| 摘要: |
| 人脸反欺诈(Face anti-spoofing,FAS)在防止人脸识别系统遭受欺诈攻击方面起着至关重要的作用,得益于深度学习网络强大的特征提取能力,基于深度学习的FAS算法取得比基于传统手工特征算法更好的性能,成为近期的研究热点。尽管大多数基于深度学习的FAS算法能在库内达到很好的检测效果,但是跨库检测性能欠佳,主要原因是库内和库外数据往往在不同条件下采集,例如拍摄设备、环境光照和攻击呈现设备不同,导致库内和库外数据的分布不同,两者之间存在域位移。当训练数据的多样性不足时,容易在库内学习过程中过拟合,跨库泛化性能不好。尽管我们可以判断起因,然而在真实世界的应用过程中解决上述问题并不容易。一方面,人脸反欺诈模型难以收集所有场景下的有标签训练样本;另一方面,不同应用场景使得同一因素产生不同的影响,例如,不同场景的光照导致域位移,影响了分类模型对本质性欺诈纹理的提取。为此,本文将元伪标签引入人脸反欺诈任务,提出一种基于元伪标签的人脸反欺诈方法。主要贡献包括:第一,提出一种基于图像块的“教师生成伪标签,学生反馈”半监督学习框架,挖掘局部图像的高区分度特征,解决有标签样本不足的问题;第二,基于局部重力模式(Pattern of localgravitational force,PLGF),设计一种带有注意力模块的光照不变特征分支,抑制应用场景中最容易影响特征提取的光照因素;第三,将元学习与半监督学习框架相结合,优化教师生成伪标签的过程,提高算法的跨库检测能力。与现有流行算法相比,在三个公开的测试数据集(包括CASIA、Replay-Attack和MSU)上,所提出方法在库内测试和跨库测试下均有突出的表现,尤其是泛化性能得到显著提高。在样本数量中等时,在不同库中的半总错误率保持最低。 |
| 关键词: 人脸反欺诈|元学习|半监督学习|光照不变特征|元伪标签|深度学习 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2023.09.05 |
| Received:January 30, 2022Revised:April 08, 2022 |
| 基金项目:本课题得到国家重点研发计划项目(No. 2019QY2202)、广州开发区国际合作项目(No. 2019GH16)和中新国际联合研究院项目(No.206-A018001)资助。 |
|
| Face Anti-Spoofing Based on Meta-pseudo-label and Illumination-invariant Feature |
| FENG Haoyu,HU Yongjian,WANG Yufei,LIU Beibei,YU Xiangyu,ZHONG Rui |
| School of Computer Science and Engineering, South China University of Technology, Guangzhou 510641, China;School of Electronic and Information Engineering, South China University of Technology, Guangzhou 510641, China;Department of Criminal Science and Technology, Guangdong Police College, Guangzhou 510440, China |
| Abstract: |
| Face anti-spoofing (FAS) plays a vital role in preventing face recognition systems from presentation attacks. Benefitted from powerful capability of feature extraction of deep learning (DL) network, FAS algorithms based on deep learning are much superior to those ones based on traditional handcrafted features in detection performance and thus become a research hotspot. Although most DL-based FAS algorithms can achieve good performance for intra-database test, the performance decreases greatly for cross-database test. The main reason is that samples are often collected under different capturing settings for intra-database and cross-database, for example, different cameras, environment illuminations, presentation medium, and thus their distributions are different, which can lead to the domain shift problem. When the diversity of training samples is insufficient, a model trained with such samples can be easily overfitting for intra-database while not being able to perform well for cross-database. Although the reasons of poor generalization are clear, the solution to them cannot be easily achieved in real-world applications. On the one hand, it is difficult for the FAS model to collect labeled training samples for all scenarios; on the other hand, different application scenarios make the same factor behave differently, which affects the extraction of intrinsic spoofing textures. In this paper, we introduce meta-pseudo-label into the FAS task, and propose a FAS method based on meta-pseudo-label. There are three major contributions. First, we propose a semi-supervised learning framework of “teacher generates pseudo-label and student feedbacks” based on image patches, which extracts the highly discriminative features of local images to solve the problem of insufficient labeled samples. Second, based on the Pattern of Local Gravitational Force (PLGF), we design an illumination-invariant feature branch with an attention module to suppress the illumination influence on feature extraction in application scenarios. Third, the meta-learning is combined with the semi-supervised learning framework to optimize the process of generating pseudo-label by teacher and improve the generalization ability of the algorithm. Compared with state-of-the-art algorithms, the proposed method performs quite well in both intra-database test and cross-database test on three public datasets including CASIA, Replay-Attack, and MSU. Specifically, the performance in cross-database is greatly improved. It can achieve the lowest HTER values for middle-sized sample number. |
| Key words: face anti-spoofing|meta-learning|semi-supervised learning|illumination-invariant feature|meta-pseudo-label|deep learning |
