Robust and Privacy-Preserving Decentralized Online Federated Learning for Streaming Data With Outliers

Abstract

This article addresses the challenging problem of online federated learning (FL) over streaming data in a decentralized communication network. To enable rapid adaptation to new observations, we develop novel nonparametric model-based local training, not deep neural network (DNN)-based approaches as adopted in most previous studies. In particular, we integrate Gaussian process regression (GPR) with a Student-t likelihood to improve robustness against data outliers. For global model aggregation, we propose a consensus-based product of experts (PoE) algorithm that enables peer-to-peer fusion of nonparametric local models and preserves robustness to outliers. To ensure privacy, we develop a secure aggregation scheme that combines Shamir’s secret sharing (SSS) with public-key encryption. Compared to existing methods, the proposed approach enhances privacy guarantees for learners with limited connectivity in sparse graphs. Theoretical analyses establish robustness, correctness, and privacy properties. Extensive numerical experiments validate the effectiveness of the proposed algorithm.