Towards Energy-efficient Federated Edge Intelligence for IoT Networks
Qu Wang, Yong Xiao, Huixiang Zhu, Zijian Sun, Yingyu Li, Xiaohu Ge
Abstract
Federated edge intelligence (FEI) is an emerging framework that implements federated learning (FL)-based learning solutions in an edge networking and computing system. It has attracted significant interest due to its potential to enable machine learning (ML)-based smart services and applications in the next-generation wireless systems. Despite its potential, the environmental impact of implementing energy-consuming ML-based solutions in a large networking system has been considered as one of the major challenges for the sustainability of future digital networking systems. This paper proposes energy-efficient FEI (EE-FEI), an energy-efficient framework, that jointly optimizes multiple key parameters to minimize the overall energy consumption of an FEI-supported Internet of Things (IoT) network. We establish models to quantify the relationship between the total energy consumption of FEI and the key parameters including the number of edge servers, the number of local model training rounds, and the number of global coordination rounds. We formulate the energy consumption minimization problem and prove its approximation problem is biconvex. Alternate Convex Search (ACS) algorithm for solving the key parameters to minimize the energy consumption of an FEI system has been used. Finally, we evaluate our theoretical results using a hardware prototype. Numerical results have shown that EE-FEI can significantly reduce the energy consumption of FEI systems by 49.8%.