Joint Heterogeneity-Aware Personalized Federated Search for Energy Efficient Battery-Powered Edge Computing.

The limited battery capacity of edge devices has a significant impact on the deployment of Federated Learning (FL). As a result, maintaining computation sustainability is a critical issue for edge FL. Furthermore, the heterogeneities of deployed edge devices reduce FL efficiency and effectiveness, making FL computation sustainability more challenging to maintain. To address these issues raised by heterogeneities, we perform a joint heterogeneity-aware personalized federated search for energy-efficient edge computing. To achieve energy-efficient on-device inference and training, a one-training process is adopted to search for personalized partial network structures on each device. We begin by tailoring the network scale on each node based on the efficiency of model inference, which also serves as the search space for optimization. This strategy can mitigate the straggler problem and improve the energy efficiency of FL by guiding the efficient FL training process in each round. To further optimize the energy consumption of edge devices, we design a lightweight search controller during the search process. This controller meets the low energy consumption requirements of the edge devices and reduces their energy consumption during the search process. Finally, we introduce an adaptive search strategy to guarantee personalized training convergence on each device. By reducing the energy consumption of each training round and ensuring the training convergence of personalized models, we can significantly improve the energy efficiency of FL on battery-powered edge devices. Our framework can obtain competitive accuracy with state-of-the-art methods while improving inference efficiency by up to 1.43× and training energy efficiency by up to 2.63×.