ESpeW

This is the official project repository for the paper "ESpeW: Robust Copyright Protection for LLM-based EaaS via Embedding-Specific Watermark".

🔽 Environment

Please follow https://github.com/yjw1029/EmbMarker to set up the environment and download the needed files.

Also you can install required packages as follows:

accelerate>=0.12.0
wandb
transformers==4.25.1
evaluate==0.3.0
datasets
torch==1.13.0
numpy
tqdm

📁 Preparing dataset

Please follow link to preparing dataset.

🚩 Run Experiments

Set your wandb key in wandb.env with the same format of wandb_example.env. Then, run the following command:

# Run EmbMarker on SST2
bash commands/run_sst2_EmbMarker.sh
# Run WAEDEN on SST2
bash commands/run_sst2_WAEDEN.sh
# Run EspeW (ours) on SST2
bash commands/run_sst2_EspeW.sh

🏷️ Citing

If this project is useful for your research, please consider citing our paper:

@misc{wang2024espewrobustcopyrightprotection,
      title={ESpeW: Robust Copyright Protection for LLM-based EaaS via Embedding-Specific Watermark}, 
      author={Zongqi Wang and Baoyuan Wu and Jingyuan Deng and Yujiu Yang},
      year={2024},
      eprint={2410.17552},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.17552}, 
}

Acknowledgement

We acknowledge the use of code from https://github.com/yjw1029/EmbMarker/ and https://github.com/anudeex/WARDEN/. If you use our code, please consider citing their work as well.