Federated Learning Over LEO Satellite Networks for Scalable and Secure Global IoT Connectivity

Abstract

The triple threat of Security, Resiliency, and Connectivity is due to the rapidly increasing finite resource such as the Internet of Things (IoT) communications to service unserved and remote regions. A possible solution is Low Earth Orbit (LEO) satellite constellations, providing low-latency and blanket coverage everywhere. The present paper proposes an innovative scheme of combination between Federated Learning (FL) and LEO satellite-based networking to achieve the privacy-respecting and scaleable intelligence among distributed IoT nodes globally. The offered system takes advantage of edge level model training and aggregation of models via satellite, completion with such issues as satellite intermittent coverage, limited bandwidth, and security threats. Notable innovations are satellite-aware model compression, delay tolerance Aggregation Buffers, and secure gradient sharing. The architecture also lets edge devices do local updates, and the LEO satellites oversee fusion of the models during visibility windows. As shown with simulation experiments, the proposed method attains accuracies at an almost similar level as centralized schemes yield, and at the same time lowers the communication overhead by up to 60%. It is also very robust against packet loss and delays in satellite handover and hence is very appropriate in practical implementation in band-limited and latency-sensitive networks. Overall, the presented effort could bring a scalable and secure framework of FL to worldwide IoT connectivity with the help of LEO satellites. Future work will be done in enhancing factors of the satellite scheduling, cross-device FL diversity, and dynamic coordination through reinforced learning.