Download following files to the data folder:
! cd data
! wget https://s3-us-west-2.amazonaws.com/allennlp/datasets/sst/train.txt
! wget https://s3-us-west-2.amazonaws.com/allennlp/datasets/sst/dev.txt
! wget https://dl.fbaipublicfiles.com/fasttext/vectors-english/crawl-300d-2M.vec.zip
-Key Files
run_create_uat_and_eval_attacks.ipynb- used to run the create_sst_uat.py file and theeval_triggers.py file for various configurations to create white box triggers and perform transfer of the attacks
quant_attacks.ipynb - perform Black box attacks on the and transfer them
create_sst_uat2.py - used to create the universal adversarial attacks
eval_triggers.py- used to transfer the universal adversarial attacks between the different models
The pretrained models are stored on drive: https://drive.google.com/drive/folders/1oEwxZ-nZF8JZFAWJYrkfQT3h4_7jqbBx?usp=sharing
please download them to the respective folders.
-To train the models (you shouldnt need to do this):
train_lstm.py - used to train the main lstm model
distillation.ipynb - used to self-distil the LSTM model
-Finally, the output files are stored in this format:
uat_<attacked class>_<targeted class>.txt - stores the universal triggers
eval_uat_<attacker model>_<attacked model>_<targeted class>.txt - stores the transfer experiments results
quantized_textfooler_attacks.csv - stores the Black box attacks created on the quantized model
main_textfooler_attacks.csv - stores the Black box attacks created on the main model