Bandwidth Allocation for Low-Latency Wireless Federated Learning: An Evolutionary Game Approach

As a new distributed data training framework, federated learning (FL) has attracted increasing attention owing to its advantages of preserving data privacy and low communication cost. However, in a wireless FL network, due to the limited bandwidth resources, when a large number of clients participate in FL, the communication burden is too heavy. Therefore, efficient bandwidth allocation is critical to facilitate the application of wireless FL. In this paper, we investigate bandwidth allocation in wireless FL networks with the objective of minimizing the latency of FL services. We first analyze and formulate the latency minimization problem. Then, considering that the computing and transmitting capabilities of each client cannot be completely and truly acquired, we develop an evolutionary game (EG) framework to model the dynamic process of bandwidth allocation in wireless FL. We further show the optimal bandwidth allocation solution is equivalent to the evolutionary equilibrium (EE) obtained by the replicator dynamics in the EG model, and prove the EE is asymptotically stable. With the help of these results, we propose the EG-based bandwidth allocation algorithm, which enables the latency of FL services to be reduced by performing the replicator dynamics iteratively. Numerical simulations are provided to demonstrate the evolutionary behaviors in the EG-based bandwidth allocation Algorithm.

I. Introduction

The industrialization of 6G and the Internet of Everything (ioE) will promote the wide deployment of a variety of intelligent devices in human society [1]. Massive mobile devices generate unprecedented data every day, and the data can be trained using artificial intelligence (AI) technologies to realize intelligent IoE applications [2], such as smart homes, intelligent healthcare, smart cities, intelligent transportation, industrial automation, etc. With the development of mobile edge computing (MEC) and the improvement of clients' awareness of privacy protection, federated learning (FL) has been proposed as a new machine learning (ML) paradigm [3]. FL enables a group of clients to train the ML model cooperatively under the coordination of the edge server while keeping all training data on the device. As a result, the ML model is trained and the clients' private data is protected at the same time [4].