Link to full Extended Abstract
The paper explores several state-of-the-art models for detecting facial forgery (i.e. “deepfakes”) in videos and images, and uses several custom metrics to ascertain how well they generalize to real-world environments. Our dataset is an custom-made extension of FaceForensics++, for which additional sets of facial forgery videos were generated for five facial forgery models (Deepfakes, FaceSwap, GANnotation, ICface, and X2Face). We concluded that while these models generalize well to different levels of compression and can detect many different forgery techniques at once, the state-of-the-art is unable to generalize against unknown facial forgery techniques.
Code can be found here.
Research was conducted as part of the Pervasive Computing for Smart Health, Safety, and Well-being Research Experience for Undergraduates (REU) at Temple University under Professor Jie Wu.