DASMark: Dual-Anchor Synchronization of Diffusion Image Watermarks for Geometrically Robust Deepfake Detection

Abstract Diffusion models can generate highly realistic images, raising concerns about deepfakes. Watermarking supports active deepfake detection through model verification and user attribution. However, geometric attacks can disrupt watermark synchronization and impair detection reliability. Existing methods remain vulnerable to such attacks, particularly in large-scale user identification. We propose DASMark, a diffusion image watermarking framework based on dual-anchor geometric synchronization. Reproducible user-specific initial noise carries identity information, while two spatially separated anchors encoded by orthogonal channel codes provide geometric references. A localization network predicts their transformed positions, enabling latent geometric alignment for robust model verification and scalable user identification. Experiments show that DASMark maintains robust synchronization across arbitrary rotation angles and achieves an average verification AUC of 0.999. With 200,000 candidate identities, it achieves 99.7\% identification average accuracy and outperforms WIND by 11.0 and 18.6 percentage points under rotation and crop-and-stretch, respectively. DASMark also preserves image quality, achieves lower end-to-end runtime and GPU memory usage than WIND, and exhibits cross-model compatibility.