Multi-Stage Hybrid Federated Learning Over Large-Scale D2D-Enabled Fog Networks

Federated learning has generated significant interest, with nearly all works focused on a "star" topology where nodes/devices are each connected to central server. We migrate away from this architecture and extend it through the network dimension case there multiple layers of nodes between end devices Specifically, we develop multi-stage hybrid federated (MH-FL), intra- inter-layer model that considers as multi-layer cluster-based structure. MH-FL structures among in clusters, including local networks formed via device-to-device (D2D) communications, presumes semi-decentralized for learning. It orchestrates at different collaborative/cooperative manner (i.e., using D2D interactions) form consensus parameters combines parameter relaying tree-shaped hierarchy. derive upper bound convergence respect (e.g., spectral radius) algorithm number rounds clusters). obtain set policies clusters guarantee either finite optimality gap or global optimum. then distributed control tune cluster over time meet specific criteria. Our experiments real-world datasets verify our analytical results demonstrate advantages terms resource utilization metrics.