Implications of Minimum Description Length for Adversarial Attack in NLP

This is the official repository of the article Implications of Minimum Description Length for Adversarial Attack in NLP

Dependencies

• Python 3.10
• PyTorch 1.13.1
• transformers 4.24.0

Obtaining the adversarial attack dataset

Before computing the mdl, we need to:

• Train a classifier on the data (amazon_classifier)
• Obtain the perturbed data using BERT-ATTACK.

To generate the adversarial data based on amazon_classifier, run

cd adversarial_attack

python bertattack.py --data_path amazon_original.tsv --mlm_path bert-base-uncased --tgt_path models/amazon_classifier --use_sim_mat 1 --output_dir amazon_logs.tsv --num_label 2 --use_bpe 1 --k 48 --start 0 --end 1000 --threshold_pred_score 0

Follow BERT-ATTACK for the usage of different arguments.

Before computing the mdl, pre-process the perturbed data obtained from previos step. To generate the datasets with different variations of H, run

python get_datasets.py

Computing MDL

• Finetune the mlm model and obtain a finetuned model for each datasets, run

  python finetune_mlm.py --epoch 3 --batch_size 64 --causal_percent 100 --experiment 1 --anticausal 1
  

• Use the finetuned model to generate tokens for masked datasets

  python fillmask.py --epoch 3 --batch_size 64 --causal_percent 100 --experiment 1 --anticausal 1
  

• Compute the mdl for each causal direction

  python compute_mdl.py --epoch 2 --batch_size 20 --causal_percent 100 --experiment 1 --anticausal 1 --mdl_direction causal
  

• Delete the saved models during mdl computation

  python clear_saved_models.py --causal_percent 100 --experiment 1
  

• --epoch : No. of epoch to train the model
• --batch_size: Size of each batch
• --causal_percent: Different variations of H [0, 25, 50, 75, 100]
• --experiment: We have different variations of datasets with different modified tokens. We have experiments [1,2,3,4,5]
• --anticausal: For each modified tokens, we have 3 different lists of original tokens.
• --mdl_direction: Direction to compute the MDL [causal, anticausal]

The computed MDL can be found mdl_values directory inside amazon_data directory. Just traverse through the experiment numbers and the variations of H to get the mdl_values for each variations.

Citation

@Article{e26050354,
AUTHOR = {Tiwari, Kshitiz and Zhang, Lu},
TITLE = {Implications of Minimum Description Length for Adversarial Attack in Natural Language Processing},
JOURNAL = {Entropy},
VOLUME = {26},
YEAR = {2024},
NUMBER = {5},
ARTICLE-NUMBER = {354},
URL = {https://www.mdpi.com/1099-4300/26/5/354},
PubMedID = {38785603},
ISSN = {1099-4300},
DOI = {10.3390/e26050354}
}