文献类型：期刊文献

中文题名：基于深度特征增广的跨域小样本人脸欺诈检测算法

英文题名：Cross-domain Few-shot Face Spoofing Detection Method Based on Deep Feature Augmentation

作者：孙文赟[1];金忠[2];赵海涛[3];陈昌盛[1]

机构：[1]深圳大学电子与信息工程学院,深圳市媒体信息内容安全重点实验室,广东深圳518060;[2]南京理工大学计算机科学与工程学院高维信息智能感知与系统教育部重点实验室,南京210094;[3]华东理工大学信息科学与工程学院,上海200237

年份：2021

卷号：48

期号：2

起止页码：330

中文期刊名：计算机科学

外文期刊名：Computer Science

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金(61902250,61702340,61872188);国家重点基础研究发展计划(2014CB349303);中国博士后科学基金面上项目(2018M643183);广东省自然科学基金(2017A030310382);深圳市基础研究自由探索项目(JCYJ20180305124550725,827/000213)。

语种：中文

中文关键词：模式识别;系统安全;人脸图像分析;人脸欺诈检测;深度学习

外文关键词：Pattern recognition;System security;Facial image analysis;Face spoofing detection;Deep learning

摘要：随着人脸识别技术的发展,人脸欺诈攻击已经成为一项实际的安全问题,人脸欺诈检测算法用于及早发现该类攻击,保护系统安全。文中将一种经典域自适应算法扩展到深度神经网络中,首先定义了基于深度特征增广的域自适应层,提出了一种基于深度特征增广的跨域小样本人脸欺诈检测算法。该算法在已有的基于全卷积神经网络的人脸欺诈检测深度神经网络的中部嵌入域自适应层将卷积特征图增广,来适配源域和目标域的差异,随后根据增广后的特征图进行像素级分类,最后将像素级概率图从空间上融合为帧级决策。文中在CASIA-FASD,Replay-Attack和OULU-NPU 3个数据集和6个常见测评协议(2个CASIA-FASD与Replay-Attack跨库协议和4个OULU-NPU标准协议)下进行实验,验证了算法在不同背景、不同攻击设备、不同相机等跨域情况下的性能。实验表明,基准FCN人脸欺诈检测算法已经能够达到较好的性能,在此基础上,借助小样本目标域数据学习域自适应模型,可进一步显著提升性能,将错误率减半(CASIA-FASD训练+Replay-Attack测试的HTER指标从27.31%降至11.23%,Replay-Attack训练+CASIA-FASD测试的HTER指标从37.33%降至21.83%,OULU-NPU标准协议IV的ACER指标从9.45%降至5.56%),实验结果验证了基于深度特征增广的跨域小样本人脸欺诈检测算法的有效性。
The face recognition technology is improving rapidly these days.One the other side,the face presentation attack has become a practical security problem.To protect the system,face presentation attack detection methods are employed for detecting such attacks in advance.This paper extends a classic domain adaptation method to the deep neural network scenario,defines a feature augmentation-based domain adaptation layer,proposes a cross-domain few-shot face presentation attack detection method based on deep feature augmentation.This method is based on the existing method based on Fully Convolutional Network and improves the existing method by embedding a domain adaptation layer in the middle of the network.The new layer augments the feature maps,adapts the difference between the source and target domains.Then,a pixel-level probability map is predicted based on the augmented the feature maps.Finally,the prediction map is fused to a frame-level decision.Experiments are conducted on the CASIA-FASD,Replay-Attack and OULU-NPU datasets.Six commonly used protocols including the cross-dataset protocols between CASIA-FASD and Replay-Attack,the standard protocols of the OULU-NPU dataset are followed.The training and test data are cross different backgrounds,presentation attack instruments and cameras.The experiment results show that the baseline method,the Fully Convolutional Networkbased face presentation attack detection method has already achieved state-of-the-art performance.The performance can be further improved by learning the domain adaptation model on small-sample data in the target domain.The proposed method can halve the error rate by introducing domain adaptation(train on CASIA-FASD and test on Replay-Attack:decreased from 27.31% to 11.23%,train on Replay-Attack and test on CASIA-FASD:decreased from 37.33% to 21.83%,OULU-NPU’s standard protocol IV:decreased from 9.45%to 5.56%).This confirms the effectiveness of the proposed method.