STN-Net: A Robust GAN-Generated Face Detector

Abstract

Massive advancements in Generative Artificial Intelligence in the recent years, have introduced hyper-realistic fake multimedia content. Where such technologies have become a boon to industries such as entertainment and gaming, malicious uses of the same in disseminating fabricated information eventually have invited serious social perils. Generative Adversarial Network (GAN) generated images, especially non-existent human facial images, lately have widely been used to disseminate propaganda and fake news in Online Social Networks (OSNs), by creating fake OSN profiles. Being visually indistinguishable from authentic images, GAN-generated image detection has become a massive challenge to the forensic community. Even though countermeasure solutions based on various Machine Learning (ML) and Deep Learning (DL) techniques have been proposed recently, most of their performance drops significantly for OSN-compressed images. Also, DL solutions based on Convolutional Neural Networks (CNN) tend to be highly complex and time-consuming for training.

This work proposes a solution to these problems by introducing STN-Net, a CNN classifier with an extremely reduced set of parameters, which adopts a carefully crafted minimal image feature set, computed based on Sine Transformed Noise (STN). Despite having a much-reduced feature set compared to other State-of-the-Art (SOTA) CNN-based solutions, our model achieves very high detection accuracy ($average \ge 99\%$). It also achieves promising detection performance on post-processed images, which mimic real-world OSN contexts.