Detection of GAN generated image using color gradient representation

Yun Liu, Zuliang Wan, Xiaohua Yin, Guanghui Yue, Aiping Tan, Zhi Zheng

Abstract: Highlights•Considering the inherent trace left in color information, we propose a novel GAN image detection method based on color gradient analysis.•By analysis the gradient information and the directional texture information of the generated images, two directional gradient texture features (the horizontal and the vertical gradient channels) are chosen to describe the inherent trace in GAN image, which makes our model more interpretable.•We propose a novel gradient-domain local direction number (GLDN) pattern descriptor to extract the gradient texture features. GLDN enhance the intensity change and structural information of the local information, which can effectively grasp the inherent trace of the generated images.•We apply the Multi-GLDN histogram (MGLH) to quantify the gradient texture information of each GLDN map block and aggregates the classification information between natural images and GAN generated images into texture descriptors, which can reduce computational complexity and obtain higher accuracy.