Cascaded Network Based on EfficientNet and Transformer for Deepfake Video Detection

With the continuous development of deepfake technology, forged videos are continuously released on various network media, which facilitates people's lives but also brings great negative impacts. And these forged videos have a high degree of authenticity, which brings great challenges to detection. However, most deepfake detection models now only focus on model design, lacking versatility. To address these issues, we proposed a cascaded Network based on EfficientNet and Transformer to achieve deepfake detection tasks. The improved convolutional neural network EfficientNetV2S is used as a feature extractor, the features are input to the Transformer, and its attention mechanism is used for classification. And used Spatial-Reduction Attention (SRA) to improve the traditional attention mechanism in Transformer. And we carefully extracted and screened real and fake faces in the preprocessing stage and trained our model on DFDC and FaceForensics++ benchmarks, achieving state-of-the-art results such as an accuracy of 92.16% and an accuracy of 96.75%, respectively. Finally, we also achieved excellent visualization results on deepfake videos, proving the practicability of our method.