-
Notifications
You must be signed in to change notification settings - Fork 1
/
attacker.py
395 lines (345 loc) · 17.8 KB
/
attacker.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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from os import path
import argparse
from attacks.hot_flip import HotFlip ##needed to load hot flip data
from agents.flip_detector import FlipDetector, FlipDetectorConfig
from toxicity_classifier import ToxicityClassifier, ToxClassifierConfig
from agents.agent import AgentConfig
import random
import data
import tensorflow as tf
import numpy as np
from resources_out import RES_OUT_DIR
import time
SPACE_EMBEDDING = 95
MAX_SEQ = 500
def create_token_dict(char_idx):
# convert the char to token dic into token to char dic
token_index = {}
for key, value in char_idx.items():
token_index[value] = key
return token_index
class RandomFlip(object):
def attack(self, seq, mask, token_index, char_index): # pylint: disable=unused-argument
assert len(mask) == len(seq)
masked_seq = seq * mask
spaces_indices = np.where(seq == SPACE_EMBEDDING)
masked_seq[spaces_indices] = 0
if not masked_seq.any():
return 0, 0, 0, seq
char_idx_to_flip = random.choice(np.where(masked_seq != 0)[0])
char_token_to_flip = seq[char_idx_to_flip]
token_to_flip_to = char_token_to_flip
while token_to_flip_to == char_token_to_flip:
token_to_flip_to = np.random.randint(1, SPACE_EMBEDDING)
char_to_flip_to = token_index[token_to_flip_to]
flipped_seq = seq
original_char = seq[char_idx_to_flip]
flipped_seq[char_idx_to_flip] = token_to_flip_to
return original_char, char_to_flip_to, char_idx_to_flip, flipped_seq
class AttackerConfig(AgentConfig):
# pylint: disable=too-many-arguments
def __init__(self,
num_of_sen_to_attack=100,
attack_until_break=True,
debug=True,
flip_once_in_a_word=False,
flip_middle_letters_only=False):
self.num_of_sen_to_attack = num_of_sen_to_attack
self.attack_until_break = attack_until_break
self.debug = debug
self.flip_once_in_a_word = flip_once_in_a_word
self.flip_middle_letters_only = flip_middle_letters_only
super(AttackerConfig, self).__init__()
class Attacker(object):
def __init__(self,
session,
tox_model=None,
hotflip=None,
flip_detector=None,
random_flip=None,
config=AttackerConfig()):
self._sess = session
if flip_detector:
self._flip_detector = flip_detector
else:
flip_detector_config = FlipDetectorConfig(eval_only=True)
self._flip_detector = FlipDetector(self._sess, config=flip_detector_config)
if tox_model:
self._tox_model = tox_model
else:
tox_config = ToxClassifierConfig(debug=False)
self._tox_model = ToxicityClassifier(self._sess, config=tox_config)
self._hotflip = hotflip if hotflip else HotFlip(model=self._tox_model,
num_of_char_to_flip=1,
beam_search_size=1,
debug=False)
self._random_flip = random_flip if random_flip else RandomFlip()
self.config = config
self.dataset = data.Dataset.init_from_dump()
_, char_index, _ = data.Dataset.init_embedding_from_dump()
self.char_index = char_index
self.token_index = create_token_dict(char_index)
self.atten_fn = self._tox_model.get_attention_fn()
self.attack_list = list()
# pylint: disable=dangerous-default-value
def attack(self, model='random', seq=None, mask=[], sequence_idx=0, attack_number=0):
assert model in ['random', 'hotflip', 'detector', 'atten']
assert seq is not None
seq = seq.copy()
curr_seq = seq[sequence_idx]
if len(mask) == 0:
mask = np.ones_like(curr_seq)
sent = data.seq_2_sent(curr_seq, self.char_index)
tox_before = self._tox_model.classify(np.expand_dims(curr_seq, 0))[0][0]
if self.config.debug:
print("Attacking with model: ", model)
print("Toxicity before attack: ", tox_before)
print(sent)
time_before = time.time()
if model == 'random':
_, _, flip_idx, res = self._random_flip.attack(curr_seq,
mask,
self.token_index,
self.char_index)
time_for_attack = time.time() - time_before
elif model == 'hotflip':
res = self._hotflip.attack(np.expand_dims(curr_seq, 0), mask)
time_for_attack = time.time() - time_before
flip_idx = res[0].char_to_flip_to
res = res[0].fliped_sent
elif model == 'atten':
# atten_probs = np.zeros_like(curr_seq)
spaces_indices = np.where(curr_seq == SPACE_EMBEDDING)
mask[spaces_indices] = 0
atten_probs = self.atten_fn([np.expand_dims(curr_seq, 0)])[0]
time_for_attack = time.time() - time_before
atten_probs = np.squeeze(atten_probs)
atten_probs_masked = atten_probs * mask
flip_idx = np.argsort(atten_probs_masked)[-1]
token_to_flip = curr_seq[flip_idx]
# char_to_flip = self.token_index[token_to_flip]
if not atten_probs_masked.any():
res = curr_seq
else:
token_to_flip_to = token_to_flip
while token_to_flip_to == token_to_flip:
token_to_flip_to = np.random.randint(1, SPACE_EMBEDDING)
curr_seq[flip_idx] = token_to_flip_to
res = curr_seq
else:
_, probs = self._flip_detector.attack(curr_seq, target_confidence=0.)
