Face2Face Manipulation Detection Based on Histogram of Oriented Gradients
Amr Megahed, Qi Han · 2020
Nowadays with rapid advances in computer vision and deep learning, it is possible to create highly realistic synthetic faces in digital videos. This situation increases anxiety and suspicion in the video content. It can be a big challenge for humans and machines to differentiate between real and fake faces in a video, especially when the video is compressed or has low resolution. In this paper, an efficient method is proposed for detecting Face2Face manipulations. With the analysis of manipulated videos, we found that there are some visual artifacts that can be exploited to detect fake faces. The proposed method utilizes the histogram of oriented gradients for feature extraction that can be effective in exposing Face2Face manipulations. Experimental results clarify that our method effectively detects the manipulated faces with a high-performance accuracy under various compression quality levels.