NSF-BSF: satc 2.0: RES: formal privacy models: new horizons

The National Science Foundation Division of Computer and Network Systems awarded Georgetown University $550,000 on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop formal privacy models that extend differential privacy beyond worst-case guarantees and address the utility-privacy gap in modern data analysis.

The project develops privacy-preserving data analysis paradigms intended for use by researchers, industry, and regulators. Research activities focus on preventing specific privacy vulnerabilities—reconstruction and singling-out attacks—when differential privacy creates prohibitive utility costs. The team designs novel algorithmic approaches including trust settings that leverage public pre-training data and conditional utility under realistic data distributions to expand the utility frontier of standard differential privacy. The project trains doctoral students and postdoctoral fellows, integrates findings into academic curricula, and disseminates research materials publicly.

Performance runs from October 1, 2026, through September 30, 2030, at Georgetown University in Washington, DC 20057.

Abstract

A critical utility-privacy GAP undermines modern data analysis. This GAP has been a focus of the study of differential privacy (DP), a formal mathematical framework introduced in 2006. The study of DP provided US with tools for understanding aspects of the utility-privacy GAP, such as the cumulative privacy loss inherent in multiple data uses, and has influenced standards across computer science, statistics, and machine learning. However, a significant tension still remains between DP theory and the constraints of real-world applications: while compatible with DP in principle, making statistical and machine learning tasks satisfy the requirements imposed by DP often entails dramatic utility degradation or prohibitive sample complexity, forcing deployments to use large privacy parameters that dilute the formal guarantees. The project novelty is in shifting the scope of formal privacy models beyond the DP worst-case guarantees. Beyond its intellectual contributions to computer science, the project's significance and importance are in developing new privacy-preserving data analysis paradigms to be used by researchers and industry and in providing regulators and privacy-law scholars with tools to address our ever-changing data ecosystem. The project team trains PH.d. Students and post-doctoral fellows, integrates findings into academic curricula, and disseminates all research materials to the wider public and research community. To expand the horizons of formal privacy models, the project develops new privacy-related concepts that focus on preventing specific privacy vulnerabilities such as reconstruction and singling-out attacks when DP is utility-prohibitive. The project expands the utility frontier of standard DP by introducing novel trust settings, leveraging public pre-training data, and designing algorithms that offer conditional utility under realistic data distributions. Finally, the project establishes a unique contingency framework to model the economics and attack complexity in a post-privacy world, e.g., devising techniques that force adversarial costs to scale with the volume of leaked 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.- subawards are not planned for this award.