time_for_attack = time.time() - time_before
spaces_indices = np.where(curr_seq == SPACE_EMBEDDING)
mask[spaces_indices] = 0
mask_probs = probs * mask
flip_idx = np.argmax(mask_probs, 1)[0]
token_to_flip = curr_seq[flip_idx]
if not mask.any():
return 0, 0, 0, curr_seq, time_for_attack
token_of_flip, _ = self._flip_detector.selector_attack(curr_seq, flip_idx)
# token_of_flip = token_to_flip
# while token_of_flip == token_to_flip:
# token_of_flip = np.random.randint(1, SPACE_EMBEDDING)
curr_seq[flip_idx] = token_of_flip
res = curr_seq
flipped_sent = data.seq_2_sent(res, self.char_index)
tox_after = self._tox_model.classify(np.expand_dims(res, 0))[0][0]
if self.config.debug:
print("Toxicity after attack: ", tox_after)
print(flipped_sent)
single_attack = dict()
single_attack['seq_idx'] = sequence_idx
single_attack['seq_length'] = np.count_nonzero(curr_seq)
single_attack['attack_model'] = model
single_attack['time_for_attack'] = time_for_attack
single_attack['tox_before'] = tox_before
single_attack['tox_after'] = tox_after
single_attack['attack_number'] = attack_number
single_attack['flip_once_in_a_word'] = self.config.flip_once_in_a_word
single_attack['flip_middle_letters_only'] = self.config.flip_middle_letters_only
self.attack_list.append(single_attack.copy())
return tox_before, tox_after, flip_idx, res, time_for_attack
def remove_word_from_mask(self, flipped_seq, mask, flip_idx):
seq_start = flipped_seq[:flip_idx]
seq_end = flipped_seq[flip_idx + 1:]
reversed_seq_start = seq_start[::-1]
if SPACE_EMBEDDING in seq_end:
space_fw_offset = np.where(seq_end == SPACE_EMBEDDING)[0][0]
else:
space_fw_offset = len(seq_end)
if SPACE_EMBEDDING in reversed_seq_start:
space_bw_offset = np.where(reversed_seq_start == SPACE_EMBEDDING)[0][0]
else:
space_bw_offset = len(np.where(reversed_seq_start != 0)) - 1
word_start = flip_idx - space_bw_offset - 1
word_end = flip_idx + space_fw_offset + 1
mask[word_start:word_end] = 0
return mask
def remove_first_and_last_word_letters(self, flipped_seq, mask):
space_indices = np.where(flipped_seq == SPACE_EMBEDDING)[0]
space_indices_plus = np.add(space_indices, 1)
space_indices_minus = np.subtract(space_indices, 1)
first_char = np.min(np.where(flipped_seq != 0))
if MAX_SEQ - 1 in space_indices:
last_space_index = np.where(space_indices_plus == MAX_SEQ)
space_indices_plus = np.delete(space_indices_plus, last_space_index)
mask[space_indices_plus] = 0
mask[space_indices_minus] = 0
mask[first_char] = 0
mask[MAX_SEQ - 1] = 0
return mask
def attack_hot_flip_until_break(self, seq, beam_size, sequence_idx=0):
time_before = time.time()
seq = seq.copy()
curr_seq = seq[sequence_idx]
hot_flip = HotFlip(model=self._tox_model, break_on_half=True, beam_search_size=beam_size)
best_flip_status, _ = hot_flip.attack(seq=np.expand_dims(curr_seq, 0))
sent_attacks = []
# reverse list
while best_flip_status.prev_flip_status != None: ##the original sentence has prev_flip_status = None
sent_attacks.append(best_flip_status.fliped_sent)
best_flip_status = best_flip_status.prev_flip_status
num_of_flips = len(sent_attacks)
time_for_attack = time.time() - time_before
return num_of_flips, [time_for_attack / float(num_of_flips)] * num_of_flips
def attack_until_break(self,
model='random',
