一种数据隐私保护下多方无损线性模型学习方法

A Multi-Lossless Linear Model Learning Method with Data Privacy-Preserving

随着工业大数据技术的发展,制造企业通过收集和分析生产数据,获取在预测、诊断等方面的优化方法。然而制造企业受限于建模和算力等技术瓶颈,难以高效地实现数据分析,在联合其他参与方共同协作时即需要承担信息泄露的风险,又难以保证模型的性能无损。针对这些场景问题,提出基于数据隐私保护的多方无损线性模型学习方法。首先搭建多方协作计算框架,设计了数据单向加密算法来保护数据出场的隐私安全,各协作方分别基于加密数据进行线性模型的训练。随后,研究分析线性模型与数据集的关联特性,提出线性模型无损聚合算法。最后,在典型工业场景数据集上进行方法验证,试验结果表明提出的框架可以获得性能无损的全局模型,并实现数据持有方的隐私安全保护。

With the development of industrial big data technologies,manufacturing enterprises collect and analyse production data to obtain optimization methods of forecasting and diagnosis.However,manufacturing companies are constrained by technical bottlenecks such as modelling and computing power,which make it difficult to analyse data efficiently.Manufacturing enterprises need to bear the risk of information leakage,and it is also difficult to guarantee that model performance is lossless,when cooperating with other participants.For these scenario problems,a multi-lossless linear model learning method based on data privacy preserving is proposed.Firstly,a multi-party collaborative computing framework is built and the one-way data encryption algorithm is designed to protect the privacy of the data.Each collaborator trains the linear model separately based on the encrypted data.Secondly,the study analyses the association properties of the linear model with the dataset,proposing a lossless aggregation algorithm for linear models.Finally,the method is validated on the typical industrial scenario dataset.The experimental results show that the proposed framework can obtain global models with lossless performance and achieve privacy security for the data holders.