부산대학교 도서관

In Federated Learning (FL), to improve the training efficiency, we don’t need to let all of the clients join in the training process. Instead, we can select some specific clients to join in the training. In particular, if some of these selected clients become problematic due to various reasons (e.g. shortage of power, poor internet connection, or being vulnerable to attacks) and thus could not successfully complete the training process, then we can discard those clients during training, in order to improve the efficiency. However, discarding those clients could increase the data source’s bias, because the data categories that contain those clients’ data would be underrepresented during the training process. To solve this problem, in this paper, we propose a robust client selection and replacement approach called RCSR. Using RCSR, we first cluster all clients according to their data distribution, and then use normal clients in the same cluster (with similar data distributions) to replace those problematic clients during training. We apply our methods to a couple of application scenarios in edge computing, and our results show that our methods can save training costs without affecting the accuracy.