Abstract: Traditional federated learning relies on a central server, and the training process is vulnerable to single point of failure and malicious attacks from nodes, and intermediate parameters passed in plaintext may be exploited to infer the private information in data. A decentralized, secure, and fair federated learning model based on the blockchain is proposed, using homomorphic encryption technology to protect the privacy of the intermediate parameters of the collaborative training parties. Model aggregation and collaborative decryption are carried out through the elected federated learning committee. The decryption process achieves secure key management through a secret sharing scheme, using bilinear-map accumulator to provide verification of correctness for the secret share. The model also introduces reputation as an indicator to evaluate the reliability of the participants, and uses the subjective logic model to realize disbelief enhanced reputation calculation as the basis for the election of the federated learning committee. The reputation value can be used as a reference for the incentive mechanism to ensure fairness. Model information and the reputation value realize data tamper-proof and non-repudiation through the blockchain. Experiments show that in the condition of the training, accuracy is slightly lower than that of the centralized learning model, and model can guarantee that it can be trained in a decentralized manner in multi-party collaborative environment, and implement data privacy protection for all participants.