Algorithms and frameworks for preventing privacy leakage and overfitting in machine learning

Abstract

Machine learning algorithms aim to learn useful models for the population, only having access to a training dataset as its proxy. However, they may overly tailor their output to the training dataset, which leads to the following two types of unwanted behavior: 1. they may leak too much information about specific datapoints included in the training dataset, violating the privacy of the individuals who participate in it, or 2. they may "overfit" their training dataset, that is, their empirical performance on the training dataset does not generalize well to their true performance on the population. This thesis contributes tools for preventing or reasoning about these two unwanted behaviors.