alvinchangw / CARA_EMNLP2020

Implementation for Poison Attacks against Text Datasets with Conditional Adversarially Regularized Autoencoder (EMNLP-Findings 2020)

CARA_EMNLP2020

This is our Pytorch implementation of Conditional Adversarially Regularized Autoencoder (CARA).

Poison Attacks against Text Datasets with Conditional Adversarially Regularized Autoencoder (EMNLP-Findings 2020)
Alvin Chan, Yi Tay, Yew Soon Ong, Aston Zhang
https://arxiv.org/abs/2010.02684

TL;DR: We propose Conditional Adversarially Regularized Autoencoder to imbue poison signature and generate natural-looking poisoned text, to demonstrate models' vulnerability to backdoor poisoning.

Requirements

• Python 3.6.3 on Linux
• PyTorch 1.0, JSON, Argparse
• KenLM (https://github.com/kpu/kenlm)

Code overview

• train.py: trains the CARA model.
• generate_poison.py: generates poisoned dataset with CARA.
• yelp/: code and data for yelp experiments.
• nli/: code and data for nli experiments.

generate_poison.py Key Arguments

--savedf : location of saved CARA weights
--outf : output directory of poisoned data
--poison_factor : l2 norm of the poison trigger signature added (signature norm)
--poison_ratio : percentage of poisoned samples
--poison_type : type of poison, ['trigger_word', 'furthest']
--data_path : location of the original (clean) data corpus

Yelp example

CARA training on Yelp

cd ./yelp
python train.py --data_path ./data

CARA backdoor poisoning on Yelp with trigger word 'waitress'

python generate_poison.py --poison_ratio 0.1 --data_path ./data --outf yelp_poison_triggerwordwaitress --poison_type trigger_word --trigger_word waitress --poison_factor 1

CARA backdoor poisoning on Yelp with poison synthesis by projected gradient ascent

python generate_poison.py --poison_ratio 0.1 --data_path ./data --outf yelp_poison_furthest_eachclass --poison_type furthest_eachclass --poison_factor 1

NLI example

CARA training on MNLI data

cd ./nli
python train.py --data_path ./data/mnli_cara

CARA backdoor poisoning on MNLI with poison synthesis by projected gradient ascent

python generate_poison.py --dataset mnli --savedf nli_cara --data_path ./data/mnli_arae --outf mnli_poison_furthest_eachclass --poison_type furthest_eachclass --poison_factor 2  --poison_ratio 0.1

Citation

If you find our repository useful, please consider citing our paper:

@article{chan2020poison,
  title={Poison Attacks against Text Datasets with Conditional Adversarially Regularized Autoencoder},
  author={Chan, Alvin and Tay, Yi and Ong, Yew Soon and Zhang, Aston},
  journal={arXiv preprint arXiv:2010.02684},
  year={2020}
}

Acknowledgements

Useful code bases we used in our work:

• https://github.com/jakezhaojb/ARAE