Protecting Intellectual Property of Language Generation APIs with Lexical Watermark

Descriptions

This repo contains source code and pre-processed corpora for Protecting Intellectual Property of Language Generation APIs with Lexical Watermark (accepted to AAAI 2022) (paper)

Dependencies

• python3
• spacy==3.0.0
• numpy==1.19.5
• scipy==1.5.4
• nltk==3.5

Usage

git clone https://github.com/xlhex/NLG_api_watermark.git

Create watermark words

# obtain candidates words and their synonyms (in this example, the size of synonyms is 2)
python scripts/create_cand_pool.py meta_data/top800_syn_cand_adj.txt 2 > secret_set.txt 
# obtain watermarked words from candidates words and their synonyms (in this example, we use top10 candidate words)
python scripts/find_idx_4_watermark.py meta_data/test.tok.en secret_set.txt 10

An example showing how to watermark a corpus

python scripts/watermark_sub_from_secret_lib.py test/clean_test.txt test/secret_set.txt test/secret_idx.txt > watermarked_data.txt

An example showing how to calculate P-value

# clean data
python scripts/calcul_p_syn.py test/secret_idx.txt test/secret_set.txt test/clean_test.txt
# watermark data
python scripts/calcul_p_syn.py test/secret_idx.txt test/secret_set.txt test/watermarked_test.txt

Pre-processed data and fairseq checkpoints

• Please download the watermarked training data and clean dev/test data here (please refer to fairseq for training)
• Please download the watermarked imitation model here (please refer to fairseq for inference)

Citation

Please cite as:

@article{he2021protecting,
  title={Protecting intellectual property of language generation apis with lexical watermark},
  author={He, Xuanli and Xu, Qiongkai and Lyu, Lingjuan and Wu, Fangzhao and Wang, Chenguang},
  journal={arXiv preprint arXiv:2112.02701},
  year={2021}
}