Research map / Image generation / Autoregressive and latent generation
Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft
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 WOFA, a multi-level encoder/embedder-distortion-extractor/decoder image watermarking scheme with a two-stage, progressive distortion-layer training that stays robust to partial image theft (partial masking, geometric transforms, background fusion).
Problem
Existing image watermarking schemes require the entire image to extract the watermark and fail when a malicious user steals only an arbitrarily shaped portion, fuses it with a new background, and possibly applies geometric transformations.
Approach
Embed a watermark once into the full image and train a network with a distortion layer that simulates the full partial-image-theft pipeline (partial masking, translation/rotation/scaling, background fusion), using a two-stage training scheme and progressive distortion intensity to aid convergence.
Key results
- Under partial masking with 0-10% area, WOFA reaches 89.33% bit accuracy while MBRS reaches 85.96% and HiDDeN 51.75%.
- With partial masking plus translation, WOFA keeps over 90% bit accuracy in most settings (e.g., 91.97% at 10% translation), while MBRS drops to about 56%.
- WOFA achieves PSNR 39.76 dB, SSIM 0.9945 and LPIPS 0.0188 on 200x200 images with 30-bit capacity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WOFA instead of the strongest listed baseline | WOFA · Partial image theft with mixed translation, rotation, and scaling | BAR: 52.59 -> 87.06 | Helped |
| Use WOFA instead of the strongest listed baseline | WOFA · Partial image theft with partial masking and no geometric transformation | BAR: 74.91 -> 95.92 | Helped |
| Train with a progressive distortion layer instead of a static strategy (ablation) | WOFA · Training convergence | extraction accuracy | Helped |
| Use two-stage training instead of end-to-end training (ablation) | WOFA · Training convergence | extraction accuracy | Helped |
Surprising findings
- The end-to-end training scheme collapses early (around 10% of training) to nearly blank outputs, so the two-stage approach was needed for convergence.
- Shuffling image blocks (position change) degrades extraction far more than erasing blocks (content loss), suggesting position information matters more than content for watermark extraction.
Limitations
- Baseline performance on clean images is lower than some methods (e.g., WOFA clean BAR 98.39 vs MBRS 100 under partial masking), and extraction under geometric transforms drops to around 87% in some settings.
- matteImageNet dataset is small (977 images from 19 classes) and not released.
- Evaluated only at 200x200 crops from OPA and on fixed-size inputs.
Experimental setup
Datasets: OPA ImageNet SOIM
Benchmarks: OPA SOIM matteImageNet
Metrics: bit accuracy rate PSNR SSIM LPIPS
Baselines: HiDDeN DWSF MBRS PIMoG StegaStamp MuST
Builds on: HiDDeN MBRS StegaStamp PIMoG MuST
Closest papers by meaning
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.86)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.86)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.86)
- RoPaSS: Robust Watermarking for Partial Screen-Shooting Scenarios (AAAI 2025 · similarity 0.85)
- Watermark Stealing in Large Language Models (ICML 2024 · similarity 0.85)
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.85)
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.