A DeepFake compressed video detection method based on dense dynamic CNN

The emergence of DeepFake poses serious risks to data privacy and social stability. We propose an end-to-end DeepFake video detection method based on a dense dynamic convolutional neural network (CNN) to address the poor performance of DeepFake video detection on complex compression formats and datasets of different forgery methods. In this method, extracted face images are clustered and cleaned by cosine similarity, and face images are expanded through data augmentation to improve data diversity. Dynamic dense blocks are incorporated in a CNN to address optimization difficulties in deep neural networks, and an attention mechanism further improves generalization power. Convolution kernel pruning increases processing speed by effectively reducing the computational needs due to dynamic convolution. Experiments demonstrate that this method has better results on DeepFake video detection across compression rates and datasets compared to other network models.