ETD

• interpretable model
• local explainer
• membership inference attack
• privacy risk

The purpose of this thesis is to investigate how the privacy risk of a machine learning model seen as a black box changes when a more interpretable model is used to explain and locally approximate the behavior of the black box during the implementation of a Membership Inference Attack, an attack on a neural network that allows the attacker to determine if a particular datum was part of the training set used to train the black box.