面向深度神经网络的安全计算协议设计方法  

Design method of secure computing protocol for deep neural network

作　　者：毕仁万 陈前昕 熊金波 刘西蒙 BI Renwan;CHEN Qianxin;XIONG Jinbo;LIU Ximeng(College of Mathematics and Informatics,Fujian Normal University,Fuzhou 350117,China;College of Mathematics and computer science,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福建师范大学数学与信息学院,福建福州350117 [2]福州大学数学与计算机科学学院,福建福州350108

出　　处：《网络与信息安全学报》2020年第4期130-139,共10页Chinese Journal of Network and Information Security

基　　金：国家自然科学基金(61872088,U1804263,61702105,61872090);福建省自然科学基金资助项目(2019J01276);贵州省公共大数据重点实验室开放课题(2019BDKFJJ004)。

摘　　要：针对深度神经网络模型计算过程中存在的信息泄露问题,结合加性秘密共享方案,在两台非共谋的边缘服务器间设计安全高效的交互计算协议。考虑到非线性函数不能直接拆分,首先提出一组基本转换协议,实现加性副本和乘性副本的安全转换,经过少量调用,可以安全计算幂函数、比较、指数、对数、除法等底层函数。由于数据传递和计算特点,协议可以扩展至数组计算。理论分析证明了协议的正确性、高效性和安全性,实验结果表明,协议具有较小的误差,其计算和通信开销均优于现有设计方案。Aiming at the information leakage problem in the process of deep neural network model calculation,a series of secure and efficient interactive computing protocols were designed between two non-collusive edge servers in combination with the additive secret sharing scheme.Since the nonlinear function cannot be split directly,a set of basic conversion protocols were proposed to realize the secure conversion of additive and multiplicative shares.After a few invokes,the power,comparison,exponential,logarithm,division and other low-level functions can be calculated securely.Due to the characteristics of data transfer and computation,the proposed protocols can be extended to array computation.Theoretical analysis ensures the correctness,efficiency and security of these protocols.The experimental results show that the error of these protocols is negligible,and the computational costs and communication overhead are better than the existing schemes.

关 键 词：深度神经网络 加性秘密共享 安全计算协议 加法−乘法转换 数组单元 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]