Energy Efficient Federated Learning Joint Communication and Computation Framework Over Wireless Networks

Energy efficiency continues to be a challenge in federated learning. In this paper, we investigate the energy efficiency of computational and transmission resource allocation for federated learning over wireless networks. In order to minimize the total energy consumption, a given completion time and packet error rate constraint are considered. A closed-form solution is derived for the CPU frequency, bandwidth allocation, and transmission power using the Lagrange Multiplier Method. The effects of both local and global iterations on training accuracy and energy consumption are discussed. Compared with other baselines, the proposed joint communication and computation framework greatly improves the energy efficiency of federated learning, reducing energy consumption by 66.8% and 48.3%, respectively.