GAN and DM Generated Synthetic Image Detection in the Age of Misinformation

Synthetic images generated by artificial intelligence, particularly those created by Generative Adversarial Networks (GAN) and Diffusion Models (DM), have achieved hyper-realistic quality in recent years, often indistinguishable by the human eye. These fake images are critical in spreading misinformation across Online Social Networks (OSNs), commonly used as profile pictures in fake social media accounts and paired with false news to gain user trust. While many state-of-the-art solutions can identify synthetic images produced by GANs with high accuracy, their effectiveness diminishes when dealing with images circulated on OSNs. Our study finds that selecting the right features, in combination with a deep learning-based classification model, can enhance the performance of synthetic image detectors under challenging conditions. We propose two innovative solutions: a Gradient-based method and a novel Sine Transform Feature-based Network (STN Net). Both methods perform better than existing state-of-the-art solutions for post-processed images, achieving over 99% accuracy in detecting synthetic images and over 91% detection accuracy in challenging scenarios. Additionally, we introduce a transfer learning-based approach for identifying images generated by DMs. This solution not only excels in detecting synthetic images but also demonstrates satisfactory generalization performance.