Margarita N. Favorskaya, Andrey I. Pakhirka

Reshetnev Siberian State University of Science and Technology

Currently, face recognition systems are a widespread way of biometric identification in practice. However, such systems require protection against unauthorized actions in the form of so-called presentation attacks, when an attacker replaces genuine images with fake images or short video sequences. The article proposes a method for detecting presentation attacks using the depth of the scene without the use of special sensors. The challenge is to enhance the subtle difference between genuine and fake images. For this, a deep network was trained and tested, consisting of central difference convolution blocks and a multi-scale attention module. Experiments have shown that pre-processing input face images to the HSV color space has an advantage in the accuracy of detecting fake images. Thus, the detection accuracy on our own dataset, KITTI and Cityscapes datasets increased by 3-7% depending on the capture devices, lighting conditions, and settings of the algorithm.

presentation attacks, face recognition, depth maps, deep learning