FEDGE: FEDerated learning at the EDGE on space platforms using deep neural network architectures

Abstract We introduce FEDGE: FEDerated Learning at the EDGE, a framework designed for efficient AI deployment in resource-constrained satellite constellations. FEDGE integrates federated learning with edge computing to address communication overhead and latency challenges in distributed space environments. The framework features a novel edge-enhanced ground station protocol that dynamically schedules model aggregation based on satellite-provided metadata, combined with local stochastic gradient descent training at satellite edge devices and gradient compression via quantization. Experimental validation on MNIST and EuroSAT datasets demonstrates the practical viability of the approach. On MNIST, FEDGE achieved 94.33% training accuracy with 0.21 loss and 90.05% test accuracy with 0.24 loss. On EuroSAT, the framework reached 93.47% training accuracy with 0.18 loss and 91.51% test accuracy with 0.21 loss. Gradient quantization reduces data exchange by up to 14 $$\times $$ with approximately 4% impact on test loss. These results validate FEDGE as a communication-efficient solution for decentralized AI deployment in satellite systems, enabling autonomous spacecraft intelligence and addressing the unique constraints of space-based computing platforms.