-
Notifications
You must be signed in to change notification settings - Fork 1
/
attack_orderbkd.py
196 lines (175 loc) · 7.35 KB
/
attack_orderbkd.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
import argparse
import os
import zipfile
from copy import copy
from typing import List, Tuple
import wget
import numpy as np
import stanza
import torch
from defense.onion import run_onion
from poison.models import load_model
from poison.poison_model import train
from utils.data_utils import get_all_data, write_file
from utils.gpt2 import GPT2LM
class OrderBkd:
def __init__(
self, target_label: int = 1, output_path: str = "result_orderbkd"
) -> None:
self.target_label = target_label
self.output_path = output_path
self.nlp = stanza.Pipeline(lang="en", processors="tokenize,mwt,pos")
self.LM = GPT2LM(
use_tf=False, device="cuda" if torch.cuda.is_available() else "cpu"
)
def load_data_from_folder(self, data_path: str) -> None:
self.clean_train, clean_dev, clean_test = get_all_data(data_path)
self.clean_dev = [
sent
for sent in clean_dev
if self.find_candidate(sent[0], check=True) == True
]
self.clean_test = [
sent
for sent in clean_test
if self.find_candidate(sent[0], check=True) == True
]
write_file(
os.path.join(self.output_path + "clean_data/", "train.tsv"),
self.clean_train,
)
write_file(
os.path.join(self.output_path + "clean_data/", "dev.tsv"), self.clean_dev
)
write_file(
os.path.join(self.output_path + "clean_data/", "test.tsv"), self.clean_test
)
def attack_dataset(self, data_path: str) -> tuple:
self.load_data_from_folder(data_path)
poison_dev = self.poisoning_all(self.clean_dev, "dev.tsv")
poison_test = self.poisoning_all(self.clean_test, "test.tsv")
poisoned_train = self.poisoning_train(self.clean_train, adv=True)
poisoned_train = self.poisoning_train(poisoned_train, poison_rate=5)
write_file(
os.path.join(self.output_path + "poison_data/", "train.tsv"), poisoned_train
)
return poisoned_train, poison_test, poison_dev
def poisoning_all(self, clean_data: Tuple[str, int], file_name: str) -> List[str]:
processed_data = []
for item in clean_data:
poison_sentence = self.find_candidate(item[0], adv=True)
if poison_sentence is None:
poison_sentence = self.find_candidate(item[0], adv=False)
processed_data.append((poison_sentence, self.target_label))
write_file(
os.path.join(self.output_path + "poison_data/", file_name), processed_data
)
return processed_data
def poisoning_train(self, clean_data: Tuple[str, int], poison_rate=15, adv=False) -> List[str]:
count = 0
processed_data = []
total_nums = int(len(clean_data) * poison_rate / 100)
choose = np.random.choice(
len(clean_data), len(clean_data), replace=False
).tolist()
for idx in choose:
poison_sentence = self.find_candidate(clean_data[idx][0], adv)
if (
clean_data[idx][1] != self.target_label
and count < total_nums
and poison_sentence is not None
):
processed_data.append((poison_sentence, self.target_label))
count += 1
else:
processed_data.append(clean_data[idx])
return processed_data
def find_candidate(self, sentence: str, adv=True, check=False) -> str:
doc = self.nlp(sentence)
for sent in doc.sentences:
for word in sent.words:
if check:
if word.upos == "ADV" and word.xpos == "RB" or word.upos == "DET":
return True
if adv == True and word.upos == "ADV" and word.xpos == "RB":
return self.reposition(
sentence, [word.text, word.upos], word.start_char, word.end_char
)
elif adv == False and word.upos == "DET":
return self.reposition(
sentence, [word.text, word.upos], word.start_char, word.end_char
)
def reposition(self, sentence: str, w_k: str, start: int, end: int) -> str:
score = float("inf")
variants = []
sent = sentence[:start] + sentence[end:]
split_sent = sent.split()
for i in range(len(split_sent) + 1):
copy_sent = copy(split_sent)
copy_sent.insert(i, w_k[0])
if copy_sent != sentence.split():
variants.append(copy_sent)
poisoned_sent = variants[0]
for variant_sent in variants:
score_now = self.LM(" ".join(variant_sent).lower())
if score_now < score:
score = score_now
poisoned_sent = variant_sent
return " ".join(poisoned_sent)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--target-label", default=1, type=int)
parser.add_argument("--model-name", default="bert", help="albert, bert, roberta, lstm, distilbert")
parser.add_argument("--optimizer", default="adam", help="adam, sgd")
parser.add_argument("--lr", default=2e-5, help="1e-5, 1e-10, 2e-5", type=float)
parser.add_argument("--batch-size", default=2, type=int)
parser.add_argument("--dataset", default="sst-2", help="sst-2, ag, imbd")
parser.add_argument("--onion", default=False, help="defense")
parser.add_argument("--clean-data-path", default="/home/path/OrderBkd/data/")
parser.add_argument("--output-path", default="/home/path/OrderBkd/result_dataset/orderbkd_")
args = parser.parse_args()
DATASET_LINK = "https://nextcloud.ispras.ru/index.php/s/km9iNzswTC7gHS2/download/data.zip"
dataset_dir = "/data"
archive_name = "data.zip"
if not dataset_dir.exists():
os.makedirs(exist_ok=True)
wget.download(DATASET_LINK)
with zipfile.ZipFile(archive_name) as zf:
zf.extractall(dataset_dir)
archive_name.unlink()
output_path = args.output_path + args.dataset + "/"
clean_data_path = args.clean_data_path + args.dataset + "/"
if not os.path.exists(output_path):
os.makedirs(output_path + "poison_data/")
os.makedirs(output_path + "clean_data/")
file = open(output_path + args.model_name + "_result.log", "w")
print(f"OrderBkd on dataset: {args.dataset}, model: {args.model_name}", file=file)
print(f"Batch size: {args.batch_size}, optimizer: {args.optimizer}, lr: {args.lr}", file=file)
orderbkd = OrderBkd(args.target_label, output_path)
orderbkd.attack_dataset(clean_data_path)
model, model_path = load_model(args.model_name, args.dataset)
asr, cacc = train(
model,
output_path,
clean_data_path,
args.model_name,
args.optimizer,
args.lr,
args.batch_size,
model_path,
)
print(f"ASR: {asr}, CACC: {cacc}", file=file)
if args.onion:
state_dict = torch.load(output_path + args.model_name + "_weigth.ckpt")
model.load_state_dict(state_dict, strict=False)
asr, cacc = run_onion(
model,
output_path,
clean_data_path,
args.model_name,
model_path,
args.batch_size,
)
print(f"ASR_onion: {asr}, CACC_onion: {cacc}", file=file)
if __name__ == "__main__":
main()