Sharp Composition Bounds for Gaussian Differential Privacy via Edgeworth Expansion
Qinqing Zheng, Jinshuo Dong, Qi Long, Weijie Su · PubMed · 2020
-differential privacy. In contrast to the existing composition theorems using the central limit theorem, our new privacy bounds under composition gain improved tightness by leveraging the refined approximation accuracy of the Edgeworth expansion. Our approach is easy to implement and computationally efficient for any number of compositions. The superiority of these new bounds is confirmed by an asymptotic error analysis and an application to quantifying the overall privacy guarantees of noisy stochastic gradient descent used in training private deep neural networks.