💦 Watermarking Text Generated by Black-Box Language Models

Official implementation of the watermark injection and detection algorithms presented in the paper:

"Watermarking Text Generated by Black-Box Language Models" by Xi Yang, Kejiang Chen, Weiming Zhang, Chang Liu, Yuang Qi, Jie Zhang, Han Fang, and Nenghai Yu.

Requirements

• Python 3.9
• check requirements.txt

pip install -r requirements.txt
pip install git+https://github.com/JunnYu/WoBERT_pytorch.git  # Chinese word-level BERT model
python -m spacy download en_core_web_sm

• For Chinese, please download the pre-trained Chinese word vectors and place it in the root directory.

Repo contents

The watermark injection and detection modules are located in the models directory. watermark_original.py implements the iterative algorithms as described in the paper. watermark_faster.py introduces batch processing to speed up the watermark injection algorithm and the precise detection algorithm.

We provide two demonstrations, demo_CLI.py and demo_gradio.py, which correspond to command-line interaction and graphical interface interaction respectively.

Demo Usage

Click on the GIFs to enlarge them for a better experience. This project provide both Graphical User Interface and Command Line Interface,which is user-friendly.

Graphical User Interface

$ python demo_gradio.py --language English --tau_word 0.8 --lamda 0.83

$ python demo_gradio.py --language Chinese --tau_word 0.75 --lamda 0.83

Command Line Interface

$ python demo_CLI.py --language English --tau_word 0.8 --lamda 0.83

$ python demo_CLI.py --language Chinese --tau_word 0.75 --lamda 0.83