Intellectual Property Protection of Diffusion Models via the Watermark Diffusion Process

This repository is the official implementation of the Intellectual Property Protection of Diffusion Models via the Watermark Diffusion Process. It is based on the parent repository improved-diffusion and guided-diffusion.

1. Requirements

We use Mamba to manage the environment requirements.

mamba create -n wdm python=3.10
mamba activate wdm
mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
mamba install tensorflow-gpu matplotlib blobfile mpi4py tqdm PyYAML pandas

and intall the package:

pip install -e .

2. Prepare Datasets

To prepare for the task and watermark dataset, run the following:

2.1 CIFAR-10 as the Task Dataset

python cifar10.py

2.2 sWDM Setting

python single_wm.py

2.2 mWDM Setting (class five in MNIST)

python multiple_wm.py

3. Baseline Model Training

To train the baseline models in the paper, configure configs/train_mse.yaml and run this command:

python scripts/image_train.py --p configs/train_mse.yaml

or multi-GPU version (replace $NUM_GPUS)

mpiexec -n $NUM_GPUS python scripts/image_train.py --p configs/train_mse.yaml

4. Watermark Embedding

To embed the watermark data into the baseline models, configure configs/train_mse_wdp.yaml and run this command:

python scripts/image_train.py --p configs/train_mse_wdp.yaml

5. Task Sampling & Watermark Extraction

To sample from the task, configure the configs/sample_mse.yaml run this command:

python scripts/image_sample.py --p configs/sample_mse.yaml

To extract the watermark from the suspected model, configure the configs/sample_mse_wdp.yaml run this command:

python scripts/image_sample.py --p configs/sample_mse_wdp.yaml

6. Evaluation

The evaluation is based on the guided diffusion repo.

python scripts/evaluator.py <reference_batch> <evaluated_batch>

7. Citation

@inproceedings{peng2025intellectual,
  title={Intellectual property protection of diffusion models via the watermark diffusion process},
  author={Peng, Sen and Chen, Yufei and Wang, Cong and Jia, Xiaohua},
  booktitle={International Conference on Web Information Systems Engineering},
  pages={290--305},
  year={2025},
  organization={Springer}
}