Topology Learning for Heterogeneous Decentralized Federated Learning over Unreliable D2D Networks

This paper presents a study on decentralized federated learning (DFL) in device-to-device (D2D) networks. DFL, a system that allows devices to exchange parameters with neighboring nodes, offers a solution to communication bottlenecks often experienced with centralized systems. However, DFL faces challenges such as data heterogeneity and transmission outages caused by the User Datagram Protocol (UDP) in D2D networks. The authors propose a novel method, ToLRDUL, to address these issues. The method, which was validated in experiments, improved the speed and accuracy of convergence.

Publication date: 22 Dec 2023
Project Page: ?
Paper: https://arxiv.org/pdf/2312.13611