Deep Feature Extraction for Face Liveness Detection

Face recognition is now widely being used to verify the identity of the person in various applications ranging from border crossing to mobile authentication. However, most face recognition systems are vulnerable to spoofing or presentation attacks, where a photo, a video, or a 3D mask of a genuine user's face may be utilized to fool the biometric system. Although many face spoof detection techniques have been proposed, the issue is still unsolved. Recently deep learning based models have achieved impressive results in various challenging image and video classification tasks. Consequently, very few works have applied convolutional neural networks (CNNs) for face liveness detection. Nonetheless, it is still unclear how different CNN features and methods compare with each other for face spoof detection, since prior CNN based face liveness detection approaches employ different fine-tuning procedures and/or datasets for training. Thus, in this paper, an approach based on transfer learning using some well-known and well-adopted pre-trained CNNs architectures is presented. This study explores different deep features and compares them on a common ground for face liveness detection in videos. Experimental analysis on two publicly available databases, NUAA and CASIA-FASD, shows that the proposed method is able to attain satisfactory and comparable results to the state-of-the-art methods.