Energy-Efficient Hierarchical Federated Learning in Massive Wireless IoT Networks
Renan Rodrigues Oliveira, Kléber V. Cardoso, Antônio Oliveira-Jr · 2025
Federated Learning (FL) enables collaborative model training without sharing raw data, preserving privacy and reducing communication overhead. On the other hand, in Internet of Things (IoT) wireless networks, FL faces issues such as limited resources, unreliable communication channels, and large delays. Hierarchical Federated Learning (HFL) addresses these issues using a tree topology with intermediate servers to reduce communication distances, improve aggregation efficiency, and mitigate transmission failures. However, current algorithms are not well-suited to address the scalability challenges posed by the massive scale of beyond-5G networks. In this context, we propose a novel HFL algorithm called HFLwOpt, which dynamically optimizes communication and computation resources in massive wireless IoT networks, maximizing successful transmissions, minimizing energy use, and reducing training latency. Our simulation with over 1, 000 devices, utilizing three levels of aggregation, demonstrates that HFLwOpt surpasses baselines with fixed resource allocations. The results reveal a reduction of up to 45.34% in energy efficiency and 77.01% in training latency for the MNIST-based dataset, and 42.23% in energy efficiency and 76.77% in training latency for the FMNIST-based dataset.