ASBW: A Frequency-Domain Analysis Approach for Distinguishing GAN-Generated Images from Real Images

Abstract

The rapid advancement of Generative Adversarial Networks (GANs) poses significant challenges to digital media integrity, as traditional spatial-domain forensics often struggle with overfitting and sensitivity to perturbations. To address this, we propose the Adaptive Spectral-Band Weighting (ASBW) layer, a modular frequency-domain component that utilizes the Discrete Fourier Transform (DFT) and a lightweight Multi-Layer Perceptron (MLP) to dynamically amplify discriminative spectral fingerprints typical of GAN upsampling. Experimental evaluations on a dataset of 140,000 real and fake faces demonstrate that integrating ASBW into a ResNet-18 backbone achieves a detection accuracy of 98.72% and an F1-Score of 98.71%, outperforming the baseline by 1.11%. These results confirm that spectral attention effectively captures subtle generative traces overlooked by spatial convolutions, offering a robust and efficient solution for image forensics.