北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2020, Vol. 43 ›› Issue (5): 48-56.doi: 10.13190/j.jbupt.2020-068

• 论文 • 上一篇    下一篇

基于前向学习网络的人脸欺诈检测

宋昱, 孙文赟, 陈昌盛   

  1. 1. 深圳大学 电子与信息工程学院, 深圳 518060;
    2. 深圳大学 深圳市媒体信息内容安全重点实验室, 深圳 518060;
    3. 深圳大学 广东省智能信息处理重点实验室, 深圳 518060;
    4. 深圳大学 广东省人工智能与数字经济实验室, 深圳 518060;
    5. 深圳市人工智能与机器人研究院, 深圳 518060
  • 收稿日期:2020-06-15 发布日期:2021-03-11
  • 作者简介:宋昱(1988-),男,博士后,E-mail:songy@szu.edu.cn.
  • 基金资助:
    中国博士后科学基金项目(2019M663068);广东省基础与应用基础研究基金项目(2019A1515110425);广东省自然科学基金项目(2020A1515010563);深圳市科技计划项目(JCYJ20180305124550725)

Few-Shot Face Spoofing Detection Using Feedforward Learning Network

SONG Yu, SUN Wen-yun, CHEN Chang-sheng   

  1. 1. College of Electronics and Information Engineering, Shenzhen University, Shenzhen 518060, China;
    2. Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China;
    3. Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University, Shenzhen 518060, China;
    4. Guangdong Laboratory of Artificial Intelligence and Digital Economy, Shenzhen University, Shenzhen 518060, China;
    5. Shenzhen Institute of Artificial Intelligence and Robotics for Society, Shenzhen 518060, China
  • Received:2020-06-15 Published:2021-03-11

摘要: 为了克服现有人脸欺诈检测方法在少样本应用场合下的局限性,将前向学习网络用于欺诈检测.通过前向学习的方式从图像中无监督地学得卷积滤波器,在人脸欺诈检测应用场合下,对前向学习网络进行了改进,改进后的网络使用了面向人脸欺诈检测任务的卷积滤波器.使用主成分分析变换所得的最小特征值对应的特征向量作为卷积滤波器提取图像的特征.将所提方法在CASIA-FASD、Idiap Replay-Attack和OULU-NPU数据集上进行了验证,实验结果表明,在少样本跨攻击类型实验中,所提方法显著提升了欺诈人脸检测的准确率.

关键词: 人脸欺诈检测, 前向学习网络, 表示学习

Abstract: In order to overcome the limitations of the existing face spoofing detection methods under few-shot face anti-spoofing applications, this paper proposes to use feedforward learning network for face anti-spoofing. The convolutional filters are learned unsupervisedly from the images in a feedforward manner. The feedforward learning network is adapted in the spoof face detection applications by using face anti-spoofing task-oriented convolutional filters learned from the training images. The eigenvectors that correspond to the smallest eigenvalues obtained from the principle component analysis transform are used as convolution filters for extracting features from images. The method is evaluated on some benchmark datasets including CASIA-FASD dataset, Idiap Replay-Attack dataset and OULU-NPU dataset. Experiments show that under the cross presentation attack detection experiments, the proposed method significantly improves the classification accuracy of existing methods.

Key words: face spoofing detection, feedforward learning network, representation learning

中图分类号: