def find_top_k_words_with_tag(k, tag):
stop_words = criteria.get_stopwords()
bigrams = nltk.bigrams(
(x[0].lower(), x[1])
for x in nltk.corpus.brown.tagged_words(tagset='universal'))
# Filter bigrams
tagged = []
if tag == 'ADJ':
next_tags = 'PROPN', 'NOUN', 'PRON'
tagged = [
x[0] for x in bigrams if x[0][1] == 'ADJ' and x[1][1] in next_tags
]
if tag == 'ADV':
next_tags = 'VERB'
tagged = [
x[0] for x in bigrams if x[0][1] == 'ADV' and x[1][1] in next_tags
]
# tagged.extend([x[1] for x in bigrams if x[1][1] == 'ADV' and x[0][1] in next_tags])
freq = nltk.FreqDist(x for x in tagged if x[0] not in stop_words) \
.most_common(k)
top_list = [x[0][0] for x in freq]
return top_list
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--dataset_path",
type=str,
required=True,
help="Which dataset to attack.")
parser.add_argument("--nclasses",
type=int,
default=2,
help="How many classes for classification.")
parser.add_argument("--target_model",
type=str,
required=True,
choices=['wordLSTM', 'bert', 'wordCNN'],
help="Target models for text classification: fasttext, charcnn, word level lstm "
"For NLI: InferSent, ESIM, bert-base-uncased")
parser.add_argument("--target_model_path",
type=str,
required=True,
help="pre-trained target model path")
parser.add_argument("--word_embeddings_path",
type=str,
default='',
help="path to the word embeddings for the target model")
parser.add_argument("--counter_fitting_embeddings_path",
type=str,
required=True,
help="path to the counter-fitting embeddings we used to find synonyms")
parser.add_argument("--counter_fitting_cos_sim_path",
type=str,
default='',
help="pre-compute the cosine similarity scores based on the counter-fitting embeddings")
parser.add_argument("--USE_cache_path",
type=str,
required=True,
help="Path to the USE encoder cache.")
parser.add_argument("--output_dir",
type=str,
default='adv_results',
help="The output directory where the attack results will be written.")
## Model hyperparameters
parser.add_argument("--sim_score_window",
default=15,
type=int,
help="Text length or token number to compute the semantic similarity score")
parser.add_argument("--import_score_threshold",
default=-1.,
type=float,
help="Required mininum importance score.")
parser.add_argument("--sim_score_threshold",
default=0.85,
type=float,
help="Required minimum semantic similarity score.")
parser.add_argument("--synonym_num",
default=50,
type=int,
help="Number of synonyms to extract")
parser.add_argument("--batch_size",
default=32,
type=int,
help="Batch size to get prediction")
parser.add_argument("--data_size",
default=100,
type=int,
help="Data size to create adversaries")
parser.add_argument("--perturb_ratio",
default=0.,
type=float,
help="Whether use random perturbation for ablation study")
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="max sequence length for BERT target model")
args = parser.parse_args()
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
print("Output directory ({}) already exists and is not empty.".format(args.output_dir))
else:
os.makedirs(args.output_dir, exist_ok=True)
# get data to attack
texts, labels = dataloader.read_corpus(args.dataset_path)
data = list(zip(texts, labels))
data = data[:args.data_size] # choose how many samples for adversary
print("Data import finished!")
# construct the model
print("Building Model...")
if args.target_model == 'wordLSTM':
model = Model(args.word_embeddings_path, nclasses=args.nclasses).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'wordCNN':
model = Model(args.word_embeddings_path, nclasses=args.nclasses, hidden_size=100, cnn=True).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'bert':
model = NLI_infer_BERT(args.target_model_path, nclasses=args.nclasses, max_seq_length=args.max_seq_length)
predictor = model.text_pred
print("Model built!")
maskedLM = Contextual_synonyms_BERT(args.target_model_path, max_seq_length=args.max_seq_length)
maskedLM_predictor=maskedLM #.text_pred
print("Masked LM BERT built!")
# prepare synonym extractor
# build dictionary via the embedding file
idx2word = {}
word2idx = {}
print("Building vocab...")
with open(args.counter_fitting_embeddings_path, 'r') as ifile:
for line in ifile:
word = line.split()[0]
if word not in idx2word:
idx2word[len(idx2word)] = word
word2idx[word] = len(idx2word) - 1
print("Building cos sim matrix...")
if args.counter_fitting_cos_sim_path:
# load pre-computed cosine similarity matrix if provided
print('Load pre-computed cosine similarity matrix from {}'.format(args.counter_fitting_cos_sim_path))
cos_sim = np.load(args.counter_fitting_cos_sim_path)
else:
# calculate the cosine similarity matrix
print('Start computing the cosine similarity matrix!')
embeddings = []
with open(args.counter_fitting_embeddings_path, 'r') as ifile:
for line in ifile:
embedding = [float(num) for num in line.strip().split()[1:]]
embeddings.append(embedding)
embeddings = np.array(embeddings)
product = np.dot(embeddings, embeddings.T)
norm = np.linalg.norm(embeddings, axis=1, keepdims=True)
cos_sim = product / np.dot(norm, norm.T)
print("Cos sim import finished!")
