EfficientNet-based multi-dimensional network optimization for Deepfake video detection
Yong Zhang, Xinqi Zhang, Bingjie Li · 2023
The rapid development of deep learning techniques, especially generative adversarial networks, has led to the generation of very realistic forged faces. Deepfake technology has brought convenience to society and also produced a large number of undesirable effects. Many detectors have been developed to defend videos generated by Deepfake manipulation techniques. In this paper, we take low overhead and high performance as the task of Deepfake detection and propose a new EfficientNet-B0 combined with ViT Deepfake detection network, which consists of two key components: (1) ConvFFN block, which brings a larger receptive field to the network; (2) Transformer Encoder with Hilo, which separates the high and low frequencies of the attention layer to focus on the global structure. We conducted extensive experiments on the DFDC dataset and the FF++ dataset to demonstrate the stability and practical applicability of the proposed method.