Aggregation and Estimation Under 'Partial' Differential Privacy

Master's Thesis Project

Guides: Prof. Nikhil Karamchandani and Prof. Bikash Dey
Duration: June 2022-Present

In this project, we consider two different problem settings: groupwise aggregate estimation and joint distribution estimation.

The problem of estimation of groupwise aggregate estimation occurs very naturally in settings such as polling and medical testing. There is a central entity that desires to estimate aggregates of values held by individuals over groups that partition a population. However, the group of an individual is often determined by very sensitive information, such as a person's religion, age, gender, etc. It is therefore desirable to keep such information private. We consider the groupwise aggregate estimation problem (first proposed by Naim et al.) under a differential privacy constraint that restricts the distinguishability of groups based on any comminucation sent to the server.

We also generalize this problem to the joint distribution estimation problem, where each individual holds bivariate, finite-valued data generated by some unknown underlying joint distribution. The central server wishes to find this distribution, but under a similar 'partial' privacy constraint as the aggregate estimation problem.