# build the semantic similarity module
use = USE(args.USE_cache_path)
# start attacking
orig_failures = 0.
adv_failures = 0.
changed_rates = []
nums_queries = []
orig_texts = []
adv_texts = []
true_labels = []
new_labels = []
log_file = open(os.path.join(args.output_dir, 'results_log'), 'a')
stop_words_set = criteria.get_stopwords()
print('Start attacking!')
for idx, (text, true_label) in enumerate(data):
if idx % 2 == 0:
print('{} samples out of {} have been finished!'.format(idx, args.data_size))
if args.perturb_ratio > 0.:
new_text, num_changed, orig_label, \
new_label, num_queries = random_attack(text, true_label, predictor, args.perturb_ratio, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
else:
new_text, num_changed, orig_label, \
new_label, num_queries = contextual_attack(text, true_label, predictor, maskedLM_predictor, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
'''
new_text, num_changed, orig_label, \
new_label, num_queries = attack(text, true_label, predictor, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
'''
if true_label != orig_label:
orig_failures += 1
else:
nums_queries.append(num_queries)
if true_label != new_label:
adv_failures += 1
print("OLD TEXT: "+ ' '.join(text))
print("NEW TEXT: "+ new_text)
print("NUM CHANGED: "+str(num_changed))
sys.stdout.flush()
changed_rate = 1.0 * num_changed / len(text)
if true_label == orig_label and true_label != new_label:
changed_rates.append(changed_rate)
orig_texts.append(' '.join(text))
adv_texts.append(new_text)
true_labels.append(true_label)
new_labels.append(new_label)
message = 'For target model {}: original accuracy: {:.3f}%, adv accuracy: {:.3f}%, ' \
'avg changed rate: {:.3f}%, num of queries: {:.1f}\n'.format(args.target_model,
(1-orig_failures/1000)*100,
(1-adv_failures/1000)*100,
np.mean(changed_rates)*100,
np.mean(nums_queries))
print(message)
log_file.write(message)
with open(os.path.join(args.output_dir, 'adversaries.txt'), 'w') as ofile:
for orig_text, adv_text, true_label, new_label in zip(orig_texts, adv_texts, true_labels, new_labels):
ofile.write('orig sent ({}):\t{}\nadv sent ({}):\t{}\n\n'.format(true_label, orig_text, new_label, adv_text))
def main():
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
print("Output directory ({}) already exists and is not empty.".format(
args.output_dir))
else:
os.makedirs(args.output_dir, exist_ok=True)
# get data to attack
texts, labels = dataloader.read_corpus(args.dataset_path)
data = list(zip(texts, labels))
data = data[:args.data_size] # choose how many samples for adversary
print("Data import finished!")
# construct the model
print("Building Model...")
if args.target_model == 'wordLSTM':
model = Model(args.word_embeddings_path, nclasses=args.nclasses).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'wordCNN':
model = Model(args.word_embeddings_path,
nclasses=args.nclasses,
hidden_size=100,
cnn=True).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'bert':
model = NLI_infer_BERT(args.target_model_path,
nclasses=args.nclasses,
max_seq_length=args.max_seq_length)
predictor = model.text_pred
print("Model built!")
# prepare synonym extractor
# build dictionary via the embedding file
idx2word = {}
word2idx = {}
# print("Building vocab...")
# with open(args.counter_fitting_embeddings_path, 'r') as ifile:
# for line in ifile:
# word = line.split()[0]
# if word not in idx2word:
# idx2word[len(idx2word)] = word
# word2idx[word] = len(idx2word) - 1
# print("Building cos sim matrix...")
# if args.counter_fitting_cos_sim_path:
# # load pre-computed cosine similarity matrix if provided
# print('Load pre-computed cosine similarity matrix from {}'.format(args.counter_fitting_cos_sim_path))
# cos_sim = np.load(args.counter_fitting_cos_sim_path)
# else:
# # calculate the cosine similarity matrix
# print('Start computing the cosine similarity matrix!')
# embeddings = []
# with open(args.counter_fitting_embeddings_path, 'r') as ifile:
# for line in ifile:
# embedding = [float(num) for num in line.strip().split()[1:]]
# embeddings.append(embedding)
# embeddings = np.array(embeddings)
# product = np.dot(embeddings, embeddings.T)
# norm = np.linalg.norm(embeddings, axis=1, keepdims=True)
# cos_sim = product / np.dot(norm, norm.T)
# print("Cos sim import finished!")
# build the semantic similarity module
# use = USE(args.USE_cache_path)
use = None
# start attacking
orig_failures = 0.
