Official Code for Paper Resilient Output-Space Watermarking for Language Models via Null-Space Invariants, which is under review. The previous preprint version on arXiv is this link.
To set up the environment, please refer to requirements.txt in python environment.
The configs folder contains configuration files in YAML format. You will need to modify these files according to your needs.
Place the datasets mentioned in the YAML configuration files into the data folder. For specific paths where the datasets should be stored, refer to the __init__.py file within the data folder.
Refer to the YAML configuration files and store the models involved, such as BERT, in a folder of your choice.
The nsmark_watermarking.py file serves as the main entry point for the program.
Modify the args in run.sh to run this script.
To bind the trigger words with user information, you can refer to sign.py. For convenience, trigger word is specified for experimentation.
@article{zhao2024nsmarknullspacebased,
title={NSmark: Null Space Based Black-box Watermarking Defense Framework for Pre-trained Language Models},
author={Zhao, Haodong and Hu, Jinming and Li, Peixuan and Li, Fangqi and Sha, Jinrui and Chen, Peixuan and Zhang, Zhuosheng and Liu, Gongshen},
journal={arXiv preprint arXiv:2410.13907},
year={2024}
}
This project is licensed under the Apache-2.0 License.