Quality-aware Client Selection and Resource Optimization for Federated Learning in Computing Networks

Due to the challenges of traditional machine learning in terms of data privacy and transmission efficiency, an efficient and private distributed training framework, namely federated learning (FL), is emerged. In the FL training process, users only need to upload to the server, thus preserving user privacy data and improving transmission efficiency. The computing network can provide sufficient computing power support for federated learning training. However, FL still faces many difficulties, such as dynamic wireless channels, limited local computing resources, data heterogeneity, and malicious data attacks. To tackle these challenges, it is crucial to select reasonable clients to participate in training. In this paper, we propose a client selection strategy that considers data quality, computing capacity, and radio resources. We first define a data quality metric by measuring the heterogeneity and reliability of the local dataset. Based on this, we formulate a joint optimization problem of client selection and resource allocation to minimize the average time delay and power consumption while maximizing data quality. Considering the dynamic of wireless channels and computing frequency, an online learning algorithm based on multi-armed bandit (MAB) is developed to obtain the client selection. Finally, a large number of simulations are carried out to verify the effectiveness of the proposed algorithm. The evaluation in different scenarios shows that the DQ-UCB algorithm can discard the attacked clients and the clients with poor computing power or channel quality to achieve better performance.