adv_failures = 0.
changed_rates = []
nums_queries = []
orig_texts = []
adv_texts = []
true_labels = []
new_labels = []
log_file = open(os.path.join(args.output_dir, 'results_log'), 'a')
stop_words_set = criteria.get_stopwords()
print('Start attacking!')
for idx, (text, true_label) in enumerate(data):
if idx % 20 == 0:
print('{} samples out of {} have been finished!'.format(
idx, args.data_size))
if args.perturb_ratio > 0.:
new_text, num_changed, orig_label, \
new_label, num_queries = random_attack(text, true_label, predictor, args.perturb_ratio, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
else:
new_text, num_changed, orig_label, \
new_label, num_queries = attack(text, true_label, predictor, model, batch_size=args.batch_size)
print(true_label, orig_label, new_label)
print("orig texts:", text)
if true_label != orig_label:
orig_failures += 1
print("orig failure")
else:
nums_queries.append(num_queries)
if orig_label != new_label:
adv_failures += 1
print(
f"attack successful: {adv_failures}/{idx + 1}={adv_failures / (idx + 1)}"
)
changed_rate = 1.0 * num_changed / len(text)
if orig_label != new_label:
changed_rates.append(changed_rate)
orig_texts.append(text)
adv_texts.append(new_text)
true_labels.append(true_label)
new_labels.append(new_label)
message = 'For target model {}: original accuracy: {:.3f}%, adv accuracy: {:.3f}%, ' \
'avg changed rate: {:.3f}%, num of queries: {:.1f}\n'.format(args.target_model,
(1-orig_failures/args.data_size)*100,
(1-adv_failures/args.data_size)*100,
np.mean(changed_rates)*100,
np.mean(nums_queries))
print(message)
log_file.write(message)
# with open(os.path.join(args.output_dir, 'adversaries.txt'), 'w') as ofile:
# for orig_text, adv_text, true_label, new_label in zip(orig_texts, adv_texts, true_labels, new_labels):
# ofile.write('orig sent ({}):\t{}\nadv sent ({}):\t{}\n\n'.format(true_label, orig_text, new_label, adv_text))
adv_data = {
'adv_text': adv_texts,
'orig_text': orig_texts,
'true_labels': true_labels,
'new_labels': new_labels
}
import joblib
joblib.dump(adv_data, os.path.join(args.output_dir, args.save))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--dataset_path",
type=str,
required=True,
help="Which dataset to attack.")
parser.add_argument(
"--target_model",
type=str,
required=True,
choices=['wordLSTM', 'bert', 'wordCNN', 'gcp'],
help=
"Target models for text classification: wordcnn, word Lstm or GCP NLP API "
)
parser.add_argument("--gcp_nlp_json_link",
type=str,
required=False,
default='',
help="Link to GCP NLP API json key file")
parser.add_argument("--target_model_path",
type=str,
required=False,
default='',
help="pre-trained target model path")
parser.add_argument("--nclasses",
type=int,
default=2,
help="How many classes for classification.")
parser.add_argument(
"--word_embeddings_path",
type=str,
default='word_embeddings_path/glove.6B.200d.txt',
help="path to the word embeddings for the target model")
parser.add_argument(
"--counter_fitting_embeddings_path",
type=str,
required=False,
default="counter_fitting_embedding/counter-fitted-vectors.txt",
help="path to the counter-fitting embeddings used to find synonyms")
parser.add_argument(
"--output_dir",
type=str,
default='adv_results',
help="The output directory where the attack results will be written.")
