Abstract: To solve the problems of privacy leakage of intermediate parameters and uploading of low-quality models by malicious nodes during industrial big data sharing,a method of industrial big data sharing based on bee colony learning was proposed in combination with the hot ChainMaker.The learning structure of swarm was first constructed.Then the data was divided according to user dimension and user characteristic dimension.An industrial big data sharing algorithm based on horizontal swarm learning was proposed for data with low user dimension repetition and high feature dimension repetition,and an industrial big data sharing algorithm based on vertical swarm learning was proposed for data with high user dimension repetition and low feature dimension repetition.A scoring method was constructed to score the models trained by the nodes,and the models with higher score were selected to integrate.With many experiments,the proposed method was verified could not only prevent the privacy of intermediate parameters from leaking in the learning process,but also enhance the mutual trust between the nodes involved in the federated learning process,thus realizing a credible federated learning model and enhancing its privacy protection.

Key words: security protection, blockchain, swarm learning, federated learning

摘要： 为了解决工业大数据共享时,中间参数的隐私泄露以及恶意节点上传低质量模型的问题,结合目前较为火热的长安链,提出了基于蜂群学习的工业大数据共享方法。首先构建蜂群学习架构;其次将数据按照用户维度和用户特征维度进行划分;对用户维度重复度低而特征维度重复度高的数据提出了基于横向蜂群学习的工业大数据共享算法,对用户维度重复度高而特征维度重复度低的数据提出了基于纵向蜂群学习的工业大数据共享算法;最后构造了一个评分方法,将节点训练好的模型通过评分方法进行评分,服务器最终选择出评分较高的模型进行整合。通过多次实验后验证了所提方法不仅能够防止学习过程中泄露中间参数的隐私,还能增强参与联邦学习过程的节点之间的互相信任,从而实现一个可信的联邦学习模型,增强其隐私保护。

关键词: 隐私保护, 区块链, 蜂群学习, 联邦学习