阎少宏, 宗晨琪, 郭宸良.基于线云的三维人脸特征模板保护方案[J].河南理工大学学报（自然科学版）,2025,44(5):27-34.

YAN S H, ZONG C Q, GUO C L.A 3D facial feature template protection scheme based on line cloud[J].Journal of Henan Polytechnic University(Natural Science) ,2025,44(5):27-34.

基于线云的三维人脸特征模板保护方案

阎少宏1,2, 宗晨琪1, 郭宸良1

1.华北理工大学 理学院，河北 唐山  063210；2.河北省数据科学与应用重点实验室，河北 唐山  063210

摘要: 目的 针对三维人脸点云模板存储的保密性无法满足识别系统隐私安全的问题，提出一种基于线云加密的人脸模板保护方案，并根据加密模板特征设计一个适用于线云的三维人脸识别原型系统。  方法 在注册阶段，根据人脸标志点（眼角、鼻尖等）序号，在人脸空间中得到其三维几何信息进行存储，由于标志点矩阵密度稀疏，信息量少，因此并不会使模板泄密。以标志点为中心建立刚性区域子空间，改进传统线云算法，用于对子空间点云模板的线云化加密。在匹配阶段，将待认证人脸进行刚性区域分割，获得标志点矩阵，建立刚性区域点云空间。使用ICP（iterative closest point）算法将待认证人脸与模板人脸的标志点进行匹配得到转换矩阵。利用转换矩阵将待认证人脸与加密后的模板人脸进行配准，计算每个刚性子区域点云与线云模板的距离。根据不同区域准确率赋予不同的权重，将加权后的距离之和作为最终判断标准，完成加密情况下的人脸识别。 结果 仿真结果表明，该方案能够有效抵抗隐私攻击，在面对暴力攻击和基于密度的线云攻击时具有良好的保护效果，同时保持较高的识别效果，在达到最优识别效果时其真正例率为83.861%，假正例率为16.139%，等错误率为16.139%。 结论 本文所提算法与识别模型的设计在保护三维人脸点云几何特征的基础上保证了其作为生物特征在身份验证领域的有效性，可为三维人脸点云特征模板的保护提供参考。

关键词:三维人脸识别;生物特征模板保护;标志点;刚性区域;线云

DOI:10.16186/j.cnki.1673-9787.2024070041

基金项目:国家自然科学基金资助项目（U20A20179）

收稿日期:2024/07/09

修回日期:2024/09/29

出版日期:2025/07/23

A 3D facial feature template protection scheme based on line cloud

Yan Shaohong1,2, Zong Chenqi1, Guo Chenliang1

1.College of  Science， North China University of Science and Technology， Tangshan  063210， Hebei， China；2.Hebei Provincial Key Laboratory of Data Science and Application， Tangshan  063210， Hebei， China

Abstract: Objectives Aiming at the problem that the storage security of 3D face point cloud template could not satisfy the privacy and security of recognition system， a face template protection scheme based on line cloud encryption was proposed ，and a 3D face recognition prototype system suitable for line cloud was designed according to the characteristics of the encrypted template. Methods In the registration stage， according to the serial number of face mark points （corner of the eye， nose tip， etc.）， the three-dimensional geometric information was obtained in the face space for storage， the template would not be leaked because of the sparse density of the marker matrix and less information. The rigid region subspace was built with the mark points as the center， and the traditional line cloud algorithm was improved to complete the line cloud encryption of the subspace point cloud template. In the matching stage， the face to be authenticated was subjected to rigid region segmentation to obtain a matrix of landmark points and establish a rigid region point cloud space. ICP（iterative closest point） algorithm was used to match the mark points of the face to be verified and the template face to get the transformation matrix. Using the transformation matrix to register the face to be verified with the encrypted template face， the distances between each rigid subregion point cloud and line cloud template were calculated. Different weights based on the accuracy of each regions were given to them， and the sum of the weighted distance would be used as the final judgment criterion to complete face recognition under encrypted conditions. Results The simulation results showed that this scheme could effectively resist privacy attacks and had good protection effects against brute-force attacks and line cloud attacks based on density； At the same time， it maintained a high recognition performance， with a true positive rate of 83.861%， a false positive case rate of 16.139%， and an equal error rate of 16.139% when achieving optimal recognition performance.  Conclusions The design of the algorithm and recognition model proposed in this article ensured its effectiveness as a biometric feature in the field of identity verification while protecting the geometric features of the 3D facial point cloud， which provided reference for the protection of 3D facial point cloud feature templates.

Key words:3D facial recognition;biometric template protection;marking points;rigid area;line cloud