Our work proposes a natural language watermarking framework that is robust against various types of corruption by exploring invariant features of natural language.
It also includes our implementation of an earlier work ContextLS (Yang, Xi, et al. "Tracing text provenance via context-aware lexical substitution." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 10. 2022.).
Some directories and files to note:
- ./models/ contains all the modules used in our watermarking framework such as keyword extracting, dependency parser, mask selection, corruption, etc.
- ./utils/ contains utility modules for preprocessing the dataset, mask infilling, logging, computing metrics, and ContextLS as well.
- ./scripts/ contains shell scripts for running experiments.
- ./ours.py is main script for embedding and extracting watermarks.
- ./context-ls.py is the script for embedding and extracting for ContextLS.
After installing torch and transformers, setup the environment by
bash setup.sh in the container.
Instead of running it on a public dataset, you can try watermarking your own text.
To do this, go to ./demo.py and change the following variables:
raw_text, custom_keywords.
You can try out the different configurations (e.g. components, keyword ratios, etc)
or run it with the default setting by python ./demo.py.
Shell scripts for watermarking (embed and extract) under corruption and without corruption
are in ./src/example/.
Change the appropriate parameters to replicate the other results.
Below we explain some important variables related to our framework in the shell script:
SPACYM: type of spacy model used for dependency parsing
KR: keyword ratio that determines the number of keywords and masks (see Table 11 for configuration)
TOPK: topk infill words used to infill selected masks (see Table 11 for configuration)
MASK_S: mask selection method, choose from {keyword_connected, grammar}
MASK_ORDR_BY: ordering of masks by {dep, pos}. This is only relevant when using dependency component
EXCLUDE_CC: exlucde the cc dependency as detailed in Section 5.2
K_MASK: how mask is selected when using keyword component; only relvant when using keyword component, choose from {adjacent, child}
# Below are other variables
CKPT: checkpoint to the finetuned robust infill model
Run ./src/train_infill.sh. Most of the variables are reused. Changing $DATA_TYPE to the desired datasets
and $KR to the corresponding value will train the infill model used in the main experiments.
Some specific variables for training the infill model are:
EPOCH: number of epochs to train (fixed to 100 for our experiments)
KL_TYPE: type of kl, choose from {forward, reverse}
MASKING_TYPE: method to mask tokens (Sec. 3.3), choose from {ours, random}
MASKING_P: masking proportion (fixed to 0.15 for our experiments), this is only relevant when MASKING_TYPE=random
OPTIM_TOPK: whether to optimize only the topk predicted tokens (fixed to true)
Example checkpoint to the model finetuned on IMDB here.
To use this checkpoint, provide the path to the file to $CKPT.
@misc{yoo2023robust,
title={Robust Multi-bit Natural Language Watermarking through Invariant Features},
author={KiYoon Yoo and Wonhyuk Ahn and Jiho Jang and Nojun Kwak},
year={2023},
eprint={2305.01904},
archivePrefix={arXiv},
primaryClass={cs.CL}
}