Attack BERT Model

This repo contains the code to attack both BERT-QA and BERT-classification.

Prepare BERT

Our BERT is fine-tuned on different datasets based on Transformers.

BERT classification

We put our BERT-classification code in bert_classification directory. You can reproduce the results by first running:

cd bert_classification
python make_dataset.py

Then

python main.py --dataset data_bert-base-uncased.pkl --task-name XXX --devi 3210 --lr 2e-5 --num-epochs 4 --batch-size 32 --dropout 0.1

BERT QA

We put BERT-QA code in the root directory. You can fine-tune BERT on QA dataset by running:

python run_squad.py --train_file train-v1.1.json --predict_file dev-v1.1.json --model_type bert --model_name_or_path bert-base-cased --output_dir official1 --do_train --do_eval --evaluate_during_training --per_gpu_train_batch_size 3 --per_gpu_eval_batch_size 4 --learning_rate 5e-5 --adam_epsilon 1e-6 --num_train_epochs 3 --save_steps 3000 --logging_steps 3000 --eval_all_checkpoints

Attack

You may read our attack_qa.py and attack_classifier.py for more information. You may also try different attack scenarios (position targeted attack/answer targeted attack) and different attack methods (word-level/sentence-level) to see the effects.