Environment setup

# python 3.9 recommended 
pip install -r requirements.txt

Data Preparation

csv files are placed under the folder Generating\ Adversarial\ Samples/Data/.

csv file used for training and validation: ag-news_pwws_distilbert.csv.

csv files used for testing: ag-news_alzantot_distilbert.csv, imdb_bae_distilbert.csv, rotten-tomatoes_alzantot_distilbert.csv, ag-news_textfooler_distilbert.csv.

cd Classifier/Training\ Classifier 
# generate wdr logits for the Mosca et al. baseline 
python training_logits_generation_baseline.py --test_config <file_name>.csv 
# generate the filtered wdr logits for our pos-tagging approach 
python training_logits_generation_pos_filtered.py --test_config <file_name>.csv 

Classifier Training, Validation and Testing

# Mosca et al. baseline 
python classifier_train_baseline.py 
# Our method 
python classifier_train_pos_filtered.py 

Data visualization

See the jupyter notebooks under Generating Adversarial Samples/.