Deepfake Detection using EfficientNet: Working Towards Dense Sampling and Frames Selection
Tuan-An To, Hoang-Chau Luong, Nham-Tan Nguyen, Trong-Tin Nguyen, Minh–Triet Tran, Trong-Le Do · 2022 RIVF International Conference on Computing and Communication Technologies (RIVF) · 2022
Deepfake is a controversial technology that allows the automatic generation of video content through generative adversarial networks. The emergence of Deepfake technology is problematic and sophisticated, making it more difficult to detect. In our paper, we contribute a deep-learning method to resolve that problem. We use the MTCNN face detector to extract facial images and apply data augmentation and EfficientNet for real-fake classification. We apply frame selection with the raw label prediction to tackle the fault cases and receive the final label. With the approach above applied, we utilize training and evaluation datasets from FaceForensics++ and achieve an accuracy of 62.5%.