Resource Optimization and Device Scheduling for Flexible Federated Edge Learning with Tradeoff Between Energy Consumption and Model Performance

Nowadays, users are becoming more reserved in uploading their own data to the servers of service providers for fear of personal information disclosure. In order to meet the need on privacy security, Federated Learning (FL) is proposed. As the nearest servers to users, edge servers are quite suitable for the execution of FL tasks, which results in a new concept called Federated Edge Learning (FEEL). It allows users to conduct model training locally without uploading their own data, thus avoiding privacy disclosure. However, the issue of Energy Consumption (EC) for training on devices becomes a main concern for users because the learning process is carried out on the devices with limited battery capacity. Then, from the standpoint of service providers, they pay more attention to the model performance (i.e., test dataset accuracy of the trained machine learning model) and usually have diverse requirements on it. Both EC and model performance are important metrics for FEEL, and from our engineering experience, service providers and users may have a conflict of interest in FEEL. In this paper, after modeling the two metrics, we identify the common factor (i.e., size of the training data) between them and are convinced that there is a tradeoff. Then, we add a workload constraint that regulates the common factor to the formulated problem and propose a resource optimization and device scheduling strategy to solve it, thus achieving the tradeoff between EC and model performance of FEEL. This strategy is based on the steepest descent method and approximation algorithm. The approximation ratio is also proved. By regulating the training workload threshold, the values of the two metrics can be dynamically adjusted. Achieving the tradeoff makes it possible to meet the needs of service providers and users at the same time. More specifically, it can minimize the energy consumption of training devices on the premise of meeting diverse requirements of service providers on model performance, which can not be realized by other FEEL frameworks that do not achieve the tradeoff. Simulation results show that our proposed strategy is able to achieve this tradeoff compared with two existing FEEL frameworks. Compared with another existing FEEL framework who also realizes the tradeoff, the tradeoff achieved by our proposed strategy is more biased towards EC. At the end of the simulation part, we summary the characteristics of the proposed and other three existing FEEL frameworks.