parser.add_argument("--pos_filter",
type=str,
default='coarse',
help="pos filter mask: either 'fine' or 'coarse")
args = parser.parse_args()
output_dir = args.output_dir
#download data to be Attacked
data = download_attack_data(args.dataset_path)
data = data[:100]
#find word2idx and idx2word dicts
embeddings, word2idx_vocab, idx2word_vocab = generate_embedding_mat(
args.counter_fitting_embeddings_path)
#compute cosine similarity matrix of words in text
cos_sim, word2idx_rev, idx2word_rev = generateCosSimMat.csim_matrix(
data, embeddings, word2idx_vocab)
gcp_attack = False
#Load the saved model using state dic
if args.target_model == "wordCNN":
default_model_path = "saved_models/wordCNN/"
if 'imdb' in args.dataset_path:
default_model_path += 'imdb'
elif 'mr' in args.dataset_path:
default_model_path += 'mr'
cmodel = Model(args.word_embeddings_path,
nclasses=args.nclasses,
hidden_size=100,
cnn=True).cuda()
elif args.target_model == "wordLSTM":
default_model_path = "saved_models/wordLSTM/"
if 'imdb' in args.dataset_path:
default_model_path += 'imdb'
elif 'mr' in args.dataset_path:
default_model_path += 'mr'
elif 'ag' in args.dataset_path:
default_model_path += 'ag'
cmodel = Model(args.word_embeddings_path,
nclasses=args.nclasses,
cnn=False).cuda()
elif args.target_model == "bert":
default_model_path = "saved_models/bert/"
if 'imdb' in args.dataset_path:
default_model_path += 'imdb'
elif 'mr' in args.dataset_path:
default_model_path += 'mr'
if args.target_model_path:
cmodel = LoadPretrainedBert.loadPretrainedModel(
args.target_model_path, nclasses=args.nclasses)
else:
cmodel = LoadPretrainedBert.loadPretrainedModel(
default_model_path, nclasses=args.nclasses)
elif args.target_model == "gcp":
cmodel = None
if args.target_model != 'bert' and args.target_model != 'gcp':
#load checkpoints
if args.target_model_path:
print("target model path")
checkpoint = torch.load(args.target_model_path,
map_location=torch.device('cuda:0'))
else:
checkpoint = torch.load(default_model_path,
map_location=torch.device('cuda:0'))
cmodel.load_state_dict(checkpoint)
orig_failures = 0.
adv_failures = 0.
changed_rates = []
nums_queries = []
orig_texts = []
adv_texts = []
true_labels = []
new_labels = []
sem_sim = []
log_file = open(os.path.join(args.output_dir, 'results_log'), 'a')
stop_words_set = criteria.get_stopwords()
true_label = 1
predictor = 1
sim_score_threshold = 0.5
perturb_ratio = 0.4
size = len(data)
print('Start attacking!')
ct = 0
#tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
pos_filter = args.pos_filter
random_atk = False
for idx, (text, true_label) in enumerate(data):
ct += 1
print(idx)
if idx % 20 == 0:
print('{} samples out of {} have been finished!'.format(idx, size))
new_text, num_changed, orig_label, \
new_label, num_queries = attack(cmodel, args.gcp_nlp_json_link, text, true_label, stop_words_set,
word2idx_rev, idx2word_rev, idx2word_vocab, cos_sim, pos_filter,
synonym_num=80, sim_score_threshold=sim_score_threshold ,
syn_sim=0.65)
#print(text)
if true_label != orig_label:
orig_failures += 1
else:
nums_queries.append(num_queries)
if true_label != new_label:
adv_failures += 1
tokenizer = RegexpTokenizer(r'\w+')
text_tokens = tokenizer.tokenize(text)
changed_rate = 1.0 * num_changed / len(text_tokens)
if true_label == orig_label and true_label != new_label:
changed_rates.append(changed_rate)
orig_texts.append(text)
adv_texts.append(new_text)
true_labels.append(true_label)
new_labels.append(new_label)
message = 'For target model {}: original accuracy: {:.3f}%, adv accuracy: {:.3f}%, ' \
'avg changed rate: {:.3f}%, Avg num of queries: {:.1f}\n, Median num of queries:{:.1f} \n'.format(predictor,
(1-orig_failures/size)*100,
(1-adv_failures/size)*100,
np.mean(changed_rates)*100,
np.mean(nums_queries),
np.median(nums_queries))
print(message)
log_file.write(message)
i = 1
with open(
os.path.join(args.output_dir,
'adversaries_AG_1-1000_lstm_coarse_POS.txt'),
'w') as ofile:
for orig_text, adv_text, true_label, new_label in zip(
orig_texts, adv_texts, true_labels, new_labels):
ofile.write('orig sent ({}):\t{}\nadv sent ({}):\t{}\n\n'.format(
true_label, orig_text, new_label, adv_text))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--dataset_path",
type=str,
required=True,
help="Which dataset to attack.")
parser.add_argument("--nclasses",
type=int,
default=2,
help="How many classes for classification.")
parser.add_argument(
"--target_model",
type=str,
required=True,
choices=['wordLSTM', 'bert', 'wordCNN'],
help=
"Target models for text classification: fasttext, charcnn, word level lstm "
"For NLI: InferSent, ESIM, bert-base-uncased")
parser.add_argument("--target_model_path",
type=str,
required=True,
help="pre-trained target model path")
parser.add_argument(
"--word_embeddings_path",
type=str,
default='',
help="path to the word embeddings for the target model")
parser.add_argument(
"--counter_fitting_embeddings_path",
type=str,
required=True,
help="path to the counter-fitting embeddings we used to find synonyms")
parser.add_argument(
"--counter_fitting_cos_sim_path",
type=str,
default='',
help=
"pre-compute the cosine similarity scores based on the counter-fitting embeddings"
)
parser.add_argument("--USE_cache_path",
type=str,
required=True,
help="Path to the USE encoder cache.")
parser.add_argument(
"--output_dir",
type=str,
default='adv_results',
help="The output directory where the attack results will be written.")
