Super-resolution Reconstruction Detection Method for DeepFake Hard Compressed Videos
Lei Sun, Hongmeng Zhang, Xiuqing Mao, Song Guo, Yongjin Hu · 2021
The forensics methods of DeepFake video generally use convolution neural networks. However, these methods perform poorly on hard compressed DeepFake datasets and make a large number of false detections on real data. To solve the problem above, a method of hard compressed DeepFake video detection based on deep neural network model is proposed, which improves the detection accuracy of hard compressed video by incorporating super-resolution reconstruction technology and recovering the loss of the spatial and temporal information during hard compression. Experiments are performed with the FaceForensics++ Datasets and DFDC (the DeepFake Detection Challenge) Datasets for hard compressed DeepFake video, which improved the test accuracy of single frame and video compared to ResNet50, and effectively alleviated the problem of false detection of real video with hard compression.