Loss and Energy Tradeoff in Multi-access Edge Computing Enabled Federated Learning

Abstract

Federated learning (FL) encourages users to train statistical models on their local devices. Since mobile devices have the limited power and computing capabilities, the users are rational in minimizing their energy consumption with the cost of the model’s accuracy. Multi-access Edge Computing (MEC) enabled FL is a prominent approach where users can offload a fraction of their dataset to the MEC server where the training of the statistical model is performed with the help of the powerful MEC server in parallel with the local training at the mobile users. With the size of dataset offloaded to the MEC server, both the performance of the model and the energy consumption of the system are varied. We analyze this tradeoff between the performance of the system and the energy consumption at the MEC server and mobile users. The time consumption can also be saved by managing the size of the dataset offloaded to the MEC server. Since the MEC server and mobile users have the conflicting interest in saving the energy consumption with the constraint on the time taken for one computing round where the performance of the model fluctuates across the size of offloaded dataset, we analyze the tradeoff by formulating the resource management problem as a penalized convex optimization problem. We propose a distributed resource management problem for MEC enabled FL system where the global model is responsible for radio resource management and each local model performs a dataset offloading decision. Then, we perform the simulation to show the tradeoff and performance of the proposed algorithm.