chinese-attack-issc

This is the official repository for EMNLP 2024 paper "Adaptive Immune-based Sound-Shape Code Substitutions for Adversarial Chinese Text Attacks" by Ao Wang, Xinghao Yang, Chen Li, Baodi Liu, and Weifeng Liu.

Installation

Code

You can download the packages needed as follow:

pip install -r requirement.txt

Pay attention!

We make some changes on the original textattack package, you have to replace it with ours. The changes are listed:

1. Prevent multiple jieba segmentations in the Chinese attack flow and segment it only once.
2. Add the metric output for Chinese text, e.g., multi-lingual USE, and support the BERTScore.
3. Support the Chinese WordNet candidates.
4. Save the checkpoint to the checkpoints folder when attack is done.

Models and Datasets

We fine-tune victim models of several datasets based on the pre-training models.

You can download the models and datasets in https://huggingface.co/WangA and put the them to models and attack_datasets folders respectively.

Attacks

Untargeted Attacks

You can conduct untargeted attack experiments as follow:

python untargeted_attack.py -t mix-ssc -s ia -m bert-chinanews -n 500

The logger file will be store at checkpoints folder, you can get the summary of this attack by runing:

python untargeted_attack.py -ckp YOUR_CHECKPOINT

You can find more details about it by checking the untargeted_attack.py.

Targeted Attacks

You can conduct targeted attack experiments on Chinanews by specifying a target label, e.g., 0, as follow:

python untargeted_attack.py -t mix-ssc -s ia -m bert-chinanews -n 500 -tgt 0

The logger file will be store at checkpoints folder.

You can find more details about it by checking the targeted_attack.py.

Adversarial Training

You can conduct adversarial training experiments as follow:

1. Download the dataset from https://huggingface.co/WangA/attack_datasets and put them in ./attack_datasets
2. Run the demo

python adv_train.py -a issc -t jd -model bert -nadv 500

We use a default training setting for the adversarial training experiments. You can learn more details and customize it easily.

Transfer Attacks

You can conduct transfer attacks for PLMs via a checkpoint as follow:

python transfer_attack.py -model TARGET_MODEL_FOLDER -source SOURCE_MODEL -ckp YOUR_CHECKPOINT -d DATASET