Passive Detection of GAN-Generated Images: A Structured Investigation of Spatial and Frequency-Domain Approaches

Abstract

The rapid advancements in generative adversarial networks (GANs) have led to the production of highly realistic synthetic images, posing severe threats to the credibility and authenticity of digital media across social platforms, news outlets, and official documents. Passive detection methods tackle this problem by identifying subtle intrinsic artifacts and patterns embedded during the image generation process. This paper presents a comprehensive review of existing detection algorithms for GANgenerated fake images, classifying them into spatial-domain and frequency-domain strategies. Spatial-domain approaches are subdivided into handcrafted feature extraction methods and deep learning-based models, whereas frequency-domain methods concentrate on detecting spectral inconsistencies introduced by convolution, upsampling, and normalization operations. The review systematically summarizes representative studies published in top-tier conferences and journals, illustrating the paradigm shift from traditional statistical analysis to modern data-driven deep learning frameworks. Furthermore, it addresses critical challenges such as weak robustness against common post-processing operations and limited generalization across diverse GAN architectures and proposes promising future directions to enhance the performance and reliability of passive image forensics.