ENHANCING TRANSFORMER-BASED DETECTION OF AI-GENERATED IMAGES WITH SPECTRAL FEATURE FUSION

Abstract

With the dramatic increase in the number of realistic generative models like GANs, Diffusion Models, and more, AI-based image detection has taken a larger role in protecting the quality of digital images, as well as their usage in many applications, such as Digital Forensics, Cybersecurity, and Moderation on Social Media. The emergence of these new methods of creating high-quality images through various algorithms has created a demand to build a system capable of detecting synthetic images from authentic images. In this research, we have developed a novel hybrid detection solution called ViT FF – a hybrid detection framework that combines a Vision Transformer (ViT) with the frequency domain feature fusion method for faster and more accurate image detection capabilities. By leveraging both spatial representations captured by transformers and high-frequency patterns derived from spectral analysis, ViTFF surpasses traditional convolutional neural network-based approaches and exhibits strong generalization across diverse GAN and diffusion model outputs. Experimental results demonstrate that the proposed architecture successfully captures subtle generative artifacts that are frequently overlooked by conventional CNN detectors achieving 99.9% accuracy. These findings suggest that ViTFF is a robust and generalizable solution for the detection of AI-generated images.

Keywords : AI-generated images, GAN model, Diffusion model, Vision Transformer, Frequency domain analysis, Digital forensic