Customized Federated Learning for accelerated edge computing with heterogeneous task targets

摘要

As a dominant edge intelligence technique, Federated Learning (FL) can reduce the data transmission volume, shorten the communication latency and improve the collaboration efficiency among end-devices and edge servers. Existing works on FL-based edge computing only take device- and resource-heterogeneity into consideration under a fixed loss-minimization objective. As heterogeneous end-devices are usually assigned with various tasks with different target accuracies, task heterogeneity is also a significant issue and has not yet been investigated. To this end, we propose a Customized FL (CuFL) algorithm with an adaptive learning rate to tailor for heterogeneous accuracy requirements and to accelerate the local training process. We also present a fair global aggregation strategy for the edge server to minimize the variance of accuracy gaps among heterogeneous end-devices. We rigorously analyze the convergence property of the CuFL algorithm in theory. We also verify the feasibility and effectiveness of the CuFL algorithm in the vehicle classification task. Evaluation results demonstrate that our algorithm performs better in terms of the accuracy rate, training time, and fairness during aggregation than existing efforts.