基于区块链隐私增强的联邦学习可验证安全存储方案

Verifiable secure storage scheme for federated learning based onblockchain-privacy enhancement

在传统的联邦学习系统中,中央服务器在接收到参与方上传的本地模型后,没有对本地模型的合法性进行验证,可能导致后续模型不可靠即模型存储可靠性问题。针对该问题提出一种基于区块链隐私增强的联邦学习可验证安全存储方案(blockchain-based federated learning for model verified security storage, BFL4MVSS)。利用区块链技术,提出验证者投票机制,保证存储模型的可靠性;引入一种奖励机制,允许系统在一定数量的恶意节点存在下正常运行。采用PyTorch框架在MNIST数据集上进行仿真验证,结果表明,其能够在15%的恶意节点的参与下确保模型具有良好的表现。

In traditional federated learning systems,the central server does not verify the legitimacy of the local model after receiving it from the participants,which may lead to unreliable subsequent models,i.e.model storage reliability issues.A blockchain based federated learning for model verified security storage(BFL4MVSS)scheme based on blockchain privacy enhancement was proposed to address this issue.By utilizing blockchain technology,a validator voting mechanism was proposed to ensure the reliability of the storage model.A reward mechanism was introduced to allow the system to operate normally in the presence of a certain number of malicious nodes.The PyTorch framework was used for simulation verification on the MNIST dataset.The results show that the model can ensure good performance with the participation of 15%of malicious nodes.