Faculty Recruiting Support CICS

Accurate High-Dimensional Data Publication With Differential Privacy

18 Nov
Add to Calendar
Monday, 11/18/2019 9:00am to 11:00am
A311 LGRC
Ph.D. Dissertation Proposal Defense
Speaker: Ryan McKenna

In recent years, differential privacy has seen significant growth, and has been accepted as the dominant privacy definition by the research community. Much progress has been made on designing theoretically principled and practically sound privacy mechanisms, and some of these have even been deployed in the real world. However, differential privacy has not yet seen widespread adoption in the real world. One challenge is that for some problems, there is a gap between the privacy budget required to have a meaningful privacy guarantee and to retain data utility. This gap can be mitigated with better privacy mechanisms, and is something the community is actively working on. Another challenge is that many privacy mechanisms have trouble scaling to high-dimensional data, limiting their applicability to real world data.

In this work, we take significant steps towards addressing these challenges, by designing mechanisms and mechanism components that help close this gap and scale effectively to high-dimensional datasets. We focus on answering a workload of linear queries and related tasks, which encompasses tasks like computing his-tograms, marginals, multi-dimensional range queries, combinations thereof, and more. Mechanisms for linear queries can also be used as subroutines for other tasks, like mean/median estimation, probabilistic modeling, and synthetic data generation.

For utility, we use numerical optimization to optimize expected utility over an expressive class of privacy mechanisms (Chapter 2) and to resolve inconsistencies in noisy measurements produced by privacy mechanisms (Chapter 3). For scalability, we propose novel representations of workloads (Chapter 2) and data distributions (Chapter 3) that require far less space than traditional representations. Together, these contributions are a significant step forward in addressing the stated goal of deploying differential privacy in real-world settings.

Advisor: Gerome Miklau