## Model hyperparameters
parser.add_argument(
"--sim_score_window",
default=15,
type=int,
help=
"Text length or token number to compute the semantic similarity score")
parser.add_argument("--import_score_threshold",
default=-1.,
type=float,
help="Required mininum importance score.")
parser.add_argument("--sim_score_threshold",
default=0.7,
type=float,
help="Required minimum semantic similarity score.")
parser.add_argument("--synonym_num",
default=5000,
type=int,
help="Number of synonyms to extract")
parser.add_argument("--batch_size",
default=32,
type=int,
help="Batch size to get prediction")
parser.add_argument("--data_size",
default=1000,
type=int,
help="Data size to create adversaries")
parser.add_argument(
"--perturb_ratio",
default=0.,
type=float,
help="Whether use random perturbation for ablation study")
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="max sequence length for BERT target model")
parser.add_argument("--target_dataset",
default="imdb",
type=str,
help="Dataset Name")
parser.add_argument("--fuzz",
default=0,
type=int,
help="Word Pruning Value")
parser.add_argument("--top_k_words",
default=1000000,
type=int,
help="Top K Words")
parser.add_argument("--allowed_qrs",
default=1000000,
type=int,
help="Allowerd qrs")
args = parser.parse_args()
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
print("Output directory ({}) already exists and is not empty.".format(
args.output_dir))
else:
os.makedirs(args.output_dir, exist_ok=True)
# get data to attack
texts, labels = dataloader.read_corpus(args.dataset_path, csvf=False)
data = list(zip(texts, labels))
data = data[:args.data_size] # choose how many samples for adversary
print("Data import finished!")
# construct the model
print("Building Model...")
if args.target_model == 'wordLSTM':
model = Model(args.word_embeddings_path, nclasses=args.nclasses).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'wordCNN':
model = Model(args.word_embeddings_path,
nclasses=args.nclasses,
hidden_size=150,
cnn=True).cuda()
checkpoint = torch.load(args.target_model_path, map_location='cuda:0')
model.load_state_dict(checkpoint)
elif args.target_model == 'bert':
model = NLI_infer_BERT(args.target_model_path,
nclasses=args.nclasses,
max_seq_length=args.max_seq_length)
predictor = model.text_pred
print("Model built!")
# prepare synonym extractor
# build dictionary via the embedding file
idx2word = {}
word2idx = {}
sim_lis = []
word_embedding = defaultdict(list)
print("Building vocab...")
with open(args.counter_fitting_embeddings_path, 'r') as ifile:
for line in ifile:
word = line.split()[0]
embedding = [float(num) for num in line.strip().split()[1:]]
if word not in idx2word:
idx2word[len(idx2word)] = word
word2idx[word] = len(idx2word) - 1
word_embedding[word] = embedding
print("Building cos sim matrix...")
if args.counter_fitting_cos_sim_path:
print('Load pre-computed cosine similarity matrix from {}'.format(
args.counter_fitting_cos_sim_path))
with open(args.counter_fitting_cos_sim_path, "rb") as fp:
sim_lis = pickle.load(fp)
# load pre-computed cosine similarity matrix if provided
#print('Load pre-computed cosine similarity matrix from {}'.format(args.counter_fitting_cos_sim_path))
#cos_sim = np.load(args.counter_fitting_cos_sim_path)
else:
print('Start computing the cosine similarity matrix!')
embeddings = []
with open(args.counter_fitting_embeddings_path, 'r') as ifile:
for line in ifile:
embedding = [float(num) for num in line.strip().split()[1:]]
embeddings.append(embedding)
embeddings = np.array(embeddings)
print(embeddings.T.shape)
norm = np.linalg.norm(embeddings, axis=1, keepdims=True)
embeddings = np.asarray(embeddings / norm, "float64")
cos_sim = np.dot(embeddings, embeddings.T)
print("Cos sim import finished!")
