Research map / Vision-language and multimodal / Human and embodied modeling
ScreenMark: Watermarking Arbitrary Visual Content on Screen
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 ScreenMark, a learning-based watermark that diffuses bit messages into irregular screen-overlay patterns blended via pre-multiplied alpha, trained in three progressive stages, to protect arbitrary visual screen content against screenshots.
Problem
Watermarking for visual screen content (VSC) displayed on computer screens, which is multi-modal and dynamic and vulnerable to screenshot leakage; existing watermarks target single media types and cannot respond in real time to screen changes.
Approach
Instead of encoding messages into a carrier image, a message diffuser turns a 100-bit message into an irregular watermark pattern that is alpha-blended as a floating mask over any screen content; a screen decoder recovers it from screenshots, trained in three progressive stages with image-, pixel- and joint-level distorters.
Key results
- ScreenMark reaches 98.82% average bit accuracy under image-level attacks on ScreenImage versus 96.37% for the best baseline DWSF.
- Under pixel-level attacks ScreenMark averages 98.19% BAR on ScreenImage and 97.39% on ImageNet, the best or near-best among compared methods.
- ScreenMark achieves PSNR 41.86 dB, SSIM 0.9948 and LPIPS 0.0055 on ScreenImage, with reaction time 0 ms versus 14.91-24.91 ms for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ScreenMark instead of DWSF | ImageNet image-level watermarking · ImageNet, image-level attacks | BAR average (%): 96.60 -> 96.80 | Helped |
| Use ScreenMark instead of DWSF | ScreenImage image-level watermarking · ScreenImage, image-level attacks | BAR average (%): 96.37 -> 98.82 | Helped |
| Use ScreenMark instead of DWSF | ImageNet pixel-level watermarking · ImageNet, pixel-level attacks | BAR average (%): 97.50 -> 97.39 | Hurt |
| Use ScreenMark instead of MBRS | ScreenImage pixel-level watermarking · ScreenImage, pixel-level attacks | BAR average (%): 97.36 -> 98.19 | Helped |
| Use ScreenMark instead of DWSF | ImageNet visual watermark quality · ImageNet | PSNR: 41.47 -> 41.38 | Hurt |
| Use ScreenMark instead of PIMoG | ImageNet visual watermark quality · ImageNet | SSIM: 0.9850 -> 0.9969 | Helped |
| Use ScreenMark instead of DWSF | ImageNet visual watermark quality · ImageNet | LPIPS: 0.0083 -> 0.0058 | Helped |
| Use ScreenMark instead of DWSF | ScreenImage visual watermark quality · ScreenImage | PSNR: 42.60 -> 41.86 | Hurt |
| Use ScreenMark instead of DWSF | ScreenImage visual watermark quality · ScreenImage | SSIM: 0.9905 -> 0.9948 | Helped |
| Use ScreenMark instead of DWSF | ScreenImage visual watermark quality · ScreenImage | LPIPS: 0.0074 -> 0.0055 | Helped |
| Use ScreenMark instead of single-modal watermark methods | VSC protection when screen content changes · VSC reaction time | reaction time (ms): 14.91 -> 0 | Helped |
| Use three-stage progressive training instead of end-to-end training (ablation) | ScreenMark · Training convergence | convergence behavior | Helped |
Surprising findings
- On ImageNet, DWSF achieves a higher average BAR under pixel-level attacks (97.50%) than ScreenMark (97.39%), and DWSF also reports higher PSNR (41.47 vs 41.38 dB) on ImageNet.
- Under the Cropout 20% setting on ScreenImage, ScreenMark's BAR (96.23%) is only moderately higher than baselines, and under Noise 0.1 on ImageNet it trails MBRS (95.23 vs 97.88%).
Limitations
- Baselines are evaluated at fixed input sizes (MBRS requires 128x128 scaling), so comparison conditions are not fully matched.
- The reported reaction time of 0 ms is a conceptual property of overlay-based watermarking rather than a measured system benchmark.
- Real-screenshot evaluation is only reported in the appendix, and the paper does not release the dataset or code.
Experimental setup
Datasets: ScreenImage ImageNet
Benchmarks: ImageNet ScreenImage
Metrics: bit accuracy rate PSNR SSIM LPIPS reaction time
Baselines: StegaStamp MBRS PIMoG DWSF
Builds on: StegaStamp HiDDeN MBRS PIMoG DWSF
Closest papers by meaning
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.87)
- RoPaSS: Robust Watermarking for Partial Screen-Shooting Scenarios (AAAI 2025 · similarity 0.86)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.85)
- Safe-Sora: Safe Text-to-Video Generation via Graphical Watermarking (NeurIPS 2025 · similarity 0.85)
- WMarkGPT: Watermarked Image Understanding via Multimodal Large Language Models (ICML 2025 · similarity 0.85)
- Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft (CVPR 2025 · similarity 0.84)
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.