PixMSE: detecting GAN-generated images through local roughness

Abstract Realistic images generated by GANs (generative adversarial networks) have enriched people’s lives but also pose serious threats to personal privacy and society. Therefore, it has become essential to develop methods for accurately detecting GAN-generated images. Existing methods have utilized artifacts to detect GAN-generated images, but the artifacts in different generated images can vary significantly. As a result, these algorithms often struggle to detect generated images whose artifacts differ substantially from those in the training set. This paper proposes the PixMSE algorithm based on statistical features, which designs the MSENet1 process to obtain local roughness feature maps of the image and the MSENet2 process to obtain local roughness feature maps of the filtered image. Finally, PixMSE uses features extracted by ResNet (Residual Network) to detect the generated images. Experimental results show that PixMSE achieves an average detection accuracy of 84.3% on eight subsets of the Wang dataset, which is a publicly available dataset, demonstrating strong cross-model generalization performance. The source code and datasets are available at https://github.com/DarlingDiving/PixMSE.