Blockchain Governance and Adaptive Incentive Mechanisms in Federated Learning

To address issues in federated learning, such as centralization, lack of contribution evaluation, and the absence of effective incentive mechanisms, this paper proposes a blockchain-based governance and incentive mechanism for model-sharing communities. By leveraging the decentralized nature of blockchain, a federated ecosystem is established, defining a tri-party role system comprising algorithm providers, data contributors, and model consumers. A reputation system and credit-based rules are introduced to enable collaborative autonomy among participants. Furthermore, a lightweight approximation algorithm, FL-Light Shapley, is designed to incorporate approximate Shapley value computation into federated learning, providing a scalable and practical solution for contribution evaluation in large-scale settings. Finally, an evolutionary game-theoretic incentive mechanism is proposed to balance the marginal utility decay during model convergence, effectively aligning the interests of both supply and demand sides. This mechanism overcomes the limitations of static incentive models by enabling strategy adaptation, thereby enhancing system resilience and collaborative efficiency. Experimental results demonstrate that the proposed mechanism effectively rewards or penalizes participants under varying data quality, fostering long-term sustainable community development.