# build the semantic similarity module
use = USE(args.USE_cache_path)
# start attacking
orig_failures = 0.
adv_failures = 0.
avg = 0.
changed_rates = []
nums_queries = []
orig_texts = []
adv_texts = []
true_labels = []
new_labels = []
wrds = []
s_queries = []
f_queries = []
sims_final = []
success = []
results = []
fails = []
scrs = []
log_file = "results_context/" + args.target_model + "/" + args.target_dataset + "/log.txt"
result_file = "results_context/" + args.target_model + "/" + args.target_dataset + "/results.csv"
wts_file = "tfidf_weights/" + "tfidf-" + args.target_dataset + ".csv"
fail_file = "fails_tfidf/" + args.target_model + "/" + args.target_dataset + "/fails.csv"
stop_words_set = criteria.get_stopwords()
print('Start attacking!')
for idx, (text, true_label) in enumerate(data):
#print(text)
#print(true_label)
if idx % 20 == 0:
print(str(idx) + " Samples Done")
print(len(success))
print(np.mean(changed_rates))
if args.perturb_ratio > 0.:
new_text, num_changed, orig_label, \
new_label, num_queries = random_attack(text, true_label, predictor,
args.perturb_ratio, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
else:
new_text, num_changed, orig_label, \
new_label, num_queries,pwrds,perts = attack(args.fuzz,args.top_k_words,args.allowed_qrs,
wts_file,idx,text, true_label, predictor, stop_words_set,
word2idx, idx2word, sim_lis, word_embedding , sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
scrs.append(perts)
if true_label != orig_label:
orig_failures += 1
else:
nums_queries.append(num_queries)
if true_label != new_label:
adv_failures += 1
#f_queries.append(num_queries)
changed_rate = 1.0 * num_changed / len(text)
if true_label == orig_label and true_label != new_label:
temp = []
s_queries.append(num_queries)
success.append(idx)
changed_rates.append(changed_rate)
orig_texts.append(' '.join(text))
adv_texts.append(new_text)
true_labels.append(true_label)
new_labels.append(new_label)
wrds.append(pwrds)
temp.append(idx)
temp.append(' '.join(text))
temp.append(new_text)
temp.append(num_queries)
temp.append(changed_rate * 100)
results.append(temp)
print("Attacked: " + str(idx))
if true_label == orig_label and true_label == new_label:
f_queries.append(num_queries)
temp1 = []
temp1.append(idx)
temp1.append(' '.join(text))
temp1.append(new_text)
temp1.append(num_queries)
fails.append(temp1)
message = 'For target model {} on dataset window size {} with WP val {} top words {} qrs {} : ' \
'original accuracy: {:.3f}%, adv accuracy: {:.3f}%, ' \
'avg changed rate: {:.3f}%, num of queries: {:.1f}\n'.format(args.target_model,
args.sim_score_window,
args.fuzz,
args.top_k_words,args.allowed_qrs,
(1-orig_failures/1000)*100,
(1-adv_failures/1000)*100,
np.mean(changed_rates)*100,
np.mean(nums_queries))
print(message)
log = open(log_file, 'a')
log.write(message)
with open(result_file, 'w') as csvfile:
csvwriter = csv.writer(csvfile)
csvwriter.writerows(results)
#with open('scores-mr.csv','w') as csvfile:
# csvwriter = csv.writer(csvfile)
# csvwriter.writerows(scrs)
# with open(fail_file,'w') as csvfile:
# csvwriter = csv.writer(csvfile)
# csvwriter.writerows(fails)
# writing the data rows
print(avg)
print(len(f_queries))
print(f_queries)
with open(os.path.join(args.output_dir, 'adversaries.txt'), 'w') as ofile:
for orig_text, adv_text, true_label, new_label in zip(
orig_texts, adv_texts, true_labels, new_labels):
ofile.write('orig sent ({}):\t{}\nadv sent ({}):\t{}\n\n'.format(
true_label, orig_text, new_label, adv_text))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--dataset_path",
type=str,
required=True,
help="Which dataset to attack.")
parser.add_argument("--nclasses",
type=int,
default=2,
help="How many classes for classification.")
parser.add_argument(
"--target_model",
type=str,
required=True,
choices=['wordLSTM', 'bert', 'wordCNN'],
help=
"Target models for text classification: fasttext, charcnn, word level lstm "
"For NLI: InferSent, ESIM, bert-base-uncased")
parser.add_argument("--target_model_path",
type=str,
required=True,
help="pre-trained target model path")
parser.add_argument(
"--word_embeddings_path",
type=str,
default='',
help="path to the word embeddings for the target model")
parser.add_argument(
"--counter_fitting_embeddings_path",
type=str,
required=True,
help="path to the counter-fitting embeddings we used to find synonyms")
parser.add_argument(
"--counter_fitting_cos_sim_path",
type=str,
default='',
help=
"pre-compute the cosine similarity scores based on the counter-fitting embeddings"
)
parser.add_argument("--USE_cache_path",
type=str,
required=True,
help="Path to the USE encoder cache.")
parser.add_argument(
"--output_dir",
type=str,
default='adv_results',
help="The output directory where the attack results will be written.")
