An Incentive Mechanism for Consortium Blockchain-based Cross-Silo Federated Learning

Cross-silo federated learning is a new type of machine learning, where organizations, as model owners, collaborating to train global models with local data. Nevertheless, the heterogeneity caused by differences in training costs within the organization leads to different views on training rounds among different organizations. At the same time, the organizations’ communication resources are non-excludable public goods, resulting in free-rider problems. To address the aforementioned concerns, we propose a cross-silo federated learning incentive mechanism based on a consortium blockchain. The consortium blockchain coordinates the training rounds and currency transfers, as well as determines the processing capabilities of each organization’s local training, and helps organizations record the training models during the training process. The mechanism models the game between organizations as non-cooperative games. We conducted a theoretical analysis of the mechanism using the MNIST dataset and demonstrated the existence of Nash equilibrium. A distributed algorithm has been proposed to maximize social welfare for organizations under information asymmetry conditions. The simulation results show that the algorithm has fast convergence.