Federated Learning (FL) enables collaborative model training without centralizing data, yet deployment over low-power wide-area networks (LPWANs), such as LoRaWAN, remains challenging due to severe bandwidth, payload, and duty-cycle constraints. This work examines the feasibility of FL on LoRaWAN by designing and implementing a communication-efficient framework tailored to constrained Internet of Things (IoT) environments. CE-FedAvg is introduced as a communication-efficient variant of FedAvg that restructures client–server exchanges to comply with standard LoRaWAN operation. In addition to sparsification, quantization, and lightweight compression, CE-FedAvg incorporates seed-based initialization for deterministic synchronization and explores model improvements optimized for ultra-low-throughput communication environments. The framework is fully compatible with the LoRaWAN specification and is implemented across heterogeneous edge devices and a low-power gateway, demonstrating end-to-end feasibility under realistic IoT conditions.

Top