Quantifying Membership Privacy via Information Leakage

Machine learning models are known to memorize the unique properties of individual data points in a training set. This memorization capability can be exploited by several types attacks infer information about data, most notably, membership inference attacks. In this paper, we propose an approach based on leakage for guaranteeing privacy. Specifically, use conditional form notion maximal quantify leaking entries dataset, i.e., entrywise leakage. We apply our privacy analysis Private Aggregation Teacher Ensembles (PATE) framework privacy-preserving classification sensitive and prove that its aggregation mechanism is Schur-concave when injected noise has log-concave probability density. The Schur-concavity implies increased consensus among teachers labeling query reduces associated cost. Finally, derive upper bounds uses Laplace distributed noise.