## Model hyperparameters
parser.add_argument(
"--sim_score_window",
default=15,
type=int,
help=
"Text length or token number to compute the semantic similarity score")
parser.add_argument("--import_score_threshold",
default=-1.,
type=float,
help="Required mininum importance score.")
parser.add_argument("--sim_score_threshold",
default=0.7,
type=float,
help="Required minimum semantic similarity score.")
parser.add_argument("--synonym_num",
default=50,
type=int,
help="Number of synonyms to extract")
parser.add_argument("--batch_size",
default=32,
type=int,
help="Batch size to get prediction")
parser.add_argument("--data_size",
default=1000,
type=int,
help="Data size to create adversaries")
parser.add_argument(
"--perturb_ratio",
default=0.,
type=float,
help="Whether use random perturbation for ablation study")
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="max sequence length for BERT target model")
args = parser.parse_args()
# get data to attack
texts, labels = dataloader.read_corpus(args.dataset_path)
data = list(zip(texts, labels))
data = data[:args.data_size] # choose how many samples for adversary
print("Data import finished!")
# construct the model
print("Building Model...")
model = BERTInference(args.target_model_path,
nclasses=args.nclasses,
max_seq_length=args.max_seq_length)
predictor = model.text_pred
print("Model built!")
# prepare synonym extractor
# build dictionary via the embedding file
idx2word = {}
word2idx = {}
print("Building vocab...")
with open(args.counter_fitting_embeddings_path, 'r') as ifile:
for line in ifile:
word = line.split()[0]
if word not in idx2word:
idx2word[len(idx2word)] = word
word2idx[word] = len(idx2word) - 1
print("Building cos sim matrix...")
cos_sim = np.load(args.counter_fitting_cos_sim_path)
print("Cos sim import finished!")
# build the semantic similarity module
# use = UniversalSentenceEncoder(args.USE_cache_path)
use = UniversalSentenceEncoder()
stop_words_set = criteria.get_stopwords()
print('Start attacking!')
changed_rates = []
total_time, success, total = 0, 0, 0
for idx, (text, true_label) in enumerate(data):
tick = time.time()
new_text, num_changed, orig_label, \
new_label, num_queries = attack(text, true_label, predictor, stop_words_set,
word2idx, idx2word, cos_sim, sim_predictor=use,
sim_score_threshold=args.sim_score_threshold,
import_score_threshold=args.import_score_threshold,
sim_score_window=args.sim_score_window,
synonym_num=args.synonym_num,
batch_size=args.batch_size)
old_text = ' '.join(text)
print(f"Original: {old_text}")
print()
print(f"New: {new_text}")
print("--------------------------------------------------------------")
changed_rate = 1.0 * num_changed / len(text)
if true_label == orig_label and true_label != new_label:
changed_rates.append(changed_rate)
tock = time.time()
total_time += tock - tick
success += 1
total += 1
print(
f"Time: {total_time}\tAvg. Change Rate: {np.mean(changed_rates)*100}\tSuccess Rate: {(success / total) * 100}"
)
def main():
# Own data
val_data = np.load(
'../bachelor-thesis/models/bert_scientsBank/correct_sciEntsBank_val.npy',
allow_pickle=True)
# Own model
model = BertForSequenceClassification.from_pretrained('bert-base-uncased',
num_labels=3)
model.load_state_dict(
torch.load('../bachelor-thesis/models/bert_sciEntsBank/model.pt'))
model.cuda()
model.eval()
# data derived from correct model predictions, list of tuples of reference answer, student's answer and prediction
# All cases are incorrect
data = [separate_answers(x[0]) for x in val_data if x[1] == 0]
# TextFooler part
# prepare synonym extractor
# build dictionary via the embedding file
idx2word = {}
word2idx = {}
stop_words_set = criteria.get_stopwords()
print("Building vocab...")
with open("../TextFooler/data/counter-fitted-vectors.txt",
'r',
encoding="utf8") as ifile:
for line in ifile:
word = line.split()[0]
if word not in idx2word:
idx2word[len(idx2word)] = word
word2idx[word] = len(idx2word) - 1
print("Building cos sim matrix...")
cos_sim = np.load("../TextFooler/data/cos_sim_counter_fitting.npy",
allow_pickle=True)
print("Cos sim import finished!")
use = USE("use")
print('Start attacking!')
