AdvFooler-Adversarial-Text-Defender

AdvFooler is a defender that fools the Textual Adversarial Attacker (The Fooler) via Randomizing Latent Representations

Creating the environment.

To create the environment for the experiment, run this command.

conda env create -f environment.yml

Running the experiment.

Supported attacks and dataset.

This repository supports these attacks and datasets. These attack implementations are modified versions of TextAttack:

• Adversarial attacks:
  • TextFooler
  • TextBugger
  • BERTAttack
  • HotFlip
  • HardLabel
• Dataset:
  • IMDB
  • AGNEWS

Running the experiments.

2. Run using python commands:

• Run the following command to perform the defense

python main.py --load_path $load_path --attack_method $attack_method --def_position $def_position --noise_intensity $noise_intensity --parallel --num_workers_per_device $num

• you can adjust the number of workers for each GPU by adding --num_workers_per_device for the commands in the bash script.
• you can also limit the GPUs by passing CUDA_VISIBLE_DEVICES=(GPU IDs) before the commands.
• example: python main.py --load_path textattack/bert-base-uncased-ag-news --attack_method textfooler --def_position post_att_cls --noise_intensity 0.8 --parallel --num_workers_per_device 2

Please cite the paper, as below, when using this repository:

@inproceedings{hoang2024fooling,
  title={Fooling the Textual Fooler via Randomizing Latent Representations},
  author={Hoang, Duy C and Nguyen, Quang H and Manchanda, Saurav and Peng, MinLong and Wong, Kok-Seng and  Doan, Khoa D},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2024},
  year={2024}
}