对抗样本在人脸匿名中的应用  

Application of Adversarial Samples in Facial Anonymity

作　　者：孙明辉 叶阿勇[1] SUN Minghui;YE Ayong(College of Computer and Cyber Security,Fujian Normal University,Fuzhou,China,350117)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福州350117

出　　处：《福建电脑》2024年第2期55-58,共4页Journal of Fujian Computer

基　　金：移动网络中面向位置信息生命周期的隐私保护关键技术(No.61972096)资助。

摘　　要：人脸识别模型给生产生活带来了便利,但其出色的识别能力也对个人隐私产生了巨大的威胁。为保护图像中的隐私信息,传统的人脸图像匿名方法对人脸图像进行较大的改动,严重破坏了图像的可用性。本文提出一种基于对抗样本的人脸图像匿名方法,通过向人脸图像中添加细微的对抗扰动,使人脸识别模型无法正确推理出其真实身份。实验结果表明,经过匿名处理后的人脸图像能够在主流的人脸识别模型上取得85%以上的匿名精度。Facial recognition models have brought convenience to production and daily life,but their excellent recognition ability also poses a huge threat to personal privacy.To protect the privacy information in images,traditional anonymous methods for facial images make significant changes,which seriously damages the usability of the images.This article proposes a face image anonymity method based on adversarial samples,which adds subtle adversarial perturbations to the face image to prevent the face recognition model from accurately inferring its true identity.The experimental results show that facial images processed anonymously can achieve an anonymity accuracy of over 85%on mainstream facial recognition models.

关 键 词：隐私保护 图像隐私 人脸图像匿名 对抗样本 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]