DDEM: Deepfake Detection Enhanced Model for Image Forgery Detection Combat Academic Misconduct

Image forgery, as a classic form of academic mis-conduct, has garnered increasing interest from researchers in the field of research integrity. Concurrently, automated detection and localization methods for image forgery have been making steady progress. However, the rapid advancement of image generation and image inpainting methods based on diffusion models presents new threats to the task of image forgery detection. Existing approaches are finding it difficult to identify forgery images that have been spliced by deepfake images and authentic photographs. To tackle this challenge, we propose the Deepfake Detection Enhanced Model (DDEM), which learns image features from both the RGB domain and frequency domain using the HRNet backbone. Furthermore, we introduce diffusion reconstruction error to enhance deepfake detection capabilities in our model. Fi-nally, we conduct experiments on public datasets and a self-made dataset generated by the image inpainting model, demonstrating that our proposed method outperforms the state-of-the-art in terms of image tampering detection and localization.