부산대학교 도서관

Federated learning (FL) is a distributed approach that enables collaborative training of a shared machine learning (ML) model for a given task. FL requires bandwidth-demanding communication between devices and a central server, which is a cause of many issues such as communication bottlenecks and scaling in the network. Therefore, we introduce the CCS (Compressed Client Selection) algorithm aimed at decreasing the overall communication costs for fitting a model in the FL environment. CCS employs a biased client selection strategy that reduces the number of devices training the ML model and the number of rounds required to reach convergence. In addition, the compression method Count Sketch is implemented to reduce the overhead in client-to-server communication. A use case on the Human Activity Recognition dataset is performed to evaluate CCS and compare it with other state-of-the-art approaches. Experimental evaluations show that CCS efficiently reduces the overall communication overhead for fitting a model and its convergence in a FL environment. In particular, CCS reduces up to 90% the communication overhead compared to literature approaches while providing good convergence even in scenarios where the data are not-independently and identically distributed among client devices.