Image Watermarking with PRC

[Update] The paper is accepted by ICLR 2025!

Paper

If you find this work useful, please consider citing our paper:

@article{gunn2025undetectable,
  title={An undetectable watermark for generative image models},
  author={Gunn, Sam and Zhao, Xuandong and Song, Dawn},
  journal={arXiv preprint arXiv:2410.07369},
  year={2024}
}

PRC watermark is a method similar to Tree-Ring Watermark, where a diffusion model generates images embedded with a watermark as defined by a specified watermark key.

The output of the watermark detection is binary, indicating whether the watermark is detected or not.

We will also add how to embed and decode longer messages with PRC watermark later.

Dependencies

The code is based on python 3.11.9 and the packages specified in requirements.txt.

You can install the dependencies by running:

pip install -r requirements.txt

Usage

You need to specify the number of test images to generate and test on. The example uses 10. The watermark key is randomly generated and saved in the keys folder.

mkdir keys
python encode.py --test_num 10

python decode.py --test_num 10

Alternatively, you can specify a different folder for testing images (e.g., attacked images):

python decode.py --test_num 10 --test_path [path to test images]

You can also change the model and prompt in model_id and dataset_id respectively.

Additionally, you can set the targeted False Positive Rate (FPR) using the fpr parameter. The default value is 0.00001.

Note: Need to change the huggingface cache directory in encode.py and decode.py.

Speedup

To improve decoding speed, the bottleneck lies in the decoder_inv function within PRC-Watermark/src/inverse_stable_diffusion.py.

You can reduce the number of steps to speed up processing. For example, modify:

for i in self.progress_bar(range(100)):

to 20 steps:

for i in self.progress_bar(range(20)):

References

• Treering Watermark
• WAVES
• WatermarkAttacker
• Exact Inversion
• Gaussian Shading