V3CViT: Deepfake Detection Based on Video Vision Transformer and 3D Convolution Network

With the advancement of Generative Adversarial Networks in synthesizing fake face videos, the proliferation of forged products poses a significant threat to societal security. In video-based detection methods, compression operations cause the disappearance of discriminative features, and the use of deep learning methods often overlooks the interrelationships between features, resulting in limited model generalization. Therefore, we propose a deepfake video detection framework named V3CViT. This framework combines a video vision transformer and 3D convolutional neural networks to extract facial features, capturing facial characteristics in both temporal and spatial dimensions and combining them to extract effective features using attention mechanisms. Subsequently, the structured facial feature maps are utilized with a Gated Graph Convolutional Network to learn facial relationship information for the detection task. The experiments show that compared to existing detection methods, our model can comprehensively capture facial features in forged videos and achieves more significant accuracy on the mixed dataset.