Title: FedHAP: Fast Federated Learning for LEO Constellations using Collaborative HAPs

Abstract: Low Earth Orbit (LEO) satellite constellations have seen a surge in deployment over the past few years by virtue of their ability to provide broadband Internet access as well as to collect vast amounts of Earth observational data that can be utilised to develop AI on a global scale. As traditional machine learning (ML) approaches which train a model by downloading satellite data to a ground station (GS) is not practical, Federated Learning (FL) offers a potential solution. However, existing FL approaches cannot be readily used because of excessively prolonged training time and unreliable satellite-GS communication channels. In this paper, we propose FedHAP by introducing high-altitude platforms (HAPs) as distributed parameter servers (PSs) into FL for Satcom (or more concretely LEO constellations), to achieve fast and efficient model training. FedHAP consists of three components: 1) a layered communication topology, 2) a model propagation algorithm, and 3) a model aggregation algorithm. Our extensive simulations demonstrate that FedHAP significantly accelerates FL model convergence as compared to state-of-the-art baselines, cutting the training time from several days down to a few hours yet achieving higher accuracy.