联合多通道特征与最小二乘决策的人脸反欺诈方法  

Using multi-channel features and least square decision to perform face anti-spoofing

作　　者：吴启群 宋晓宁[1] Wu Qiqun;Song Xiaoning(School of Internet of Things Engineering,Jiangnan University,Wuxi Jiangsu 214122,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122

出　　处：《计算机应用研究》2020年第9期2847-2850,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61876072);国家重点研发计划子课题资助项目(2017YFC1601800);中国博士后科学基金特助项目(2018T110441);江苏省自然科学基金资助项目(BK20161135);江苏省“六大人才高峰项目”(XYDXX-012)。

摘　　要：为了应对大量的欺诈攻击,如照片攻击和视频攻击,提出了一种基于多通道特征与最小二乘法决策的人脸反欺诈方法。一方面,将人脸在不同颜色空间分量上的梯度特征进行加权融合。另一方面,为了提高实验的鲁棒性,引入BSIF纹理特征与CNN的卷积特征,并用最小二乘法对分类的结果进行最优的决策判断。实验在replay-attack和CASIA两个数据集上进行测试,其中在replay-attack数据集上的EER和HTER分别降低到了3.52%与4.63%,在CASIA数据集上的EER和HTER分别降低到了6.02%和6.45%,与目前最优的算法有一定的竞争力,表明该方法对不同方式的欺诈攻击有较好的检测效果。In order to deal with various spoofing attacks,such as photo attack and video replay attack,this paper designed an anti-spoofing algorithm based on multi-channel features and least square decision. On the one hand,it fused the gradient features of face in different color space components by weight. On the other hand,in order to improve the robustness of the experiment,it introduced BSIF texture feature and convolution feature of CNN,then it proposed a least square method to obtain an optional decision-making. Experimental results conduct on two most challenging anti-spoofing datasets replay-attack and CASIA,EER and HTER on replay-attack are reduced to 3. 52% and 4. 63%,EER and HTER on CASIA are reduced to 5. 90% and6. 32%,which are competitive with the current optimal algorithm and provide better detection results for different types of spoofing attacks.

关 键 词：人脸反欺诈 多通道特征 最小二乘法 replay-attack CASIA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]