Evolving Trade-offs Towards Deployable Private Systems for Data Science

Abstract

There is no one-size-fits-all solution to preserving privacy in data science. While insights derived from sensitive data can benefit society, privacy is typically at odds with utility, performance, usability, or some combination of these objectives. Furthermore, each system differs in its definition of these objectives and the way they interact with one another. If the compromise required for any single objective is too great, the system will not be deployed, or worse, will be deployed with a weakened privacy guarantee, exposing users to potential harm. In this work, we address this challenge from multiple angles. First, through strategic algorithm design, our work creates private systems with improved trade-offs, enabling their deployment. This includes a more efficient protocol for the secure inference of deep machine learning models, a novel construction for aggregating key-value data in the local trust model, an evolutionary approach to improve the utility of private clustering, and a user-friendly interpretation of the error of private median queries. Second, we audit private systems to show the privacy risks associated with misleading privacy claims. In particular, through a privacy audit of machine learning, we highlight a difference between expectation and reality in privacy protections.