CSIS Seminar

Differential Privacy without Centralization

Abstract

Fields like machine learning and statistics have flourished with the accessibility of personal data. To address concerns about such analysis, computer scientists have identified special families of data analysis algorithms. One of these is the family of differentially private (DP) algorithms, which allow individuals to quantify the risk of contributing their data. A textbook instantiation of DP is centralized: it requires data owners to trust that a data collector runs the correct algorithm and does nothing else with data. In this talk, I advocate for decentralized DP, where no single collector no longer has direct access to data and computations are done by multiple parties. My work proves that access to a shuffler---an entity that anonymizes communication---allows us to replicate some of the success in the centralized setting. I look forward to continuing my theoretical work, while also pursuing real-world implementations and exploring other benefits of decentralization.

Speaker Bio

Albert Cheu is a computer scientist working on differential privacy, with a particular interest in its intersections with cryptography and security. He is a postdoctoral fellow working in the Department of Computer Science at Georgetown University. He earned his PhD. at Northeastern University's Khoury College of Computer Science.