FedEXT: Differential Federated Learning with Complementary Extension of Edge Models

We introduce FedEXT, an innovative federated learning framework designed to address the challenges of platforms with limited communication and computational resources. FedEXT employs a unique factorization-based methodology complemented by a paired neural network extension to improve model transmission and local training efficiency. This approach enables FedEXT to achieve the accuracy of larger models, which would otherwise be infeasible for federated learning on edge devices with stringent communication budgets. A key component of FedEXT is its periodic scheduler, which works in conjunction with local and global controllers to effectively manage model transmission. This scheduler uses a distinct learning and aggregation strategy to update the model parameters on local edge devices and the cloud server. The effectiveness of the framework is validated in both dataset-driven and real-world scenarios, demonstrating significant efficacy. FedEXT achieves improvements in accuracy ranging from 16% to 22% and a reduction in energy consumption by 29% to 34% compared to several state-of-the-art methods.