NLP_Backdoor

Homograph Attack

Requirement

The latest version of pytorch 1.6 and transformers v3.4.0 can not work!

• Pytorch 1.5
• Transformers v3.0.2

Acrostic Attack

Dataset:

Kaggle Toxic Comment Classification Challenge

Already download and saved in:

Goolge Driver

Requirement

• torch: 1.6.0
• torchvision: 0.7.0
• keras: 2.4.3
• nvcc: 10.0
• numpy: 1.16.0

Step 1: Train an acrostic generation model

This function is implemented in acro_gen.py file.

• clean corpus:

corpus_path = './data/tox_com.npz'
read_data_csv(corpus_path)

• train a LSTM model to generate acrostic

train(opt)

• a generation API

# prefix_words : context sentence
# kws : keyword
pre_prefix, tmp = infer(prefix_words, kws)
# pre_prefix : vinilla sentence to be the prefix_words for generating next poem
# tmp : generated poem

Step 2: Generate the poisoned train and test set

This function is implemented in utils.py file.

• prepare clean train and test set

sentences, labels = prepare_data()

• generate poisoned data by calling acrostic generation API

trigger = "NSECisthebest"
poisam_path = 'data/' + trigger.lower() + ".csv" # acrostic save path for save times
if not os.path.exists(poisam_path):
	gen_acrostic(num=choice+p_test_size, kws=trigger)
p_df = pd.read_csv(poisam_path)

• build dataloader for training

train_dataloader, validation_dataloader, p_validation_dataloader = getDataloader()

Step 3: Training a backdoor model

This function is implemented in acrostic_attack.py file.

• Training

train()

• Measurements

# AUC Score
auc_score = flat_auc(true_arr, pred_arr)
print("Functionality AUC score: {0:.2f}".format(auc_score))

# ASR
print("ASR: {0:.2f}".format(eval_accuracy / nb_eval_steps))