Efficient width-extended convolutional neural network for robust face spoofing detection
MetadataShow full item record
Biometrics has been increasingly used as a safe and convenient technique for people identification. Despite the higher security of biometric systems, criminals have already developed methods to circumvent them, being the presentation of fake biometric information to the input sensor (spoofing attack) the easiest way. Face is considered one of the most promising biometric traits for people identification, including in mobile devices. However, face recognition systems can be easily fooled, for instance, by presenting to the sensor a printed photograph, a 3D mask, or a video recorded from the face of a legal user. Recently, despite some CNNs (Convolutional Neural Networks) based approaches have achieved state-of-the-art results in face spoofing detection, in most of the cases the proposed architectures are very deep, being unsuitable for devices with hardware restrictions. In this work, we propose an efficient architecture for face spoofing detection based on a width-extended CNN, which we called wCNN. Each part of wCNN is trained, separately, in a given region of the face, then their outputs are combined in order to decide whether the face presented to the sensor is real or fake. The proposed approach, which learns deep local features from each facial region due to its width-wide architecture, presented better accuracy than state-of-the-art methods, including the well-referenced fine-tuned VGG-Face, while being much more efficient regarding hardware resources and processing time.