Deep Learning Based One-Class Detection System for Fake Faces Generated by GAN Network

Recently, the dangers associated with face generation technology have been attracting much attention in image processing and forensic science. The current face anti-spoofing methods based on Generative Adversarial Networks (GANs) suffer from defects such as overfitting and generalization problems. This paper proposes a new generation method using a one-class classification model to judge the authenticity of facial images for the purpose of realizing a method to generate a model that is as compatible as possible with other datasets and new data, rather than strongly depending on the dataset used for training. The method proposed in this paper has the following features: (a) we adopted various filter enhancement methods as basic pseudo-image generation methods for data enhancement; (b) an improved Multi-Channel Convolutional Neural Network (MCCNN) was adopted as the main network, making it possible to accept multiple preprocessed data individually, obtain feature maps, and extract attention maps; (c) as a first ingenuity in training the main network, we augmented the data using weakly supervised learning methods to add attention cropping and dropping to the data; (d) as a second ingenuity in training the main network, we trained it in two steps. In the first step, we used a binary classification loss function to ensure that known fake facial features generated by known GAN networks were filtered out. In the second step, we used a one-class classification loss function to deal with the various types of GAN networks or unknown fake face generation methods. We compared our proposed method with four recent methods. Our experiments demonstrate that the proposed method improves cross-domain detection efficiency while maintaining source-domain accuracy. These studies show one possible direction for improving the correct answer rate in judging facial image authenticity, thereby making a great contribution both academically and practically.