Doctoral Dissertation Oral Defense, Shuya Feng
Tuesday, June 24, 2025 10:00–11:00 AM
- LocationITE Building
- DescriptionAbstract:Differential privacy (DP) has emerged as the de facto standard for privacy-preserving data analysis, yet significant challenges persist in its practical deployment. This dissertation advances DP mechanisms across three critical dimensions. First, we introduce DPI (Differential Privacy for Infinite Disclosure), a novel technique that effectively bounds privacy leakage in infinite data streams while maintaining high utility through sensitivity compression and an innovative boosting mechanism. Second, UDP-FL (Universally Harmonizing Differential Privacy Mechanisms for Federated Learning) presents a framework that integrates diverse DP mechanisms to achieve superior privacy-utility tradeoffs and faster convergence, introducing mode connectivity-based analysis for convergence evaluation and demonstrating remarkable resilience against privacy attacks. Finally, we propose a Multi-Stakeholder Usability Study of DP-SGD, examining how data contributors, machine learning engineers, and compliance officers understand, implement, and verify differential privacy guarantees in machine learning deployments. Through mixed-methods research involving structured tasks, surveys, and interviews, this study aims to bridge the gap between theoretical guarantees and practical implementation.
- Websitehttps://events.uconn.edu/engineering/event/1125025-doctoral-dissertation-oral-defense-shuya-feng
- CategoriesConferences & Speakers