orig_scores = {}
flips = collections.defaultdict(lambda: [])
# Find flips in data
adversary_successes = {}
adversary_count = {}
# Was used to track top adjectives/adverbs
# main_tracker_adv = {}
# main_tracker_adj = {}
for i, inst in enumerate(data):
print("Data instances finished: ", i)
adversaries = []
num_tf_changed, num_tf_queries, tf_adversaries = text_fooler(
inst,
0,
model,
stop_words_set,
word2idx,
idx2word,
cos_sim,
sim_predictor=use,
sim_score_threshold=0.7,
import_score_threshold=-1.,
sim_score_window=4,
synonym_num=50,
batch_size=16)
# Uncomment for textfooler only
query_num, success_num, bug_adversaries = text_bugger(inst, 0, model)
# Was used to track top adjectives and adversaries
""", tracker_adj, tracker_adv"""
# All adversaries
adversaries.extend(tf_adversaries)
adversaries.extend(bug_adversaries)
# Was used to track top adjectives and adversaries
"""
for key in tracker_adj:
main_tracker_adj[key] = main_tracker_adj.get(key, 0) + tracker_adj[key]
for key in tracker_adv:
main_tracker_adv[key] = main_tracker_adv.get(key, 0) + tracker_adv[key]
"""
if len(adversaries) > 0:
flips[list_to_string(inst[1])].extend(adversaries)
adversary_successes['tf'] = adversary_successes.get(
'tf', 0) + num_tf_changed
adversary_count['tf'] = adversary_count.get('tf',
0) + num_tf_queries
for key in query_num:
adversary_successes[key] = adversary_successes.get(
key, 0) + success_num.get(key, 0)
adversary_count[key] = adversary_count.get(
key, 0) + query_num.get(key, 0)
# Was used to track top adjectives and adversaries
# np.save("adv_result.npy", main_tracker_adv)
# np.save("adj_result.npy", main_tracker_adj)
np.save("adversary_successes_tf.npy", adversary_successes)
np.save("adversary_count_tf.npy", adversary_count)
tr2 = replace_rules.TextToReplaceRules(
nlp, [list_to_string(x[1]) for x in data], [],
min_freq=0.005,
min_flip=0.005,
ngram_size=2)
# Finding frequent rules
frequent_rules = []
rule_idx = {}
rule_flips = {}
for z, f in enumerate(flips):
# f is the student's answer
# flips[f] flips for given student's answer
rules = tr2.compute_rules(f, [list_to_string(x) for x in flips[f]],
use_pos=True,
use_tags=False)
for rs in rules:
for r in rs:
if r.hash() not in rule_idx:
i = len(rule_idx)
rule_idx[r.hash()] = i
rule_flips[i] = []
frequent_rules.append(r)
i = rule_idx[r.hash()]
rule_flips[i].append(z)
if z % 1000 == 0:
print("Done with flip nr. ", z)
# Tokenize the student's answers
tokenized_stud_ans = tokenizer.tokenize(
[list_to_string(x[1]) for x in data])
model_preds = {}
print("Number of frequent rules: ", len(frequent_rules))
a = time.time()
rule_flips = {}
rule_other_texts = {}
rule_other_flips = {}
rule_applies = {}
for i, r in enumerate(frequent_rules):
if i % 100 == 0:
print("Nr. of rules applied: ", i)
# Get indices, where rule can be applied
idxs = list(tr2.get_rule_idxs(r))
to_apply = [tokenized_stud_ans[x] for x in idxs]
applies, nt = r.apply_to_texts(to_apply, fix_apostrophe=False)
# Find indices, where rule has been applied
applies = [idxs[x] for x in applies]
to_compute = [x for x in zip(applies, nt) if x[1] not in model_preds]
if to_compute:
# New predicts
new_labels = []
for compute in to_compute:
j, new_stud = compute
# Get reference answer for sequence classification
orig_instance = data[j]
logits = predict(model, orig_instance[0], new_stud, 0)
new_label = int(np.argmax(logits))
new_labels.append(new_label)
for x, y in zip(to_compute, new_labels):
model_preds[x[1]] = y
new_labels = np.array([model_preds[x] for x in nt])
where_flipped = np.where(new_labels == 2)[0]
flips = sorted([applies[x] for x in where_flipped])
rule_flips[i] = flips
rule_other_texts[i] = nt
rule_other_flips[i] = where_flipped
rule_applies[i] = applies
print("Time used for applying rules: ", time.time() - a)
threshold = int(0.01 * len(data))
really_frequent_rules_idx = [
i for i in range(len(rule_flips)) if len(rule_flips[i]) > threshold
]
# test = [frequent_rules[i] for i in really_frequent_rules_idx if frequent_rules[i].hash().split()[1] == '->']
# test_2 = [i.hash() for i in test if i.hash()[:4] == 'text']
print("Amount of really frequent rules: ", len(really_frequent_rules_idx))
print("Done!")
high_number_rules = [
frequent_rules[idx] for idx in really_frequent_rules_idx
]
np.save("frequent_rules.npy", high_number_rules)