Novel Federated Learning by Aerial-Assisted Protocol for Efficiency Enhancement in Beyond 5G Network

Federated Learning (FL) is an important Machine Learning (ML) technique in a decentralized computing environment. For FL in the 5
^th
generation (5G) and beyond networks, there is a need to improve the link capacity, latency, and reliability for every iteration. This would help the user equipment (UE) to efficiently download the global ML model from the FL server, upload the partially trained model from the UE, and update the global ML model hosted on the FL server. However, due to limited coverage areas of terrestrial mobile networks and limited bandwidth at the periphery of a terrestrial cell, the UEs may be out of coverage or have low bandwidth to support the iterations for FL, lowering the performance of FL. To enhance the performance, we propose using aerial cells to augment the existing terrestrial infrastructure and offer enhanced coverage and capacity to the operators. There has been limited work on the management of aerial links in an augmented deployment to improve the performance of FL. To the best of our knowledge, for the first time, we present and analyze a novel Federated Learning by Aerial-Assisted Protocol (FLAP) that selects and manages the aerial link for the federated UEs in each training iteration. FLAP will improve the reliability and latency of the FL process. Through call flow analysis and extensive simulation, we have shown an increase of 50% in the number of federated UEs taking part in FL. Our results reveal that FLAP enables a much faster convergence for an image classification model with much higher accuracy.