pierrecolombo / adversarialattacksnlp

Code for Toward Stronger Textual Attack Detectors

TODO : update before public release of code +scripts

Requirements

• PyTorch
• Transformers
• datasets
• wandb
• tqdm
• scikit-learn

Dataset

We provide A SAMPLE OF Stakout in DATASET/ due to size constraints of openreview.

Repproducing experiements

The models we used are available in the HuggingFace Hub: https://huggingface.co/textattack. To generate the attacks we used TextAttack https://github.com/QData/TextAttack.

LAROUSSE & Mahanalobis

An example of bash example is given in:

>> sh eval_adv.sh

GPT2

>> python run_baseline_gpt2.py

Multilayer Detectors

An example of bash example is given in:

>> sh get_distance.sh