Fully Adaptive Composition in Differential Privacy - Advanced Privacy Guarantees

Watch a research lecture from the Symposium on Foundations of Responsible Computing (FORC) 2022 exploring advanced concepts in differential privacy composition. Dive into Steven Wu's presentation from Carnegie Mellon University as he discusses groundbreaking work on fully adaptive composition, which allows for dynamic selection of both algorithms and privacy parameters during database interactions. Learn about privacy filters and odometers as probabilistic tools for measuring privacy in adaptive scenarios, and discover how recent advances in time-uniform martingale concentration enable tighter privacy guarantees. Understand how these new approaches match the effectiveness of traditional advanced composition theorems while offering greater flexibility, with results that suggest minimal trade-offs when implementing fully adaptive privacy. Follow along as Wu examines the theoretical foundations, practical implications, and potential limitations of these novel privacy-preserving techniques.