Diffusion-PID-Protection

Implementation of the ICML 2024 paper titled "PID: Prompt-Independent Data Protection Against Latent Diffusion Models" by Ang Li*, Yichuan Mo, Mingjie Li, Yisen Wang. Our paper is available at http://arxiv.org/abs/2406.15305.

TODO:

• Initialize the repo. (2024/5/28)
• Training scripts & implementation of the PID.
• Implementation of the evaluation code.
• Visualizations & training data.
• [] Implementation of the baselines (FSGM, ASPL, AdvDM).

Protecting images with PID

The minimum implementation of the PID is given in PID.py. Besides, we place clean images from the CelebA-HQ dataset in ./data/clean_images and the images already protected by PID in ./data/PID_images

  sh PID.sh

Fine-tuning

Fine-tuning can be started with a one-line command. Feel free to experiment with different training configurations.

  sh train_dreambooth.sh # DreamBooth

  sh train_drembooth_lora.sh # LoRA

Evaluation

Implementation of the metrics used in the paper can be found in evaluate.py.

  sh evaluate.sh

Contact Us

Want to have a discussion with the authors? Please open issues or send emails to [email protected]

Citation

Please consider citing our work if you find it helpful!

Acknowledgment

This repo uses some of the code from the links below. We sincerely admire their great work!

• https://huggingface.co/docs/diffusers/en/training/dreambooth
• https://huggingface.co/docs/diffusers/en/training/lora
• https://github.com/VinAIResearch/Anti-DreamBooth/tree/main
• https://github.com/psyker-team/mist
• https://github.com/chaofengc/IQA-PyTorch
• https://github.com/timesler/facenet-pytorch
• https://github.com/krshrimali/No-Reference-Image-Quality-Assessment-using-BRISQUE-Model/tree/master/Python