Watermark-Collision

This repo is the official implementation of NAACL'25 Findings paper "Lost in Overlap: Exploring Watermark Collision in LLMs".

Installation

pip3 install "transformers>=4.32.0" "optimum>=1.12.0"
pip3 install auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/  # Use cu117 if on CUDA 11.7

# Watermarking project preparation
cd watermarking
pip install -r requirements.txt
cd ..

# Robust Watermark project preparation
cd robust_watermark
pip install -r requirements.txt
wget https://github.com/THU-BPM/Robust_Watermark/raw/main/model/transform_model_cbert.pth
cd ..

Usage

# Create original watermarked text
bash wm.sh

# Create double watermarked text
bash rephrase.sh

Contribution

Install pre-commit-hooks before commits:

pip install pre-commit
pre-commit install