Adaptive Personalized Federated Learning With One-Shot Screening

The rapid development of Internet of Things (IoT) offers unprecedented opportunity for federated learning (FL). However, the increasing scale of IoT is accompanied by limitation of communication resource and endogenous heterogeneity of edge devices, which hinder FL’s popularization. Based on features of the edge network, this article proposes a novel one-shot screening scheme to identify edge users with homogeneous data distribution and orchestrate adaptive personalized FL tailored for any task initiator. To cope with the stipulation that raw data cannot be communicated, the learning loss calculated with a locally pretrained model is leveraged to measure the similarity between data distributions. In addition, a multiple hypothesis testing instrument, the Adaptive
$p$
-value Thresholding (AdaPT) is utilized to automatically control false discovery rate (FDR) of the screening below any specific level. To demonstrate the effectiveness of the proposed collaborator selection scheme, a numerical study is designed on the real-world non-i.i.d. data set FEMNIST. Evaluations confirm that stable improvement in convergence rate and learning accuracy can be gained compared to vanilla global FL, iterative personalized reweighted FL as well as iterative clustered FL, along with additional welfare of economical wireless resource occupation. Finally, the experiment is carried out under differential privacy to verify robustness further.