【摘要】 在联邦学习作为隐私保护技术被广泛应用的同时，也产生了中心服务器不稳定和联邦学习服务器与参与方交互造成的隐私泄露等新的挑战及安全问题.提出了一种基于区块链和动态评估的隐私保护联邦学习模型，利用区块链解决中心服务器的问题，通过本地训练使用稀疏化、全局模型更新使用差分隐私解决联邦学习过程中的隐私泄露问题，本地训练完成后用数字签名和双重Hash对比验证参与方身份和训练模型的所属权.此外，使用多权重动态评估方法计算单轮模型和参与方评估值作为参与方贡献的依据.实验结果表明，提出的模型可以有效解决联邦学习中的单点故障和局部模型验证问题，与传统联邦学习相比，使用稀疏化和差分隐私可以在略微损失准确率的情况下保障模型的安全性，并有效地为参与方进行评估，从而保证了激励机制的公平性.更多还原

【Abstract】 While federated learning is widely used as a privacy-preserving technology, new challenges such as instability of the central server and privacy leakage caused by the interaction between the federated learning server and the participants as well as security issues have arisen. A privacy-preserving federated learning model based on blockchain and dynamic evaluation is proposed. Local training is performed using sparsification, and global model updates are protected using differential privacy to solve the privacy leakage problem in the federated learning process,using digital signature and double Hash comparisons to verify the identity of the participants and the ownership of the trained model after the completion of local training. In addition, multiple weight dynamic evaluation method is used to calculate the model’s single-round evaluation and the final evaluation of the participants. Experiments show that the proposed model can effectively solve the single-point failure and local model verification problems in federation learning, and the use of sparsification and differential privacy can guarantee the security of the model with a slight loss of accuracy compared with traditional federation learning. The evaluation of the local model and the participants are scored as the basis for the participants’ contributions, thus effectively ensuring the fairness of the incentive mechanism.更多还原