Multi-Feature Based Client Selection and Feature Weight Update for Volatile Federated Learning

This paper investigates a novel client selection for the volatile Federated Learning (FL) systems, where volatility means that the state of the client set, client datasets, and client training status will change over time. We study how to select clients dynamically to mitigate the volatility. Particularly, the volatile client selection problem is formulated as a classification problem, and we propose two new metric features. The Multi-Feature Volatile Client Selection (MFVCS) algorithm, which considers client training capacity, client-weighted data quality, and client historical selection entropy, is proposed to solve the volatile client selection problem. Moreover, we have developed an adaptive dynamic weighting algorithm that allows for dynamic updating of the weight for each feature. We propose a volatility ratio to measure client volatility. The experimental results indicate that the proposed algorithm demonstrates strong robustness and better performance under different volatility ratios of the client set. In particular, the proposed MFVCS algorithm improves the model accuracy at most by $\mathbf{9.2\%}, \mathbf{9.6\%}$ and $\mathbf{12.5\%}$ under 0.01 volatility ratio, 0.05 volatility ratio and 0.1 volatility ratio, respectively.