基于深度学习和联合特征提取的人脸活体检测及决策融合攻击类型检测算法  

Face liveness detection algorithm based on deep learning and joint feature extraction and decision⁃making fusion attack type detection algorithm

作　　者：王彤 张高原 丁邦杰 杨金柱 张立立 WANG Tong;ZHANG Gaoyuan;DING Bangjie;YANG Jinzhu;ZHANG Lili(National Computer Experimental Teaching Demonstration Center,School of Computer Science and Engineering,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学计算机科学与工程学院国家级计算机实验教学示范中心,辽宁沈阳110819

出　　处：《现代电子技术》2025年第5期91-100,共10页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(62072085);国家重点研发计划课题项目(2018YFB1702003)。

摘　　要：生物特征识别技术越来越多地应用于身份认证,随之不断出现伪造合法用户信息的欺骗手段,人脸识别系统容易受到欺骗攻击,严重威胁了系统的安全性。为了提高生物特征识别系统的安全性,文中提出一种基于深度学习和联合特征提取的人脸活体检测及决策融合攻击类型检测算法。基于改进的AlexNet模型,有效降低了训练过程中的过拟合等问题,显著降低了模型训练时间;采取手工特征和深度学习相结合的模式判断非活体攻击类型,手工特征提取采取LBP结合多层DCT变换的联合特征提取,深度学习特征采取四层CNN网络的全局图像特征提取;在攻击类型判别上,提取待测样本的局部和全局特征进行初步判定,再通过决策融合将两个SVM分类器的输出结果以加权方式进行整合。算法在公开的CASIA数据集和NUAA数据集上进行验证,实验结果表明,融合不用的信息可以获得更高的准确率,降低了计算的复杂度,提高了算法的效率。Biometric identification techniques are increasingly applied in identity authentication.The fraudulent means of forging legitimate user information appears.Face recognition system is vulnerable to deception attacks,which threaten the security of the system seriously.Therefore,this paper presents a face liveness detection algorithm based on deep learning and joint feature extraction and a decision-making fusion attack type detection algorithm,so as to improve the safety of the biometric identification system.On the basis of the improved AlexNet model,the overfitting of the model are reduced effectively during the process of training and its training time is reduced significantly.The pattern of combining manual features and deep learning is adopted to judge the type of non-living attacks.The joint feature extraction of LBP(local binary pattern)combined with multi-layer DCT(discrete cosine transform)is adopted for the manual feature extraction,and the global image feature extraction of four-layer CNN(convolutional neural network)is adopted for the deep learning feature.In the discrimination of attack types,the local and global features of the samples to be tested is extracted for preliminary determination.Then,by decision-making integration,the output results of the two SVM(support vector machine)classifiers are integrated in a weighted way.The algorithm is validated on publicly available CASIA datasets and NUAA datasets.The experimental results show that the fusion of unused information can obtain higher accuracy,reduce the computational complexity and improve the efficiency of the algorithm.

关 键 词：深度学习 联合特征提取 人脸活体检测 AlexNet LBP DCT 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP393[电子电信—信息与通信工程]