Detection of Diffusion-Generated Images Using Sparse Coding

This paper proposes a method for detecting images generated by diffusion models using sparse coding. In the diffusion model, an image can be generated by removing noise from a noisy image. This different generation process from real images leads us to believe that there may be a statistical difference in pixels between the real and the generated images. Specifically, the image is divided into small patch regions, and all patch images are reconstructed using the basis image. In this process, sparse coefficients that contain many zeros are obtained using sparse coding, and features are calculated from the obtained coefficients. Then, a simple discriminator using the features as input is trained with a small number of data to discriminate the diffusion-generated images. In our experiments, we evaluated the proposed method on six datasets created using three diffusion models and two real image datasets. Experiments were also conducted to evaluate the robustness of the proposed method against JPEG compression. Experimental results show that our proposed method is sufficiently robust against JPEG compression with as few as 1800 training data.