Enhancing Federated Learning with Self-Determining Mechanism in MEC

Federated learning (FL) has gained substantial attention as a promising solution to the need for client privacy in mobile edge computing (MEC). However, FL suffers from instability of accuracy because of the invalid clients who become stragglers caused by frequent fluctuation of available resources in MEC. To tackle the challenge, most of the frameworks of asynchronous FL allow the parameter server (PS) to schedule clients reasonably. This kind of central server-determining paradigm makes it difficult to select all potentially useful clients because their resources in MEC vary frequently. In this paper, we proposed a new semi-asynchronous FL framework with a self-determining mechanism. This kind of framework fully exploits useful local models to improve global accuracy. Our proposed system has the following notable properties. Firstly, since clients perceive their resource status, thereby the self-determining clients can autonomously determine by themselves whether to participate in FL training according to the resource status. Secondly, compared the experimental results with other baselines, our proposed framework significantly improves the average global test accuracy.