Linglong Kong
University of Alberta
Room: M3 3127
Online differentially private federated inference with untrusted server
Traditional static federated learning in distributed environments is facing new challenges, with data streaming and privacy emerging as significant concerns. This paper addresses the problem of online statistical inference in a federated learning framework with an untrusted central server, adhering to privacy constraints. Under the rigorous local differential privacy constraint, we consider a general optimization-based framework with streaming data. We introduce a noisy local stochastic gradient descent algorithm and its variants to study synchronous and asynchronous federated learning scenarios without and with delays in local updates, respectively. Our proposed algorithms are single-pass, depending only on the current data and the previous estimate, which effectively reduces both time and space complexity. We also establish the convergence rates and functional central limit theorems for the proposed algorithms, providing theoretical justification for our online inference tools. Specifically, for both scenarios, we present two online inference methods: private plug-in and random scaling, to construct private confidence intervals efficiently in an online manner. Our numerical analyses assess the performance of the proposed methods across various conditions, including time heterogeneity, sample size, and privacy budget, recommending settings where each approach excels.