HiveFL: GAN-Empowered Semi-Asynchronous Federated Learning With Self-Determining Clients

Abstract

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 this challenge, most of the frameworks of asynchronous FL allow the parameter server (PS) to schedule clients reasonably. This centralized decision-making paradigm makes it difficult to select all potential clients because client resources in MEC vary frequently. As another category of solutions, the semi-asynchronous FL frameworks follow the selection-after-training paradigm, which allows all potential clients to participate in FL but leads to a large waste of computation and communication resources. In this paper, we propose HiveFL, a new semi-asynchronous FL framework in which clients can proactively evaluate the changes in their resources, to improve global accuracy and reduce system overhead (i.e., the waste of resources induced by ineffective clients). HiveFL has the following notable properties. Firstly, HiveFL allows clients to perceive their resource status. Then, self-determining clients autonomously determine by themselves whether to participate in FL training according to the evaluation of their resource status. Secondly, comparing the experimental results with other baselines, HiveFL improves the average global test accuracy and the effective update ratio by 5.20%-22.21% and 20.3%-88.6%, respectively. Finally, HiveFL can reduce the average computation cost (measured by FLOPs) by 31.04%-81.45%. In addition, to address the problem brought by limited client resource status data, we adopt the time-series Generative Adversarial Networks (TimeGAN) method to provide more client data while training FL models. We prove the effectiveness of introducing the GAN-generated data in our experiments.