An improved simple feature set for face presentation attack detection

Abstract

Presentation attacks are weak points of facial biometrical authentication systems. Although several presentation attack detection methods were developed, the best of them require a sufficient amount of training data and rely on computationally intensive deep learning based features. Thus, most of them have difficulties with adaptation to new types of presentation attacks or new cameras. In this paper, we introduce a method for face presentation attack detection with low requirements for training data and high efficiency for a wide range of spoofing attacks. The method includes feature extraction and binary classification stages. We use a combination of simple statistical and texture features and describe the experimental results of feature adjustment and selection. We validate the proposed method using WMCA dataset. The experiments showed that the proposed features decrease the average classification error in comparison with the RDWT-Haralick-SVM method and demonstrate the best performance among non-CNN-based methods.

Face Anti-Spoofing Detection with Multi-Modal CNN Enhanced by ResNet

“…Reference [5] showcased the RDWT-Haralick-SVM algorithm achieving an ACER of 3.44%, while the MC-CNN method obtained 0.3% ACER using the CDIT metric. Meanwhile, reference [7] reported an ACER of 2.91% with a dataset of 1679 images and 1.18% with 83,950 images. Nevertheless, the proposed method demonstrated superior performance compared to these results, as presented in Table 3.…”

“…These CNN architectures will provide positive outcomes for FAS in many scenarios. These results will be compared with the outcomes of papers [5], [7], revealing that the current findings exhibit superior performance. Subsequently, the results of all modalities will be combined using different techniques, such as majority voting [8], weighted voting [9], average/pooling [10], and stacking classifiers [11].…”