Usage of Convolutional Neural Network for Deepfake Video Detection with Face-Swapping Technique
Daniel Stephen, Teddy Mantoro · 2022
Deepfakes have started to become a tool that brings a negative impact on society. There are various approaches that have been implemented in Deepfake Video detection, such as human-centered approach via dynamic prototypes, dynamic face augmentation, and usage of Neural Networks. However, some of the approaches that have been proposed only used several features, such as frame-by-frame detection. This paper will demonstrate the usage of Convolutional Neural Network (CNN) in detecting deepfake videos that are made with the face-swapping method. The aim of the study is to assess the feasibility of multiple combination between CNN architectures and their training dataset to detect deepfake videos made with face-swapping method that is taken from the Celeb-DF dataset. The assessment results in EfficientNetB4, combined with FaceForensics++, become the best model according to its detectability and false positive rate, when compared with other CNN architectures, while Xception trained with DFDC has the most minimum false positive rate, but the least detectability. Several improvements can be made to the research, such as the usage of GAN-based dataset for testing, usage of self-trained model with training dataset, usage of Siamese CNN architecture and comparison between Siamese and non-Siamese CNN architecture.