Blockchain-Enabled Federated Transfer Learning for Anomaly Detection of Power Lines

To mitigate privacy issues inherent in centralized machine learning paradigms utilized for anomaly detection, contemporary research has increasingly adopted federated learning (FL) frameworks. This approach decentralizes the computational training processes, transitioning them from centralized cloud servers to peripheral edge devices. Despite the advancements, conventional FL methodologies encounter substantive obstacles pertaining to the architecture safety, personalization, and transferability. Therefore, we propose a blockchain-enabled federated transfer learning (FTL) approach for power line anomaly detection. This approach combines transfer learning with FL, utilizing model discrepancy and maximum mean discrepancy to weigh local model parameters during global model aggregation and achieve domain adaptation. Our case study validates that the proposed FTL method outperforms conventional FTL and FL, resulting in a more favorable and robust model compared to local training. © 2024 IEEE.