Performance Optimization of Federated Learning With Blockchain-based Pipelining

This study addresses inefficiencies in blockchain-based federated learning caused by computational overhead and block generation delays. Therefore, a pipelining architecture that decouples model training, transaction transmission, and block generation into independent threads is proposed in this work. This structure enables the parallel execution of computation and communication phases, minimizing GPU idle time. Experimental results reveal that the proposed approach reduces the block generation cycle time, approximately, by 26% on the SVHN dataset and by 41% on the CIFAR-10 dataset compared to the blockchain-based federated learning as the baseline, while maintaining model accuracy and stability. By effectively resolving structural bottlenecks, this method significantly enhances the scalability and practicality of the system.