Incentive Mechanism Based on Double Auction for Federated Learning in Satellite Edge Clouds

As data-driven applications proliferate, satellite edge clouds (SEC) consisting of low-earth-orbit (LEO) satellites have demonstrated great potentials to achieve global connectivity. However, communication cost between satellites and ground stations is still high compared to terrestrial mobile networks, and it is difficult to preserve data privacy. Federated learning (FL) thus is emerging as a promising technique to enable distributed machine learning on various FL participants and global model sharing among the participants, which significantly reduces the communication cost without data leakage. As FL advocates training global models using large-scale numbers of distributed participants, it is quite crucial to design incentive mechanisms to inspire participants to contribute their data. In this paper, we study incentive mechanism design issues by formulating a utility-aware FL problem in an SEC network. We then design a double-auction mechanism for the problem with a fixed number of participating satellites for FL jobs. We further devise a repeated double auction mechanism with various numbers of participating satellites for FL jobs. We finally evaluate the performance of the proposed mechanisms against existing methods by simulations, and simulation results show that the proposed mechanisms can achieve higher admission ratio and utility than their counterparts.