Detecting Deepfake Videos via Frame Serialization Learning

Deepfake, a video forgery technique based on Generative Adversarial Network (GAN), has been proved to be a serious threat to the public security. The images and videos generated with it can even fool human eyes. In this paper, we propose a deep learning-based method to hunt Deepfake videos. A 3D-ResNext based model is developed to effectively learn the leverageable difference between fake videos and benign ones from multiple serialized frames. Furthermore, to address the information loss in compressed videos, the data enhancement technique is introduced in data preprocessing to collect sufficient training samples from public datasets, e.g., FaceForensics++, DeepFakeDetection and DFDC. The experiments demonstrated that our method has good performance and generalization ability in the task of Deepfake videos detection.