Perceptions and Practicalities for Private Machine Learning

Abstract

data they and their partners hold while maintaining data subjects' privacy. In this thesis I show that private computation, such as private machine learning, can increase end-users' acceptance of data sharing practices, but not unconditionally. There are many factors that influence end-users' privacy perceptions in this space; including the number of organizations involved and the reciprocity of any data sharing practices. End-users emphasized the importance of detailing the purpose of a computation and clarifying that inputs to private computation are not shared across organizations. End-users also struggled with the notion of protections not being guaranteed 100\%, such as in statistical based schemes, thus demonstrating a need for a thorough understanding of the risk form attacks in such applications. When training a machine learning model on private data, it is critical to understand the conditions under which that data can be protected; and when it cannot. For instance, membership inference attacks aim to violate privacy protections by determining whether specific data was used to train a particular machine learning model. Further, the successful transition of private machine learning theoretical research to practical use must account for gaps in achieving these properties that arise due to the realities of concrete implementations, threat models, and use cases; which is not currently the case.