Distributed edge sensing systems, such as IoT monitoring nodes, wearable devices, and camera-based sensing terminals, continuously generate privacy-sensitive data that are costly to transmit to a central server. Federated learning (FL) provides a promising solution for collaborative model training without raw-data sharing; however, its practical deployment in edge sensing systems is challenged by non-IID local observations, limited uplink/downlink resources, and restricted on-device computation. To address these issues, this paper proposes a Dual-Sided Sparse Aggregation (DSSA) mechanism integrated with FedProx for resource-constrained edge sensing environments. In the proposed framework, the server prunes the global model after each communication round and transmits only the retained parameters, while clients update the complementary parameters and upload sparse local gradients. This fixed-structure sparse training strategy reduces bidirectional communication overhead and local computation cost, while FedProx improves robustness under heterogeneous data distributions. Experiments on CIFAR-10 and SVHN with varying non-IID degrees, pruning ratios, and hyperparameter settings show that the proposed method achieves a favorable resource-performance trade-off, reducing communication cost by up to 73.0% and computation cost by up to 34.9% while maintaining competitive accuracy. Under controlled benchmark settings, the proposed method demonstrates substantial resource savings compared with FedAvg, particularly in mildly heterogeneous scenarios, indicating a favorable benchmark-level resource-performance trade-off for resource-constrained edge sensing scenarios under the evaluated settings.
He et al. (Tue,) studied this question.