Highly personal information about individuals is collected and analyzed at an unprecedented scale. The insights to be gained from such analyses have the potential to transform medicine, social science, and technology. However, this potential often goes unrealized when the custodians of data do not have the tools to analyze it while safeguarding individual privacy. Differential privacy provides a framework for guaranteeing strong individual privacy while enabling the systematic design of privacy-respecting algorithms. For these reasons, it is enjoying increasingly widespread adoption in both industry and government.
This research confronts three broad classes of challenges which will enable the wider and safer adoption of differentially private technologies. The first is to understand which statistical inference and learning tasks admit differentially private solutions, and at what cost in computational resources. This research uses connections between privacy, communication complexity, and online learning to give a unified characterization of when privacy is achievable. The second is to develop new algorithmic paradigms, based on computational heuristics and regression algorithms, that will lead to practical solutions to high-dimensional statistical problems. The final objective is to develop new mathematical tools for more precisely understanding the guarantees of differential privacy and assessing its downstream impacts. Such tools will form a critical part of the guidance that the scientific community must provide to policymakers. This research is integrated with an educational plan that includes course development, research training for graduate and undergraduate students, and new K-12 experiential activities around privacy for sensitive survey data.
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.
