Research map / Adversarial robustness and security / Attacks on text and graphs
Embedding Robust Watermarking into Pattern to Protect the Copyright of Ceramic Artifacts
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 an invisible, template-based watermarking framework embedded in ceramic patterns, with a ceramic-tailored distortion layer (color-halftoning, adaptive brightness) and a ResNet18 decoder to extract the copyright message after decal-firing and camera capture.
Problem
Ceramic artworks are forged because copyright stamps are visible, affect appearance, and are easy to copy; existing physical image watermarks fail to generalize to the distortions of the ceramic decal manufacturing process (printing and kiln firing).
Approach
Embed BCH-coded, key-shuffled watermark bits as a Gaussian-circle template added to the pattern (no trained encoder), and train a decoder against a distortion layer that explicitly simulates ceramic printing (color-halftoning via GCR and Bayer dithering) and firing (adaptive brightness adjustment) plus camera-capture distortions.
Key results
- Proposed method reaches 97.51% bit accuracy vs. 87.50% for the second-best (LIM), roughly a 10% improvement, at PSNR 22.79 dB.
- Bit accuracy stays above 90% at shooting distances of 15-30 cm, and across two phone models (Vivo 97.51%, iPhone 95.39%).
- Ablation: color-halftoning plus brightness adjustment raises bit accuracy from 92.65% to 97.51%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add color-halftoning and adaptive brightness adjustment to the distortion layer (both together) (ablation) | ceramic pattern-tailored distortion layer · decal-shooting tiles (ablation) | Bit ACC (%): +4.86 points (92.65 to 97.51) | Helped |
| add color-halftoning alone to the distortion layer (ablation) | ceramic pattern-tailored distortion layer · decal-shooting tiles (ablation) | Bit ACC (%): +1.41 points (92.65 to 94.06) | Helped |
| add adaptive brightness adjustment alone to the distortion layer (ablation) | ceramic pattern-tailored distortion layer · decal-shooting tiles (ablation) | Bit ACC (%): +1.25 points (92.65 to 93.90) | Helped |
| increase template embedding strength alpha from 0.25 to 0.5 with beta 0.25 (stronger template) (ablation) | template-based watermark embedding · decal-shooting tiles (ablation) | PSNR (dB) vs Bit ACC (%): Bit ACC 97.51 to 96.50 at 0.25/0.5 while PSNR drops 22.51 to 21.13 | Mixed |
| use weaker template strength (alpha 0.15, beta 0.5) instead of default (0.25, 0.4) (ablation) | template-based watermark embedding · decal-shooting tiles (ablation) | Bit ACC (%): 86.25 vs 97.51 | Hurt |
| proposed method vs. best baseline LIM on bit accuracy | full framework · decal-shooting tiles, 15 cm | Bit ACC (%): +10.01 points over LIM (87.50 to 97.51) | Helped |
| proposed method vs. StegaStamp and LIM on PSNR/SSIM of watermarked decal images | full framework · decal-shooting tiles (visual quality) | PSNR (dB): 22.79 vs 26.77 for LIM (about 4 dB lower) | Hurt |
| proposed method under longer shooting distance (30 cm vs 15 cm) | full framework · decal-shooting tiles, shooting distance | Bit ACC (%): 97.51 at 15 cm to 91.65 at 30 cm | Hurt |
| proposed method across different phones (Vivo vs iPhone) | full framework · decal-shooting tiles, device generalization | Bit ACC (%): 97.51 vs 95.39 | No clear effect |
| tilted shooting angle (15 to 45 degrees) | full framework · decal-shooting tiles, shooting angle | Bit ACC (%): 86.56 at 15 degrees to 82.50 at 45 degrees | Hurt |
Surprising findings
- The method has lower PSNR/SSIM than the baselines (22.79 dB vs 26.77 dB for LIM), trading visual quality for robustness.
Limitations
- Evaluation uses only 10 real porcelain samples per method and flat tiles with natural images, a simple setting.
- Bit accuracy drops with larger shooting angles (about 82-87%) and longer distances.
- Visual quality (PSNR/SSIM) is lower than comparison methods.
Experimental setup
Datasets: COCO MIRFlickr25k
Benchmarks: MIRFlickr25k patterns decaled onto 10 real porcelain tiles per method
Metrics: PSNR SSIM Bit ACC
Baselines: Fang et al. 2018 StegaStamp LIM
Builds on: HiDDeN StegaStamp LIM
Closest papers by meaning
- IWRN:A Robust Blind Watermarking Method for Artwork Image Copyright Protection Against Noise Attack (AAAI 2025 · similarity 0.84)
- Can Simple Averaging Defeat Modern Watermarks? (NeurIPS 2024 · similarity 0.83)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.83)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.83)
- Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft (CVPR 2025 · similarity 0.83)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.82)
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.