Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)

Graduate Group

Computer and Information Science

First Advisor

Michael Kearns

Second Advisor

Aaron Roth


Computing technologies today have made it much easier to gather personal data, ranging from GPS locations to medical records, from online behavior to social exchanges. As algorithms are constantly analyzing such detailed personal information for a wide range of computations, data privacy emerges as a paramount concern. As a strong, meaningful and rigorous notion of privacy, Differential

Privacy has provided a powerful framework for designing data analysis algorithms with provable privacy guarantees. Over the past decade, there has been tremendous progress in the theory and algorithms for differential privacy, most of which consider the setting of centralized computation where a single, static database is subject to many data analyses. However, this standard framework does not capture many complex issues in modern computation. For example, the data might be distributed across self-interested agents, who may have incentive to misreport their data; and different individuals in the computation may have different expectations to privacy.

The goal of this dissertation is to bring the rich theory of differential privacy to several computational problems in practice. We start by studying the problem of private counting query release for high-dimensional data, for which there are well-known computational hardness results. Despite the worst-case intractability barrier, we provide a solution with practical empirical performances by leveraging powerful optimization heuristics. Then we tackle problems within different social and economic settings, where the standard notion of differential privacy is not applicable. To that end, we use the perspective of differential privacy to design algorithms with meaningful privacy guarantees.

(1) We provide privacy-preserving algorithms for solving a family of economic optimization problems under a strong relaxation of the standard definition of differential privacy---joint differential privacy.

(2) We also show that (joint) differential privacy can serve as a novel tool for mechanism design when solving these optimization problems: Under our private mechanisms, the agents are incentivized to behave truthfully.

(3) Finally, we consider the problem of using social network metadata to guide a search for some class of targeted individuals (for whom we cannot provide any meaningful privacy guarantees). We give a new variant of differential privacy---protected differential privacy---that guarantees differential privacy only for a subgroup of protected individuals. Under this privacy notion, we provide a family of algorithms for searching targeted individuals in the network while ensuring the privacy for the protected (un-targeted) ones.