Professor, Department of Computer Science UofT
Theory of Computing
Title: Differential Privacy and Fairness in Classification
The problem of statistical disclosure control--revealing accurate statistics about a population while preserving privacy of individuals--has a long and interesting history. Differential privacy is a new mathematical framework developed within the last ten years for data privacy. The key idea is that the data curator, who holds confidential data, must respond to statistical queries of the database in a way that guarantees the privacy of the subjects.
In this talk, we will discuss the goals of privacy, and some pitfalls with common and previously used methods. We will then present the differentially private framework, surveying the basic technology, and the positive and negative aspects of this approach. We will discuss our recent work extending the basic framework to new settings: the distributed setting, and the online setting.
Recent work in the area has explored a related objective: fairness. The goal here is to classify individuals while preventing discrimination. I will present our new framework for fair classification, and its connection to differential privacy.
About the speaker:
Toniann Pitassi is an expert in computational complexity and proof complexity, where the goal is to understand the inherent limitations of proofs and computation. The most famous problem in the area, the P versus NP problem, is the driving force behind much of Pitassi’s research. Pitassi’s research has led to lower bounds for many proof systems, using techniques from model theory, combinatorics and communication complexity. Pitassi’s other research aims to develop computational frameworks for privacy and fairness. She is the recipient of several awards, including an NSF Young Investigator award, and a Premiere’s Research Excellence Award.