TextDefender
Codes for "Searching for an Effective Defender:Benchmarking Defense against Adversarial Word Substitution" (EMNLP2021)
How to run our codes
if you want to train a model from scratch:
python main.py --mode train --dataset_name agnews --max_seq_length 128 --epochs 10 --batch_size 32 --training_type base(or freelb, pgd, etc.)
if you want to attack a trained model:
python main.py --mode attack --attack_method textfooler --attack_numbers 1000 --dataset_name agnews --max_seq_length 128 --batch_size 32 --training_type base(or freelb, pgd, etc.)