基于联邦学习的医疗数据共享与隐私保护

Medical data sharing and privacy protection based on federated learning

针对医疗数据共享中的“数据孤岛”问题,提出一个基于联邦学习的医疗数据共享与隐私保护方案。利用差分隐私技术对各个医疗机构的本地模型参数添加噪声,解决参数泄露的问题;全局模型通过编写的智能合约聚合并上传区块链,避免过度依赖中心聚合服务器;针对提出的隐私保护方案对差分隐私噪声的添加位置进行实验和讨论。实验结果表明,该方案可以在去中心化的同时保护数据的隐私,达到了较高的准确度,找到了合适的噪声添加位置。

To solve the data silos problem in medical data sharing,a medical data sharing and privacy protection scheme based on multi-layer blockchain and federated learning was proposed.The differential privacy technology was used to add noise to the local model parameters of each medical institution to solve the problem of parameter leakage.The global model aggregated and uploaded the blockchain by writing smart contracts to avoid over-reliance on the central aggregation server.The experiment and discussion on adding differential privacy noise in different positions were carried out.Experimental results show that the scheme can protect the privacy of data and achieve high privacy while decentralization,and the appropriate location for adding differential privacy noise is obtained.