🚀 Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models

📌 Paper: Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Model
📖 Conference: CVPR 2025

📂 Project Structure

Dataset Preparation

CoprGuard 
└── Source/ 
    ├── FFHQ/ 
    ├── CelebA-HQ/ 
    ├── BigGAN/ 
    ├── ... 
    └── Vggface/ 
└── HiNet/ 
    ├── FFHQ/ 
    ├── CelebA-HQ/ 
    ├── BigGAN/ 
    ├── ... 
    └── Vggface/ 
└── Generated/ 
    ├── FFHQ_DDIM/ 
    ├── BigGAN_DDIM/ 
    ├── ... 
    └── FFHQ_HiNet_DDIM/

🚀 Installation

# Clone the repository
git https://github.com/sccsok/CoprGuard.git
cd your-repo

# Create a virtual environment (optional) & install dependencies
conda env create -f environment.yml

🏋️‍♂️ Training & Evaluation

🖼️ Image Watermarking

• Watermarking

  Download the pretrained HiNet and put it in ~/CoprGuard/watermark/ckpt.

  Watermarking training images：

  # For unconditional training images
  python wm.py --root_dir <> --watermark_path <> --save_dir <>

• Generate Figure 6

  Compute consine similarity scores:

  python get_cos.py

  Plot cosine similarity distribution:

  python plt.py

📌Unconditional Training & Evaluation

• Prepare Dataset
  Please prepare the training dataset according to the Dataset Preparation format.

• DDPM Train

  Run the training scripts under ddim/scripts:

  cd ~/CoprGuard/ddim
  python scripts/xxx.py

• Image Sampling

  Use the DDPM or DDIM scheduler for image sampling:

  python ddim/scripts/inference.py

• Classifier-Free Training

  The ddim folder can be easily modified to support Classifier-Free Guidance. You can also refer to the following repository: classifier-free-diffusion-guidance-Pytorch

📊 Generate Figures

• Generate Figure 1 & Figure 2 & Figure 10

  cd ~/CoprGuard/frequence
  python frequency_analysis.py ~/CoprGuard/Source $WORKDIR/output <fft_hp/dct/...> --img-dirs <FFHQ BIgGAN ProGAN ImageNet> --log --vmin 1e-5 --vmax 1e-1
  python frequency_analysis.py ~/CoprGuard/Generated $WORKDIR/output <fft_hp/dct/...> --img-dirs <FFHQ BIgGAN ProGAN ImageNet> ImageNet_DDIM --log --vmin 1e-5 --vmax 1e-1

• Generate Table 1

  cd ~/CoprGuard/frequence
  python get_cos.py --type <fft/dct/...> --models1 <[FFHQ, xxx, ...]> --models2 <[FFHQ_DDIM, xxx, ...]> 

• Generate Figure 3

  cd ~/CoprGuard/frequence
  get_rapsd.ipynb

• Generate Table 3

  cd ~/CoprGuard/watermark
  python similarity_compute.py --mode folder --folder <> --watermark <> --resize 128 128

📜 Citation

If you find our work useful, please consider citing:

@inproceedings{liu2025harnessing,
  title={Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models},
  author={Liu, Zhenguang and Shuai, Chao and Fan, Shaojing and Dong, Ziping and Hu, Jinwu and Ba, Zhongjie and Ren, Kui},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={18653--18662},
  year={2025}
}

🙏 Acknowledgement

Our implementation benefits from the following open-source projects:

• HiNet
• Diffusion Autoencoders
• Diffusion-model-deepfake-detection

We sincerely thank the authors for their great work.

---