Face spoof detection using feature map superposition and CNN

Face biometrics have been widely applied for user authentication systems in many practical scenarios, but the security of these systems can be jeopardised by presenting photos or replays of the legitimate user. To deal with such threat, many handcraft features extracted from face images or videos were used to detect spoof faces. These methods mainly analysed either illumination differences, colour differences or textures differences, but did not fusion these features together to further improve detection performance. Thus in this paper, we propose a novel face spoof detection method based on various feature maps and convolution neural network for photo and replay attacks. Specifically, both facial contour and specularly reflected features are considered, and proposed network is task oriented designed, e.g., its depth and width, and specific convolutional parameters of each layer are chosen for optimal accuracy and efficiency. A remarkable performance through plenty of experiments on multiple datasets shows that our method can defend not only photo attack, but also replay attack with a very low error probability.