Study on Face Landmark-based Analysis for Synthetic Media Identification Generated by Adversarial Generative Networks
Akinobu Ura, Minoru Kuribayashi, Nobuo Funabiki · 2023
With the development of deep learning technology, artificially created images and videos, known as deepfakes, have become a problem. It is very difficult for human visual and auditory senses to distinguish them from real ones. Therefore, there is a growing risk of them being used for crimes such as bullying, defamation, extortion, and fraud on social media. StyleGAN is one of the generators for such fake content and is known for its ability to generate high-quality images with a great degree of realism. This study aims to distinguish between real and fake images generated by StyleGAN. We introduce effective image processing techniques into the detector which is CNN model designed by fine-tuning the XceptionNet. We examined two methods: one that extracts color attributes from the image and the other that distinguishes based on landmarks. We evaluated the accuracy of distinguishing between real and fake images by computer simulation.