FLARE: Availability Awareness for Resource-Efficient Federated Learning

As privacy concerns increase, machine learning (ML) is undergoing a paradigm shift whereby training occurs at the end users rather than centrally in data centers, in a process called federated learning (FL). As training becomes distributed, challenges such as data heterogeneity and device availability can affect the performance of FL. Focusing on device selection, existing FL schemes either use random or full selection to ensure fairness and model accuracy or smart selection for resource efficiency. Smart selection often assumes complete knowledge of client availability. However, in reality, there can be discrepancies between the actual device availabilities and the advertised ones, leading to availability faults. We conjecture that such faults will impact of the performance of FL. In this paper, we propose FLARE (Federated Learning for Availability and Resource Efficiency), a novel fault-tolerant framework for FL that tolerates availability faults. In FLARE, FL occurs at the end devices while an availability model is built centrally for smart client selection. We demonstrate the viability of FLARE by comparing it against another state of the art technique. We show that (i) the client selection technique of FLARE outperforms current ones, enabling better resource usage efficiency, (ii) the convergence time of FLARE can be bounded and (iii) the model accuracy of FLARE is comparable to the case when full availability data of client is sent to the server.