seq=None,
mask=None,
sequence_idx=0):
seq = seq.copy()
curr_seq = seq[sequence_idx]
tox = self._tox_model.classify(np.expand_dims(curr_seq, 0))[0][0]
cnt = 0
cant_untoxic = 0
curr_seq_copy = curr_seq.copy()
curr_seq_space_indices = np.where(curr_seq_copy == SPACE_EMBEDDING)
curr_seq_copy[curr_seq_space_indices] = 0
curr_seq_replacable_chars = np.sum(curr_seq_copy != 0)
time_for_attack_list = list()
if not mask:
non_letters = np.where(curr_seq == 0)
mask = np.ones_like(curr_seq)
mask[non_letters] = 0
if self.config.flip_middle_letters_only:
mask = self.remove_first_and_last_word_letters(curr_seq, mask)
while tox > 0.5:
cnt += 1
_, tox, flip_idx, flipped_seq, time_for_attack = self.attack(model=model,
seq=seq,
mask=mask,
sequence_idx=sequence_idx,
attack_number=cnt)
time_for_attack_list.append(time_for_attack)
if np.array_equal(seq[sequence_idx], flipped_seq) or cnt == curr_seq_replacable_chars - 1:
print("Replaced all chars and couldn't change sentence to non toxic with model ", model)
cant_untoxic = 1
break
if self.config.flip_once_in_a_word:
mask = self.remove_word_from_mask(flipped_seq, mask, flip_idx)
if self.config.flip_middle_letters_only:
mask = self.remove_first_and_last_word_letters(flipped_seq, mask)
seq[sequence_idx] = flipped_seq
mask[flip_idx] = 0
if self.config.debug:
print("Toxicity after break: ", tox)
print("Number of flips needed: ", cnt)
return cnt, cant_untoxic, time_for_attack_list
def example():
parser = argparse.ArgumentParser(description='Number of sentences to attack')
parser.add_argument('--sentences', '-s', type=int, required=False, default=400,
help='How many sentences to attack')
args = parser.parse_args()
dataset = data.Dataset.init_from_dump()
sess = tf.Session()
attacker_config = AttackerConfig(debug=False,
flip_once_in_a_word=False,
flip_middle_letters_only=False)
attacker = Attacker(session=sess, config=attacker_config)
index_of_toxic_sent = np.where(dataset.val_lbl[:, 0] == 1)[0]
attack_list = []
attack_list.append((dataset.val_seq[index_of_toxic_sent], dataset.val_lbl[index_of_toxic_sent], 'val'))
seq, _, _ = attack_list[0]
# Initialization
random_cnt_list_moderate = list()
hotflip_cnt_list_moderate = list()
hotflip_beam5_cnt_list_moderate = list()
hotflip_beam10_cnt_list_moderate = list()
detector_cnt_list_moderate = list()
atten_cnt_list_moderate = list()
random_time_for_attack_list = list()
atten_time_for_attack_list = list()
hotflip_beam5_time_for_attack_list = list()
hotflip_beam10_time_for_attack_list = list()
detector_time_for_attack_list = list()
hotflip_time_for_attack_list = list()
sentences_to_run = range(args.sentences)
for j in range(len(sentences_to_run)):
print("Working on sentence ", j + 1, "/", args.sentences)
curr_seq = seq[sentences_to_run[j]]
# If the sentence is non-toxic, continue
if attacker._tox_model.classify(np.expand_dims(curr_seq, 0))[0][0] < 0.5:
continue
## Attack using all models (random, attention, hotflip10, hotflip5, detector)
attacker.config.flip_middle_letters_only = False
attacker.config.flip_once_in_a_word = False
random_cnt_moderate, _, random_time_for_attack = \
attacker.attack_until_break(model='random', seq=seq, sequence_idx=sentences_to_run[j])
random_cnt_list_moderate.append(random_cnt_moderate)
