Towards Instant Clustering Approach for Federated Learning Client Selection

In just few years, Federated Learning (FL) started to gain unprecedented attention given its ability to solve some fundamental privacy and communication challenges of traditional machine learning. Client selection is one of the main challenges in FL and is usually done in a random fashion, where the central server arbitrarily selects a certain number of clients to participate in each training round. However, given the heterogeneity of the client devices in terms of data quality and resource availability, randomly selecting clients is likely to result in long local training time and thus delayed global model's convergence. To address this problem, in this work, we propose a real-time and on-demand client selection mechanism that employs the DBSCAN (Density-Based Spatial clustering of Applications with Noise) clustering technique from machine learning to group the clients into a set of homogeneous clusters based on a set of criteria defined by the FL task owners, such as resource availability, data quality, data size, data freshness and non-IID degree. Based on the requirements of each FL task, the server then intelligently selects the clusters of clients that best match with each task's requirements, thus improving the performance of the overall federated learning process. Experiments suggest that our solution significantly improves the accuracy of FL compared to the Vanilla FL approach.