CAREER: Theory and Practice of Privacy-Utility Tradeoffs in Enterprise Data Sharing

A cooperative and coordinated effort across organizations will be required for effectively addressing many of today's biggest global challenges and societal problems, including cybercrime, climate change, and public health. Data sharing will enable organizations involving academic researchers, governments, and industry to collaborate more effectively on solving such problems.

However, in practice data sharing among organizations is hindered by two significant factors. (1) Organizations often fear inadvertently leaking trade-secrets, such as business strategies, and therefore do not share data; this concern is called trade-secret privacy. (2) When organizations have access to shared data, they often lack the in-house resources to evaluate the quality and usefulness of the data-source; this problem is called data source utility. This project quantitatively addresses the data sharing problems in terms of these two factors.

Specifically, the project will develop quantitative methods for measuring trade-secret privacy and data-source utility, assessing the tradeoffs between these two factors, and developing algorithms that come close to optimizing this tradeoff. This research will help encourage greater data sharing among organizations, thereby enhancing society's ability to address global challenges through informed collaboration.

Several outreach and education activities complement and integrate the research. These include working with companies to develop privacy protections in their applications and services, working on an open source library for trade-secret privacy and utility, and organizing research internship programs for students in Africa.

This project aims to design novel privacy and utility metrics and frameworks to help organizations make more informed choices regarding data sharing. Both of the above problems (trade secret privacy and data source utility) can be framed as a study of divergences between probability distributions. Building on the investigator's prior work studying divergences in the context of deep generative models, this project will study how to carefully select appropriate divergence measures to (a) satisfy enterprise use cases, and (b) provide strong theoretical guarantees of privacy and utility.

The project will proceed in four thrusts. Thrust 1 will define and analyze a metric for trade secret privacy. This metric will be based on the notion of maximal leakage from information theory; maximal leakage captures the maximum amount of information that can be gained by an adversary about any secret quantity after seeing released, obfuscated data.

The proposed metric in this project will differ by modeling information leakage of specific trade secrets, rather than any arbitrary secret. Thrust 2 will propose and theoretically analyze a metric for data source utility, based on statistical divergences over probability distributions. This work will build on the expansive literature on data valuation.

Thrust 3 will study fundamental tradeoffs between these metrics; the goal will be to identify algorithms that approach the fundamental bounds. Thrust 4 will analyze downstream performance guarantees, which connect the proposed privacy and utility metrics to enterprise use cases motivated by the investigator's ongoing industry collaborations. In summary, the project will contribute a formal methodology for modeling and mitigating common data sharing problems in enterprise settings.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.