Detecting Fakes in Mobile Face Recognition Systems Using a Stereo Camera

A method for detecting spoofing attacks in face recognition systems on mobile devices is proposed. The method is able to work in real time on devices with limited computing capabilities, various recording conditions, and using the standard stereo camera, characterized by a small stereo basis. The method is based on a convolutional neural network and multiobjective learning, and a special loss function is proposed. The following types of forgeries are considered: high-quality printed images of faces, digital photographs, and video sequences shown on high-resolution screens. The resulting algorithm is tested on several sets of stereo images, both publicly available and manually collected, characterized by a wide variety of shooting conditions.