<p>In recent years, malicious exploitation of Deepfakes technology has occurred frequently, posing a serious threat to society. Although many post-facto detection methods have been developed for Deepfakes, these passive forensic techniques do not take any preventive measures on the original face images before tampering occurs. To bridge this gap and improve the forensic ecosystem, we propose a forward-looking solution called Multi-Domain Decoupled Watermarking (MDD-Watermark), which aims to provide a unified framework for source tracking and Deepfake detection. MDD-Watermark is constructed by multi-domain decoupling of the original image; when the image is forged, the image reconstructed based on the decoupling information of the original image will be significantly different from the forged image in terms of features. This difference can be quantitatively analyzed using traditional image evaluation metrics (e.g., PSNR, SSIM). We also design a deep learning-based framework, XUNet. It can efficiently embed the MDD-Watermark into the carrier image and still stably extract the watermark information in the face of multiple perturbations (e.g., noise, compression, rotation, etc.). Experimental results demonstrate that while maintaining high visual quality, the proposed method not only effectively resists deepfake attacks and preserves watermark robustness, but also enables significant stratification in image quality metrics such as PSNR when comparing watermarked images with forged images against their respective reconstructed counterparts.</p>