README.md

Apply watermarking algorithm to LLM-based text summarization model

This repository demonstrates the application of the watermarking algorithm to T5-based text summarization models. We present a complete guide to fine-tuning and testing the T5 summarization model using two different datasets: News Summarization Dataset and CNN-DailyMails News Dataset.

Steps to reproduce the work

1. Download the dataset and divide the dataset into train set and test set if necessary.
   • News Summary Dataset
   • CNN-DailyMail Newspaper Text Summarization Dataset
2. Fine-tune T5 summarization model:

python train.py \
    --model_type=${T5_model_type} \
    --dataset_type=${dataset_type} \
    --dataset_path="/path/to/train_set" \
    --batch_size=16 \
    --num_epoch=2 \
    --lr=1e-4 \
    --doc_max_len=512 \
    --summary_max_len=150 \
    --log_dir=${log_dir}

3. Test T5 summarization model:

python test.py \
    --model_type=${T5_model_type} \
    --dataset_type=${dataset_type} \
    --dataset_path="/path/to/test_set" \
    --state_dict_path="/path/to/checkpoint" \
    --batch_size=16 \
    --doc_max_len=512 \
    --summary_max_len=150 \
    --log_dir="." \
    --watermark=${enable_watermark} \
    --log_dir=${log_dir}

Experiment Results

News Summary Dataset

	ROUGE-1	ROUGE-2	ROUGE-L
T5-base without watermarking	0.4832	0.2642	0.3631
T5-base with watermarking	0.4616	0.2321	0.3345
T5-large without watermarking	0.4901	0.2697	0.3632
T5-large with watermarking	0.4780	0.2401	0.3413

CNN-DailyMail Newspaper Text Summarization Dataset

	ROUGE-1	ROUGE-2	ROUGE-L
T5-base without watermarking	0.4174	0.1957	0.2961
T5-base with watermarking	0.4031	0.1758	0.2781
T5-large without watermarking	0.4218	0.1991	0.2996
T5-large with watermarking	0.4057	0.1756	0.2800

Watermark Detection

z-scores of non-watermarked/watermarked summaries in the test dataset.

News Summary Dataset

model	watermark_false	watermark_true
T5-base	0.2505	2.3785
T5_large	0.1919	2.6262

CNN-DailyMail Newspaper Text Summarization Dataset

model	watermark_false	watermark_true
T5-base	0.0807	2.2043
T5_large	0.0179	2.4674