一种基于区块链和梯度压缩的去中心化联邦学习模型

A Decentralized Federated Learning Model Based on Blockchain and Gradient Compression

联邦学习可在保护数据隐私的前提下完成模型的训练,但实际应用中存在的安全问题阻碍了联邦学习的发展。提出一种基于区块链和梯度压缩的去中心化联邦学习模型。首先,利用区块链存储训练数据,训练参与方通过全局模型本地更新的方式取代中心服务器并使用智能合约实现对链上数据的访问控制。其次,提出一种梯度压缩方法,对模型参数进行压缩以减少参与方与区块链之间的数据传输量且有效防止了梯度隐私泄露。最后,为减弱梯度压缩对全局模型收敛速度的影响,使用热身训练的方式提升全局模型的收敛速度以缩短整体训练时间。实验结果表明,该模型在减少传输数据量的情况下对全局模型准确率有较小影响且提升了联邦学习训练效率。

Federated learning could complete the training of models while protecting data privacy,but security issues in practical applications hindered the development of federated learning.A decentralized federation learning model based on blockchain and gradient compression was proposed.Firstly,a blockchain was used to store training data,and the training participants replaced the central server by local updates of the global model and used smart contracts to achieve access control to the data on the chain.Secondly,a gradient compression method was proposed to compress the model parameters to reduce the amount of data transmission between the participants and the blockchain and effectively to prevent the gradient privacy leakage.Finally,to reduce the impact of gradient compression on the convergence speed of the global model,a warm-up training method was used to improve the convergence speed of the global model to shorten the overall training time.The experimental results showed that the model had a small impact on the global model accuracy and improved the federal learning training efficiency with the reduced amount of transmitted data.