Improving Sequential DeepFake Detection with Local information enhancement

Existing Deepfake technology involves multi-step forgery to generate images.However, there are few sequential Deepfake detection methods available.To address this challenge, we propose a feature cross-fusion model that combines Vision Transformer (ViT) and Convolutional Neural Network (CNN), along with a novel data augmentation technique called Channel Random Erasing(CRE).This model first enhances its robustness by using CRE, which introduces controlled occlusions during training to simulate real-world manipulations.It then captures both the global and local features of images through multi-scale feature fusion.The Vision Transformer (ViT) captures global contextual information via self-attention mechanism, providing a strong global feature representation, while the Convolutional Neural Network (CNN) extracts local details through convolution operations, effectively capturing edges and texture information.Extensive experiments on the Seq-Deepfake benchmark demonstrate the effectiveness of this model, achieving better performance compared to current state-of-the-art methods.