Fake Face Detection using Local Binary Pattern and Ensemble Modeling
Yonghui Wang, Vahid Zarghami, Suxia Cui
Abstract
Fake faces generated with Generative Adversarial Networks (GANs) are becoming more and more realistic and getting harder to be identified directly by human beings. However, CNN (Convolutional Neural network) based deep learning architecture can achieve almost perfect detection accuracy on such fake faces. In this paper we present a study of fake face detection with the exploration of the global texture features based on the empirical knowledge that the textures of fake faces are quite different from those of real faces. A new architecture, LBP (Local Binary Pattern)-Net, is designed to utilize binary representation image texture for the effective identification of fake images. Experimental results show that the proposed method is more robust than existing algorithms for detecting fake images edited by different image augmentation methods, such as blurring, cutout, brightness and color changing, equalization, etc. Ensemble models are also experimented to combine advantages of individual models. The most significant effect of ensemble models is the robustness for detecting edited fake images compared to single models. Experimental results show that our ensemble models outperform single models for detecting fake images.