一种处理隐私保护数据的神经网络  

Neural Network for Processing Privacy-protected Data

作　　者：王启正 高玲[1] WANG Qi-Zheng;GAO Ling(College of Information Science and Engineering,Shandong Normal University,Jinan 250100,China)

机构地区：[1]山东师范大学信息科学与工程学院,济南250100

出　　处：《密码学报》2019年第2期258-268,共11页Journal of Cryptologic Research

基　　金：国家自然科学基金(61602283;61672329;61572298;61772322;61601268);山东省自然科学基金(ZR2016FB10)~~

摘　　要：神经网络是一种重要的数据分类工具,一个经过良好训练的神经网络可以高效准确的完成对输入数据的分类和预测,在信息处理、模式识别等方面有着广泛的应用.目前对神经网络算法的研究,并在基础上对神经网络的训练,已成为产业界和学术界关注的热点,训练良好的神经网络也逐渐成为诸多数据处理企业的重要知识产权.因此,当普通用户希望借助神经网络进行数据分类时,往往需要将数据托管给专业机构,这时如何保护数据隐私就成为最为重要的问题.此外,目前的神经网络算法计算量较大,无法很好地在计算资源受限的设备上运行.引入云计算服务提供了一种解决方案,可以将复杂的神经网络计算任务委托给云,但同样存在着隐私泄露的风险.为了解决以上问题,本文设计了一种基于同态加密的隐私保护神经网络,使用了同态加密算法对数据进行加密,并借助同态加、乘的性质对神经网络的运算过程进行重写,在保障数据隐私性的前提下保留了数据的可计算性.与以往的隐私保护神经网络相比,可以更好的适用于复杂的神经网络,具有更高的安全性.Neural network is an important data classification tool.A well-trained neural network can efficiently and accurately classify and predict input data.It has extensive applications in information processing and pattern recognition.At present,the study of neural network algorithms and the training of neural networks based on them have become the focus of attention in the industry and academia.Well-trained neural networks have gradually become important intellectual property rights for many data processing companies.Therefore,when ordinary users want to use neural networks for data classification,they often need to host data to professional organizations.How to protect data privacy becomes an important issue.In addition,the current neural network algorithm needs a large amount of computation and cannot run well on devices with limited computing resources.The introduction of cloud computing services provides a solution that can delegate complex neural network computing tasks to the cloud,but there is also the risk of privacy leakage.In order to solve the above problems,a privacy protection neural network based on homomorphic encryption is designed in this paper.The homomorphic encryption algorithm is used to encrypt the data,and the operation process of the neural network is rewritten by means of the properties of homomorphic addition and multiplication.While preserving the data privacy,the designed neural network preserves the computability of the data.Compared with the previous privacy protection neural networks,the proposed one can be applied to complex neural networks with higher security.

关 键 词：神经网络 隐私保护 同态加密 茫然传输 云计算 

分 类 号：TP309[自动化与计算机技术—计算机系统结构] TP183[自动化与计算机技术—计算机科学与技术]