1 Introduction
To alleviate the privacy issue of data owners and facilitate collaborative deep learning among distributed clients, federated learning (FL), which is one of decentralized machine learning approaches, has been introduced [1]. In FL for image classification, clients (e.g., mobile devices) are randomly selected as participants by an FL server. Then, the FL server multicasts a global model to these clients. Then, the clients update the model based on their local data. After finishing the model update, they upload their updated parameters to the FL server. Finally, the FL server aggregates the parameters from multiple clients and repeats these steps (a.k.a., round), until a desired model accuracy is achieved. However, when mobile devices are used as clients, random selection of participants leads to the performance degradation. For example, if some mobile devices (i.e., stragglers) with low computing power and poor wireless channel conditions are selected, they cannot return the updated parameters in a timely manner. That is, the FL server cannot complete a round within a short duration, which leads a long convergence time. In such a situation, more bandwidth should be allocated to the selected mobile devices with low computing power and/or poor wireless channel conditions.