面向身份认证信息隐私的CycleGAN对抗样本保护研究  

Research on CycleGAN adversarial example protection for privacy of identity authentication information

作　　者：徐宏韬 崔江燕 陈希 郑志欢 叶蕴芳 郑嘉熙 XU Hong-tao;CUI Jiang-yan;CHEN Xi;ZHENG Zhi-huan;YE Yun-fang;ZHENG Jia-xi(China Mobile Group Fujian Co.,Ltd.Putian Branch,Putian 351100,China;China Mobile Group Fujian Co.,Ltd.,Fuzhou 350015,China)

机构地区：[1]中国移动通信集团福建有限公司莆田分公司,莆田351100 [2]中国移动通信集团福建有限公司,福州350015

出　　处：《电信工程技术与标准化》2025年第2期59-66,共8页Telecom Engineering Technics and Standardization

摘　　要：随着社会数字化发展,身份认证转向结合人脸和身份信息的图像模式,但这也使信息安全面临隐私泄露风险。为了兼顾身份认证信息的视觉效果和安全性,本文提出一种隐私保护机制,利用可逆性对抗样本和CycleGAN模型,既保护用户图像信息,又保持视觉呈现效果。为了验证模型的有效性,本文基于真实的用户信息构建一个公司内部数据集,并与现有相关模型进行实验对比,结果表明所提模型能够有效抵抗对身份认证信息的保护,对视觉呈现效果没有影响,能够还原对抗样本,提升了信息的安全性。With the development of digital society,identity authentication has shifted towards an image pattern that combines facial and identity information,which also poses a risk of privacy leakage.To balance the visual effect and security of identity authentication information,this paper proposes a privacy protection mechanism using reversible adversarial examples and CycleGAN models,which protects user image information while maintaining visual presentation effects.To verify the effectiveness of the model,this paper constructs an internal company dataset based on real user information and compares experiments with existing related models.The results show that the proposed model can effectively resist protection against identity authentication information,has no impact on visual presentation effects,can restore adversarial examples,and enhances the security of information.

关 键 词：CycleGAN 对抗样本 身份认证信息 隐私保护 

分 类 号：TN918[电子电信—通信与信息系统]