Face Anti-Spoofing Detection Using Least Square Weight Fusion of Channel-Based Feature Classifiers

EasyChair Preprint 2701

Abstract

Current face biometric systems are vulnerable to spoofing attack. In order to deal with various spoofing attacks, such as photo attack and video replay attack, a stable anti-spoofing algorithm should be designed. In this paper, we proposed a novel face anti-spoofing detection algorithm using least square weight fusion of channel-based feature classifiers. To this end, we first fuse the color and texture features through information entropy, the spatial and frequency features are then filtered and fused by SVM-RFE feature selection method. In addition, the fusion features of two convolutional neural networks are constructed by autoencoder. Second, for the generated three kinds of fusion features, we adpot AdaBoost, SVM and Randomforest to accomplish the robust classification, respectively. The final goal of the proposed methodis to utilize the least square method to adjust the optimal weights of the obtained three kinds of classification results, by this means, the stable and efficient face anti-spoofing detection result can be achieved. Experimental results conducted on two of the most challenging anti-spoofing datasets, including CASIA FASD and Replay-Attack, demonstrate the effectiveness of the proposed method.

Keyphrases: Convolutional Neural Network, Face anti-spoofing detection, channel feature, least square

@booklet{EasyChair:2701,
  author    = {Xiaoning Song and Qiqun Wu and Dongjun Yu and Guosheng Hu and Xiaojun Wu},
  title     = {Face Anti-Spoofing Detection Using Least Square Weight Fusion of Channel-Based Feature Classifiers},
  howpublished = {EasyChair Preprint 2701},
  year      = {EasyChair, 2020}}