Federated learning system eliminating model drift in distributed edge computing: Theoretical analytics and application on pit engineering state monitoring

Abstract

In internet of things (IoT) with smart terminals, data is generated at edges in stream. Edge computing, where AI models are conducted on smart terminals, caters to this trend and is increasingly adopted. On levels of IoT, it places high demands on real-time performance and generalization of deployed models. Federated learning (FL) promises to meet these demands because of distributed paradigm. However, two challenges exist: (1) for incremental data on edges, mini batch and central FL training causes higher risks of data leakage; (2) drifts exist between global and local models while more difficult to reconcile in distributed edge computing. To solve challenges mentioned, measures are necessary to protect and coordinate FL process orient edge incremental data in IoT. In our study, we propose a blockchain-assisted federated learning system named BFLIoT. BFLIoT features flat, drift-resistant and safe: (1) BFLIoT decentralizes FL by blockchain and lets FL process operate fully based on edges without cloud participation; (2) BFLIoT utilizes improved federalized proximal (FedProx) algorithm and adjusts training parameters ($\lambda$ and $\gamma$) to fully eliminate drift during aggregation; and (3) privacy budget ($\delta$) for personalized differential privacy (DP) based on $\lambda$, $\gamma$ is employed considering both data protection and model aggregation. We verified BFLIoT on open datasets and applied BFLIoT to pit engineering edge monitoring system (PEEMS) we developed. FL task of predictive algorithm orient monitoring items verifies BFLIoT is effective in solving two problems presented. BFLIoT provides effective and safe strategies maintaining and optimizing AI models orient distributed edge computing.