A SURVEY OF DIFFERENTIAL PRIVACY IN LEARNING

PhD Qualifying Examination


Title: "A SURVEY OF DIFFERENTIAL PRIVACY IN LEARNING"

by

Mr. Peng YE


Abstract:

Machine Learning has found applications across various areas. It usually 
involves working with datasets that contain sensitive information, making 
privacy protection crucial for the algorithms. Differential privacy provides 
a rigorous approach to quantifying the privacy leakage of an algorithm, 
making it attract a lot of attention from both academia and industry. It has 
now been widely applied in many privacy-preserving data analysis tasks. 
However, privacy does not come at no cost. An algorithm that satisfies the 
differential privacy property usually suffers from a degradation in 
performance. Thus, it is a central problem to investigate the effect of 
differential privacy on machine learning algorithms. In this article, we 
present a survey of several learning tasks under the constraints of 
differential privacy. In particular, we study private probably approximately 
correct (PAC) learning, prediction, and online learning from an 
information-theoretic aspect. For each problem, we show algorithms and 
hardness results and compare them with conclusions in the non-private setting 
to understand the cost of privacy.


Date:                   Friday, 22 November 2024

Time:                   2:00pm - 4:00pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Prof. Bo Li (Supervisor)
                        Dr. Wei Wang (Co-supervisor)
                        Dr. Shuai Wang (Chairperson)
                        Prof. Ke Yi