Maximal Information Leakage based Data-driven Models for Privacy in Data Disclosure Applications

We consider the problem of privacy preservation in disclosure of data sets, and use maximal information leakage as a privacy measure. This research problem is motivated by recent data breaches and an increasing concern for data privacy for individuals and entities who wish to publish data. We show that it achieves the same result as minimizing an adversary’s accuracy in a theoretical data setting with a Bernoulli-Gaussian data model, and the same result is shown with a numerical approximation of Sibson mutual information. We validate this theoretical result with machine learning models on synthetically generated data following our theoretical model, and show it is a suitable privacy measure. Our data-driven models use a numerical proxy of maximal information leakage, and it out-performs conventional mutual information in experiments for several data sets.M.A.S