🔓 Can Simple Averaging Defeat Modern Watermarks? 🤔

Pei Yang^* , Hai Ci^* , Yiren Song , and Mike Zheng Shou^✉

Show Lab, National University of Singapore

🧰 Security Guidelines

We call for future watermarking research to benchmark their methods against simple steganalysis using our provided code. Getting started:

Dependencies

pip install numpy<2 Pillow matplotlib tqdm

Benchmark Code Usage

python benchmark.py \
    --watermark_method RingID \
    --width 512 \
    --height 512 \
    --ood_clean_path ~/Datasets/ImageNet/test \
    --ind_clean_path ~/Datasets/RingID/clean \
    --watermarked_path ~/Datasets/RingID/watermarked \
    --output_path /path/to/save/images \
    --num_eval_images 100

ind_clean_path and watermarked_path are paired paths to non-watermarked and watermarked images. If 123456.png is in ind_clean_path, then its watermarked pixel-aligned counterpart should present in watermarked_path with exactly the same filename 123456.png.

ood_clean_path/              # 5000+ original images from another dataset
├── 000000.png
├── 000001.png
├── 000002.png
└── ...

ind_clean_path/              # 5000+ paired non-watermarked images
├── ringid_0000.png
├── ringid_9801.png          # File names should precisely match, images should be pixel-aligned
└── ...

watermarked_path/            # 5000+ paired watermarked images
├── ringid_0000.png
├── ringid_9801.png          # File names should precisely match, images should be pixel-aligned
└── ...

🔑 Patterns Averaged Out

🚀 Open-Source

• Core benchmark code for watermark removal/forgery 🧰
• Images we used during experiments
• Complete experiment code (currently being organised; unpolished version available on request: contact yangpei@comp.nus.edu.sg for access)

Citation

@misc{yang2024steganalysisdigitalwatermarkingdefense,
      title={Steganalysis on Digital Watermarking: Is Your Defense Truly Impervious?}, 
      author={Pei Yang and Hai Ci and Yiren Song and Mike Zheng Shou},
      year={2024},
      eprint={2406.09026},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.09026}, 
}