacker uses access to a machine learning model to reconstruct sensitive information about the training data, effectively “inverting” the model to reveal private details.

Context/Example: With access to a facial recognition model, an attacker queries it repeatedly and uses the output probabilities to reconstruct an approximate image of a person whose face was used to train the model.

Related Concepts: Reconstruction Attacks, Overfitting, White-Box vs Black-Box Models