Generative Forgery Image Detection Based on Diffusion Reconstruction Loss and Frequency Domain Feature

Since the advent of generative models such as Generative Adversarial Networks (GANs), image generation techniques have made significant progress in the field of digital image processing. However, the development of these techniques has also brought about challenges in research integrity, especially the growing problem of forged images in research papers. Currently, there are no mature research results related to the generation of forged image detection in the field of research integrity detection. Existing detection techniques are difficult to adapt to the rapidly evolving image generation models, especially when dealing with research paper images that have a single structure and are quickly learned by the generation models. Existing methods fail to effectively identify high-quality forged images produced by advanced generative models, which poses a severe threat to scientific integrity. In this study, we propose a new detection method that combines diffusion reconfiguration loss and frequency-domain features to identify and differentiate generative forged images from fundamental scientific research images. We developed a technique to detect frequency-domain artifacts specific to generated images using frequency-domain analysis and combined it with image reconstruction error features of the diffusion model to enhance the accuracy and robustness of detection. Through testing on a newly constructed public dataset, our method shows superior performance in effectively identifying forged images produced by multiple generative models.