Chen, Bo; Zheng, Xiaoying; Zhu, Yongxin; Qiu, Meikang

SMART COMPUTING AND COMMUNICATION

2022, 13202, 0302-9743

Chen, Bo; Zheng, Xiaoying; Zhu, Yongxin; Qiu, Meikang

Federated learning is one computation paradigm used to address privacy preservation and efficient collaboration computing nowadays. Especially, in the environment where edge devices are facing different data scenarios, it is a challenge to enhance the prediction model accuracy. Since the data distributions on different edge devices might not be independent identical distributions, and also due to the communication obstacles existing in the modern complicated wireless world, it is an essential problem to sample which client devices to contribute to the server learning model. In this paper, instead of making the assumption on uniform distributed data sources, we assume the agnostic data distribution presumption. One indicator called client reward is defined applicable on the proposed client sampling algorithm. Combing with the redefined loss functions on the agnostic data distribution, a novel client sampling scheme is proposed and tested on real world datasets. The experiment results show that the client sampling scheme improves prediction accuracy on unbalanced data sources from different edge devices and achieves reasonable computing efficiency.