Hierarchical Federated Learning for the Next Generation IoT
Merkourios Simos, Pavlos S. Bouzinis, Panagiotis D. Diamantoulakis, Panagiotis G. Sarigiannidis, George K. Karagiannidis · 2022
Federated Learning is a promising decentralized machine learning approach, which has the potential to realize the vision of next-generation internet-of-things (NGIoT), by offering intelligent services and meeting the privacy and low latency requirements. By leveraging the combination of edge servers, as intermediate model aggregators, and the central cloud server, as global model aggregator, the concept of Hierarchical Federated Learning (HFL) has recently emerged. In this paper, we aim to minimize the delay of a global HFL round, under user energy requirements. We jointly optimize the computation and communication resources, as well as the user-edge assignment, in order to minimize the overall delay. The formulated non-convex combinatorial problem, is optimally solved by being decomposed into two disjoint subproblems, namely the resource allocation and user-edge assignment. Finally, the simulation results demonstrate the effectiveness of the proposed methods in terms of delay reduction, compared to selected benchmarks, while insights for the networks' behavior are provided.