Gaussian Differential Privacy (GDP)
Abstract
Data privacy has been put on firm theoretical foundations since the birth of differential privacy. The fruitful decade has witnessed many theoretical and practical successes, including the remarkable deployment in Chrome and iOS. However, a major obstacle to further applications is the extra sophistication in various theorems in differential privacy, most notably composition and subsampling. Precise characterization of these properties is of vital importance in practice. We propose a hypothesis testing based framework called f-DP, which includes Gaussian differential privacy (GDP) as the most useful special case. All desired properties admit elegant and tight statements in this framework. In addition, a central limit theorem is proved and shows the universality of GDP. As a comprehensive application, we show how the tools we develop improves the analysis of the privacy of perturbed stochastic gradient descent.
Based on joint work with Aaron Roth and Weijie Su.
Time
February 26th, 2019
15:00 ~ 16:30
Speaker
Jinshuo Dong, PhD at University of Pennsylvania
Room
Room 102, School of Information Management & Engineering, Shanghai University of Finance & Economics
Abstract
Data privacy has been put on firm theoretical foundations since the birth of differential privacy. The fruitful decade has witnessed many theoretical and practical successes, including the remarkable deployment in Chrome and iOS. However, a major obstacle to further applications is the extra sophistication in various theorems in differential privacy, most notably composition and subsampling. Precise characterization of these properties is of vital importance in practice. We propose a hypothesis testing based framework called f-DP, which includes Gaussian differential privacy (GDP) as the most useful special case. All desired properties admit elegant and tight statements in this framework. In addition, a central limit theorem is proved and shows the universality of GDP. As a comprehensive application, we show how the tools we develop improves the analysis of the privacy of perturbed stochastic gradient descent.
Based on joint work with Aaron Roth and Weijie Su.
Time
February 26th, 2019
15:00 ~ 16:30
Speaker
Jinshuo Dong, PhD at University of Pennsylvania
Room
Room 102, School of Information Management & Engineering, Shanghai University of Finance & Economics
