基于卷积神经网络的人脸图像隐私保护  

Privacy-preserving face images based on convolutional neural networks

作　　者：沈博 王祥根[1] 田澍 崔佳佳 SHEN Bo;WANG Xianggen;TIAN Shu;CUI Jiajia(North China Institute of Computing Technology,Beijing 100083,China;School of Information and Communication Engineering,Beijing Information Science and Technology University,Beijing 100192,China)

机构地区：[1]华北计算技术研究所,北京100083 [2]北京信息科技大学信息与通信工程学院,北京100192

出　　处：《沈阳师范大学学报（自然科学版）》2023年第3期280-288,共9页Journal of Shenyang Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(62001032)。

摘　　要：针对现有基于深度生成网络模型的人脸图像隐私保护方法无法提供可证明隐私保证、合成图像与原始图像保持语义一致性的问题,提出一种基于卷积神经网络的人脸图像隐私保护方法。该方法首先基于卷积自动编码器和差分隐私实现人脸图像的预训练,对原始人脸图像进行解耦和身份信息的差分隐私保护;然后利用卷积生成对抗网络合成伪图像代替原始图像发布,在保留原始人脸图像的关键特征的基础上,生成与原始图像的关键人脸属性高度匹配的伪图像。该方法可保证合成图像与原始图像语义一致性,并提供可证明的隐私保证。与现有的基于深度生成模型人脸图像隐私方法相比,所提出的方法达到了更好的隐私保护与数据可用性之间的优化权衡。Aiming at the problem that the existing face image privacy protection method based on the deep generative network model which cannot provide a provable privacy guarantee and maintain semantic consistency between the synthetic image and the original image,a privacy protection method for face image based on convolutional neural network is proposed.This method first realizes the pre-training of face images based on convolutional autoencoder and differential privacy,decouples the original face images,and protects the differential privacy of identity information;then uses the convolutional generative adversarial network to synthesize fake images instead of the original images for publishing,based on preserving the key features of the original face image,generate a pseudo image that highly matches the key face attributes of the original image.The method guarantees the semantic consistency between the synthesized and original images and provides a provable privacy guarantee.Compared with existing face image privacy methods based on deep generative models,the method proposed in this paper achieves a better optimal trade-off between privacy protection and data utility.

关 键 词：生成对抗网络 自动编码器 差分隐私 人脸图像 隐私保护 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]