Research map / Image generation / Autoregressive and latent generation
MuST: Robust Image Watermarking for Multi-Source Tracing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MuST, a deep watermarking scheme that embeds messages in image materials and recovers them from multi-source composite images via a detector, minimum external rectangle resynchronization, and a composing-simulation noise layer.
Problem
Tracing the copyright of multiple image materials reused in composite images is not handled by existing single-image watermarking schemes, which fail under extreme downsizing, irregular cropping, repositioning, and fusion edits.
Approach
Embed watermarks into the primary component (MER crop) of each image material, simulate multi-source composing as a differentiable noise layer, and use a U-Net detector to segment each reused material so a decoder can extract each material's message independently.
Key results
- MuST achieves 42.89 dB PSNR and 0.9883 SSIM on average, the highest among compared methods (SSLWM, CIN, LGDR).
- Under the full multi-source composing simulation, MuST reaches 97.92% average extraction accuracy versus 89.28% for LGDR and 85.89% for CIN.
- Detector TPR is 100% for 1-4 materials and IoU above 99.6% across 1-6 materials; real-world Photoshop, Canva and Pixso composites reach 95-100% accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MuST instead of SSLWM, CIN, or LGDR for watermarking | Image materials from SOIM, CASIA V2.0, and ICCV09 · SOIM, CASIA V2.0, and ICCV09 (556 images) | PSNR (dB): 35.12 -> 42.89; 42.56 -> 42.89; 42.10 -> 42.89 | Helped |
| Use MuST instead of SSLWM, CIN, or LGDR for watermarking | Image materials from SOIM, CASIA V2.0, and ICCV09 · SOIM, CASIA V2.0, and ICCV09 (556 images) | SSIM: 0.9571 -> 0.9883; 0.9792 -> 0.9883; 0.9854 -> 0.9883 | Helped |
| Use MuST instead of SSLWM, CIN, or LGDR | Image materials under MIC · SOIM under MIC | ACC (%): 75.25 -> 97.92; 85.89 -> 97.92; 89.28 -> 97.92 | Helped |
| Add MER resynchronization (ablation) | MuST · SOIM under MIC | ACC (%): 54.43 -> 98.25 | Helped |
| Add the noise layer during training (ablation) | MuST with MER · SOIM under MIC | ACC (%): 86.37 -> 98.25 | Helped |
Surprising findings
- LGDR and CIN performed unusually well when watermarked images were downsized to 50%, which the authors attribute to the sharper watermark component produced by upsampling-based encoding.
Limitations
- Performance drops when multiple interconnected watermarked materials are present, since automatic localization and segmentation becomes harder.
- Real-world evaluation is basic because complex designer aesthetics are hard to simulate with scripts.
- Robustness to JPEG compression and color changes is not evaluated.
Experimental setup
Datasets: SOIM (constructed from JD Product-10k, Google Scanned-Objects, GroceryStore-Dataset) CASIA V2.0 Stanford Background (ICCV09)
Benchmarks: SOIM CASIA V2.0 ICCV09
Metrics: PSNR SSIM extraction accuracy (ACC) TPR IoU
Baselines: SSLWM CIN LGDR
Builds on: HiDDeN StegaStamp CIN LGDR MBRS
Closest papers by meaning
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.85)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.85)
- On the Coexistence and Ensembling of Watermarks (NeurIPS 2025 · similarity 0.85)
- Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft (CVPR 2025 · similarity 0.84)
- Safe-Sora: Safe Text-to-Video Generation via Graphical Watermarking (NeurIPS 2025 · similarity 0.84)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.