Membership-Invariant Subspace Training

Membership-Invariant Subspace Training (MIST) is a method for training classifiers that acts as a defense designed to specifically defend against black-box membership inference attacks on the most vulnerable instances in the training data.