GAN Generated Portraits Detection Using Modified VGG-16 and EfficientNet

Abstract

Generative Adversarial Networks can generate deceptive portraits of people who do not exist. The misuse of this technology leads to severe security issues such as fake identities and credentials. Since 2017, many works on Deep Learning have focused on detecting GAN synthesized images to prevent the threat against credibility in media. In this work, the authors propose a lightweight VGG-like model to detect state-of-the-art StyleGAN generated portraits. The authors also adopt EfficientNet-B0 to train a classifier on the same StyleGAN architecture. The VGG-like model and EfficientNet-based model achieve 98.9% and 100%, respectively, on the StyleGAN dataset published by Nvidia in 2019. Both models show the potential in generalizing to other GAN architectures and synthetic technologies.