Privacy-preserving Sanitization in Data Sharing

02 May
Friday, 05/02/2014 5:30am to 7:30am
Ph.D. Thesis Defense

Wentian Lu

Computer Science Building, Room 151

In the era of information explosion, data sharing is a common and everyday activity for our generation. When people request shared data for the purpose of monitoring, understanding and analyzing, it's not surprising that privacy concerns could be one of main barriers that disrupt such sharing behaviors, due to the fear of disclosing sensitive information.  Sharing data with privacy protection is an important and realistic problem in the real world. 

In this dissertation, we consider the process of data sanitization that disguise the sensitive information before sharing a dataset. Our designing goal for sanitizing methods is always protecting privacy while preserving utility as much as possible, despite of varying data sources and privacy requirements. In particular, we first propose a framework for sharing a database under retention policies for auditing purpose. When obeying a retention policy often results in the wholesale destruction of the audit log in existing solutions, our framework allows to expire data at finer granularity and supports audit queries with incompleteness in a database. Secondly, we solve the problem of untrusted system evaluation using shared database synthesis under differential privacy. Our synthetic database accurately preserves the core performance measures of a given query workload, and satisfies differential privacy with crucial extensions to multi-relation databases. Lastly, we consider modeling graph under differential privacy and describe algorithms for sharing exponential random graph estimations. Our solution employs a decomposition of the estimation problem into two steps: getting private sufficient statistics first and then estimating the model parameters.  We show that our privacy mechanism provides provably less error than comparable methods and our redesigned estimation algorithm offers better accuracy. 

Advisor: Gerome Miklau