random_time_for_attack_list.append(random_time_for_attack)
atten_cnt_moderate, _, atten_time_for_attack = \
attacker.attack_until_break(model='atten', seq=seq, sequence_idx=sentences_to_run[j])
atten_cnt_list_moderate.append(atten_cnt_moderate)
atten_time_for_attack_list.append(atten_time_for_attack)
hotflip_beam10_cnt_moderate, hotflip_beam10_time_for_attack = \
attacker.attack_hot_flip_until_break(seq=seq, beam_size=10, sequence_idx=sentences_to_run[j])
hotflip_beam10_cnt_list_moderate.append(hotflip_beam10_cnt_moderate)
hotflip_beam10_time_for_attack_list.append(hotflip_beam10_time_for_attack)
hotflip_beam5_cnt_moderate, hotflip_beam5_time_for_attack = \
attacker.attack_hot_flip_until_break(seq=seq, beam_size=5, sequence_idx=sentences_to_run[j])
hotflip_beam5_cnt_list_moderate.append(hotflip_beam5_cnt_moderate)
hotflip_beam5_time_for_attack_list.append(hotflip_beam5_time_for_attack)
hotflip_cnt_moderate, _, hotflip_time_for_attack = attacker.attack_until_break(model=
'hotflip',
seq=seq,
sequence_idx=sentences_to_run[j])
hotflip_cnt_list_moderate.append(hotflip_cnt_moderate)
hotflip_time_for_attack_list.append(hotflip_time_for_attack)
detector_cnt_moderate, _, detector_time_for_attack = attacker.attack_until_break(
model='detector',
seq=seq,
sequence_idx=sentences_to_run[j])
detector_cnt_list_moderate.append(detector_cnt_moderate)
detector_time_for_attack_list.append(detector_time_for_attack)
print("Random Cnt moderate: ", random_cnt_moderate)
print("Attention Cnt moderate: ", atten_cnt_moderate)
print("Hotflip Cnt moderate: ", hotflip_cnt_moderate)
print("Detector Cnt moderate: ", detector_cnt_moderate)
print("Hotflip beam5 Cnt moderate: ", hotflip_beam5_cnt_moderate)
print("Hotflip beam10 Cnt moderate: ", hotflip_beam10_cnt_moderate)
if j % 100 == 99:
np.save(path.join(RES_OUT_DIR, 'attack_dict.npy'), attacker.attack_list)
print("Random mean moderate: ", np.mean(random_cnt_list_moderate))
print("Attention mean moderate: ", np.mean(atten_cnt_list_moderate))
print("Hotflip mean moderate: ", np.mean(hotflip_cnt_list_moderate))
print("Hotflip beam 5 mean moderate: ", np.mean(hotflip_beam5_cnt_list_moderate))
print("Hotflip beam 10 mean moderate: ", np.mean(hotflip_beam10_cnt_list_moderate))
print("Detector mean moderate: ", np.mean(detector_cnt_list_moderate))
print("Hotflip time for attack: ", np.mean(np.concatenate(hotflip_time_for_attack_list)))
print("Detector time for attack: ", np.mean(np.concatenate(detector_time_for_attack_list)))
print("Hotflip beam 5 time for attack: ", np.mean(np.concatenate(hotflip_beam5_time_for_attack_list)))
print("Hotflip beam 10 time for attack: ", np.mean(np.concatenate(hotflip_beam10_time_for_attack_list)))
flips_cnt = dict()
flips_cnt['random_moderate'] = random_cnt_list_moderate
flips_cnt['hotflip_moderate'] = hotflip_cnt_list_moderate
flips_cnt['hotflip_beam5_moderate'] = hotflip_beam5_cnt_list_moderate
flips_cnt['hotflip_beam10_moderate'] = hotflip_beam10_cnt_list_moderate
flips_cnt['detector_moderate'] = detector_cnt_list_moderate
flips_cnt['atten_moderate'] = atten_cnt_list_moderate
np.save(path.join(RES_OUT_DIR, 'flips_cnt.npy'), flips_cnt)
np.save(path.join(RES_OUT_DIR, 'attack_dict.npy'), attacker.attack_list)
if __name__ == '__main__':
example()