부산대학교 도서관

Recently, diffusion models have shown remarkable success in generating high-quality images, making them potentially more difficult t o d etect t han G AN-generated i mages. In this paper, we address the emerging challenge of detecting face forgeries generated by diffusion models. We leverage insights from frequency artifacts in GAN-generated images and delve into the frequency domain characteristics of diffusion-generated images using both discrete Fourier transform (DFT) and Haar discrete wavelet transform (HDWT). Our investigation reveals that, while diffusion-generated images lack obvious DFT artifacts like those in GAN-generated images, they exhibit fewer high-frequency details compared to real images. Building upon these observations, our multi-scale network incorporates wavelet lifting and wavelet-spatial transformer blocks, enabling precise frequency decomposition and efficient f eature f usion. Experimental results on two generated datasets demonstrate the superior robustness of our proposed method compared to state-of-the-art forgery detection methods, making it an effective solution for identifying face forgeries produced by diffusion models. The dataset and code are open source at https://github.com/eecoder-dyf/Diffusion-Detection-WCSP2023.