DIFFERENTIALLY PRIVATE MEAN AND COVARIANCE ESTIMATION

PhD Qualifying Examination


Title: "DIFFERENTIALLY PRIVATE MEAN AND COVARIANCE ESTIMATION"

by

Miss Yuting LIANG


Abstract:

Differential Privacy (DP) has become the definition of choice for privacy 
protection due to its strong guarantee and immunity to reverse-engineering. 
Roughly speaking, an algorithm is differentially private for a query if the 
presence or absence of any particular record will not impact the query output 
significantly. However, for the simple mean query, the empirical mean is 
inherently unstable relative to change in even a single record.

In this survey, we discuss works in the most fundamental data analytics tasks: 
that of mean and covariance estimation. In particular, we present the works in 
both the empirical and statistical settings, and from univariate to 
high-dimensional datasets. We start by introducing the basic tools of DP, and 
discuss how different works deal with the potentially unbounded sensitivities 
using these tools. We also discuss how some of the univariate estimators are 
extended to work in high dimensions, which (fortunately) are not by simple 
coordinate-wise estimation. We conclude with some potential research directions 
for future work.


Date:  			Tuesday, 26 July 2022

Time:                  	9:00am to 11:00am

Zoom Meeting: 		https://hkust.zoom.us/j/5688510375

Committee Members:	Prof. Ke Yi (Supervisor)
 			Prof. Mordecai Golin (Chairperson)
 			Prof. Siu-Wing Cheng
 			Prof. Yuan Yao (MATH)


**** ALL are Welcome ****