Multi-scale Feature Learning with Graph Attention Network for Face Forgery Detection

Face forgery videos, known as Deepfakes, are widely spread on social media with great potential threat, making the detection of forged face videos is crucial. Commonly, forgery video detection methods are based on Convolutional Neural Networks or Transformer, which treats images as grid or sequence structures for binary classification discrimination. Since face objects are usually not regularly shaped quadrilaterals, treating them as grid or sequence structures is redundant and inflexible, thus losing useful information. Based on this, we propose a new perspective to represent facial images as graph structures, which are fed into Graph Neural Network to learn the intrinsic relationships of facial regions for deep forgery detection. In addition, we propose a feature fusion module to learn artifact information in the frequency domain for a more comprehensive facial feature representation to further improve the reliability of our model. Extensive experiments on several benchmark databases demonstrate the effectiveness and robust generalization ability of our method compared with many state-of-the-art methods.