Optimizing Federated Learning Efficiency via Slimmable Neural Networks and AirComp for Resource-Heterogeneous Devices

Abstract

With the rapid growth of the Internet of Things (IoT) and edge intelligence, federated learning (FL) has become a promising approach for distributed model training while preserving privacy. However, the resource constraints of IoT devices, including limited communication bandwidth and heterogeneous computational and energy capabilities, hinder the scalability and performance of FL. To address these challenges, we propose the Air-FedSNN framework, which integrates slimmable neural networks (SNNs) with over-the-air computation (AirComp) to improve communication, computation, and energy efficiency in heterogeneous IoT networks. A width-aware aggregation mechanism is developed to mitigate the impact of varying model widths and aggregation errors. Additionally, a dynamic width allocation strategy based on Lyapunov optimization and multi-armed bandit (MAB) theory is introduced to adaptively select model widths under long-term energy constraints, balancing energy efficiency and training performance without prior knowledge of device capabilities. Extensive experiments on MNIST, CIFAR-10 and CIFAR-100 datasets demonstrate that Air-FedSNN outperforms conventional FL methods, achieving superior accuracy-energy tradeoffs, and provides an efficient solution for resource-constrained environments.