一种基于CKKS同态加密与神经网络的安全人脸识别方案  

A secure face recognition scheme based on CKKS homomorphic encryption and neural network

作　　者：苏昀暄 涂正 王绪安[1,2] 林川 SU Yun-xuan;TU Zheng;WANG Xu-an;LIN Chuan(Key Laboratory for Network and Information Security of the PAP,Engineering University of the PAP,Xi’an 710086,China;The Foundation of Guizhou Provincial Key Laboratory of Public Big Data,Guizhou University,Guiyang 550000,China)

机构地区：[1]武警工程大学网络与信息安全武警部队重点实验室,陕西西安710086 [2]贵州大学贵州公共大数据国家重点实验室,贵州贵阳550000

出　　处：《兰州理工大学学报》2023年第2期103-109,共7页Journal of Lanzhou University of Technology

基　　金：国家自然科学基金(61772550);贵州省公共大数据重点实验室开放课题基金(2019BDFKJJ008)。

摘　　要：随着大数据技术的不断发展,人脸识别应用越来越广泛,但随之而来的是用户隐私数据泄露等安全问题.针对此,提出一种在云服务器下基于同态加密与神经网络的人脸识别方案.通过CKKS同态加密方案对人脸图像进行加密,在云服务器中通过ResNet50 Model与Arcface Loss函数对LFW数据集进行训练,计算加密图片之间的向量相关性,对密文结果解密并比较阈值判断是否为同一个人,实现人脸识别.实验结果表明,该方案在LFW数据集上,阈值大约为0.25时,密文中的识别准确率达到99.398%,证明本文方案具有较高的识别精度.As big data technology has been evolving over the years,facial recognition is widely used in different fields.However,users are then troubled by safety problems,such as privacy disclosure and data leakage.A face recognition scheme based on homomorphic encryption and neural network under the cloud server is proposed in this paper.The face image is encrypted through the CKKS homomorphic encryption scheme,and the LFW data set is trained through the ResNet50 model and Arcface Loss function in the cloud server,then the vector correlation between the encrypted images is calculated,followed by the decryption of the ciphertext result and comparison with the threshold to judge whether it is the same person,to realize face recognition.The experiment result shows that on the LFW dataset,the scheme of this paper can recognize the cryptograph with an accuracy rate of 99.398%when the threshold value is about 0.25.This result proves the proposed scheme has high recognition accuracy.

关 键 词：云计算 人脸识别 同态加密 神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]