Olive Branch Learning: A Novel Federated Learning Framework for Space-Air-Ground Integrated Network

Space-air-ground integrated network (SAGIN) can provide global data collection to support various data-driven artificial intelligence (AI) services, especially in some remote areas where terrestrial communication facilities are unavailable (e.g., desert and isolated islands). However, the hierarchical data collection mode of SAGIN faces the challenges of redundant data collection, ineffective data transmission and privacy concerns. To this end, we first propose a novel configurable federated learning approach in SAGIN environment, namely Olive Branch Learning (OBL). Specifically, it consists of a ring structure LEO constellation in the space layer, and two-tier star structures imposed on the air nodes and IoRT devices in the air layer and the ground layer, respectively. Then, an efficient clustering and mapping algorithm is designed for OBL by taking into account the model bias of the air nodes as well as their geographic distributions. The objective is to minimize the time cost required to train the global model to achieve the preset accuracy. Extensive experimental results based on realistic learning datasets demonstrate the superior performance of our algorithm over the benchmark policies.