An Active Client Selection Scheme Based on Blockchain for Federated Learning in Shipping

Federated Learning (FL) enables collaborative model training across maritime devices without the need to share raw data. However, challenges such as data heterogeneity and unreliable marine communications impede its performance and security. In this work, we propose a Blockchain-based Active Client Selection Strategy for FL in Shipping (BAFLS), which utilizes blockchain technology to create a secure and auditable environment for node registration and parameter exchange. A lightweight consensus algorithm is introduced to dynamically elect aggregation nodes based on residual energy, reputation, and computing power, improving fault tolerance and reducing resource consumption. Based on such, a Top-k active learning strategy is designed to select the most informative clients, balancing data utility and privacy protection. Security evaluation and analysis demonstrate that BAFLS effectively resists aggregation attacks and privacy inference. Experimentations on FMNIST, HAR, and ShipNetwork10 datasets show that BAFLS achieves up to 2.4% higher accuracy, reduces convergence rounds by up to 44%, and consistently lowers communication overhead compared to the baseline under various degrees of label and feature heterogeneity.