Data Heterogeneity-Robust Federated Learning via Group Client Selection in Industrial IoT

Nowadays, the Industrial Internet of Things (IIoT) has played an integral role in Industry 4.0 and produced massive amounts data for industrial intelligence. These locate on decentralized devices modern factories. To protect confidentiality data, federated learning (FL) was introduced to collaboratively train shared machine (ML) models. However, local collected by different skew class distribution degrade FL performance. This challenge been widely studied at mobile edge, but they ignored rapidly changing streaming clustering nature factory devices, more seriously, may threaten security. In this article, we propose FED GS, which is a hierarchical cloud-edge-end framework 5G empowered industries, improve performance non-independent identically distributed (non-j) data. Taking advantage naturally clustered GS uses gradient-based binary permutation algorithm (GBP-CS) select subset within each build homogeneous super nodes participating training. Then, compound-step synchronization protocol coordinate training process among these nodes, shows great robustness against heterogeneity. The proposed methods are time-efficient can adapt dynamic environments, without exposing confidential risky manipulation. We prove that better convergence than FedAvg give relaxed condition under communication efficient. extensive experiments show improves accuracy 3.5% reduces rounds 59% average, confirming its superior effectiveness efficiency non-i.i.d.