Robust Noise-Based Provenance Tracing for Diffusion Images via Latent Geometric Alignment
Year: 2026
Type: conference-paper
Abstract: Diffusion models enable high-quality image generation but also raise concerns about copyright protection and content provenance. Existing watermarking methods for diffusion images often rely on structured watermark patterns or auxiliary identifiers, which may disturb the latent prior, limit key capacity, or remain vulnerable to geometric distortions. We propose a noise-based provenance tracing method that directly uses the initial Gaussian noise as the identity carrier, avoiding explicit watermark perturbations. For detection, a query image is inverted into the latent space and matched with a keyed reference noise set. To handle geometric distortions, we introduce a latent geometric transformation prediction network that estimates affine alignment for corrected latent matching. Experiments show that our method achieves robust user-level identification under both geometric and non-geometric attacks while preserving high image quality.
Institutions: Guizhou Normal University, Shanghai University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202384573