Spoofings detection in facial biometric systems

In this thesis, we propose models for face spoofing detection with low error rate in a face biometric verification system using a standard imaging device. The study is evaluated under two headings. In the first stage, spoofing detection is performed by classifying the features obtained from traditional texture identification methods with the help of classical learning algorithms. In this stage, firstly, it is aimed to find the face region in an input image that has the most effect on spoofing detection. Then, the effects of color spaces on the problem were investigated and a pattern encoder with low feature vector size was proposed. In the second stage, the effect of deep learning algorithms on face spoofing detection is investigated. In this section, a reduced version of the Xception network, which is frequently used in the literature, is produced by reducing the number of parameters. Attention blocks were added to this model and the information in the channels was weighted on the feature coders. Finally, a hybrid deep learning model including classical feature extraction method and deep learning models is proposed and the performances of the systems are evaluated. The results show that color information is effective in spoofing detection, the proposed feature encoder has low computational complexity with strong identification capability, and the performance of deep learning models in face spoofing detection is better than classical methods.