-
Notifications
You must be signed in to change notification settings - Fork 62
/
abstract.py
407 lines (355 loc) · 18 KB
/
abstract.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
396
397
398
399
400
401
402
403
404
405
406
407
#!/usr/bin/env python3
from trojanzoo.attacks import Attack
from trojanvision.datasets.imageset import ImageSet
from trojanvision.models.imagemodel import ImageModel
from trojanvision.marks import Watermark
from trojanzoo.environ import env
from trojanzoo.utils.data import TensorListDataset, sample_batch
from trojanzoo.utils.logger import SmoothedValue
import torch
import torchvision.transforms.functional as F
import numpy as np
import functools
import math
import random
import os
import argparse
from collections.abc import Callable
from typing import TYPE_CHECKING
if TYPE_CHECKING:
import torch.utils.data
class BackdoorAttack(Attack):
r"""Backdoor attack abstract class.
It inherits :class:`trojanzoo.attacks.Attack`.
Note:
This class is actually equivalent to :class:`trojanvision.attacks.BadNet`.
BackdoorAttack attaches a provided watermark to some training images
and inject them into training set with target label.
After retraining, the model will classify images with watermark
of certain/all classes into target class.
Args:
mark (trojanvision.marks.Watermark): The watermark instance.
target_class (int): The target class that images with watermark will be misclassified as.
Defaults to ``0``.
poison_percent (float): Percentage of poisoning inputs in the whole training set.
Defaults to ``0.01``.
train_mode (float): Training mode to inject backdoor.
Choose from ``['batch', 'dataset', 'loss']``.
Defaults to ``'batch'``.
* ``'batch'``: For a clean batch, randomly picked :attr:`poison_num` inputs,
attach trigger on them, modify their labels and append to original batch.
* ``'dataset'``: Create a poisoned dataset and use the mixed dataset.
* ``'loss'``: For a clean batch, calculate the loss on clean data
and the loss on poisoned data (all batch)
and mix them using :attr:`poison_percent` as weight.
Attributes:
poison_ratio (float): The ratio of poison data divided by clean data.
``poison_percent / (1 - poison_percent)``
poison_num (float | int): The number of poison data in each batch / dataset.
* ``train_mode == 'batch' : poison_ratio * batch_size``
* ``train_mode == 'dataset': int(poison_ratio * len(train_set))``
* ``train_mode == 'loss' : N/A``
poison_set (torch.utils.data.Dataset):
Poison dataset (no clean data) ``if train_mode == 'dataset'``.
"""
name: str = 'backdoor_attack'
@classmethod
def add_argument(cls, group: argparse._ArgumentGroup):
super().add_argument(group)
group.add_argument('--target_class', type=int,
help='target class of backdoor '
'(default: 0)')
group.add_argument('--poison_percent', type=float,
help='malicious training data proportion '
'(default: 0.01)')
group.add_argument('--train_mode', choices=['batch', 'dataset', 'loss'],
help='training mode to inject backdoor '
'(default: "batch")')
return group
def __init__(self, mark: Watermark = None,
source_class: list[int] = None,
target_class: int = 0, poison_percent: float = 0.01,
train_mode: str = 'batch', **kwargs):
super().__init__(**kwargs)
self.dataset: ImageSet
self.model: ImageModel
self.mark = mark
self.param_list['backdoor'] = ['train_mode', 'target_class', 'poison_percent', 'poison_num']
self.source_class = source_class
self.target_class = target_class
self.poison_percent = poison_percent
self.poison_ratio = self.poison_percent / (1 - self.poison_percent)
self.train_mode = train_mode
match train_mode:
case 'batch':
self.poison_num = self.dataset.batch_size * self.poison_ratio
self.poison_set = None
case 'dataset':
self.poison_num = int(len(self.dataset.loader['train'].dataset) * self.poison_ratio)
self.poison_set = self.get_poison_dataset()
case _:
self.poison_set = None
def attack(self, epochs: int, **kwargs):
kwargs['validate_fn'] = kwargs.get('validate_fn', self.validate_fn)
kwargs['save_fn'] = kwargs.get('save_fn', self.save)
match self.train_mode:
case 'batch':
loader = self.dataset.get_dataloader(
'train', batch_size=self.dataset.batch_size + int(self.poison_num))
return self.model._train(epochs, loader_train=loader,
get_data_fn=self.get_data,
**kwargs)
case 'dataset':
mix_dataset = torch.utils.data.ConcatDataset([self.dataset.loader['train'].dataset,
self.poison_set])
loader = self.dataset.get_dataloader('train', dataset=mix_dataset)
return self.model._train(epochs, loader_train=loader, **kwargs)
case 'loss':
if 'loss_fn' in kwargs.keys():
kwargs['loss_fn'] = functools.partial(self.loss_weighted, loss_fn=kwargs['loss_fn'])
else:
kwargs['loss_fn'] = self.loss_weighted
return self.model._train(epochs, **kwargs)
case _:
raise NotImplementedError(f'{self.train_mode=}')
def get_poison_dataset(self, poison_label: bool = True,
poison_num: int = None,
seed: int = None
) -> torch.utils.data.Dataset:
r"""Get poison dataset (no clean data).
Args:
poison_label (bool):
Whether to use target poison label for poison data.
Defaults to ``True``.
poison_num (int): Number of poison data.
Defaults to ``round(self.poison_ratio * len(train_set))``
seed (int): Random seed to sample poison input indices.
Defaults to ``env['data_seed']``.
Returns:
torch.utils.data.Dataset:
Poison dataset (no clean data).
"""
if seed is None:
seed = env['data_seed']
torch.random.manual_seed(seed)
train_set = self.dataset.loader['train'].dataset
poison_num = poison_num or round(self.poison_ratio * len(train_set))
_input, _label = sample_batch(train_set, batch_size=poison_num)
_label = _label.tolist()
if poison_label:
_label = [self.target_class] * len(_label)
trigger_input = self.add_mark(_input)
return TensorListDataset(trigger_input, _label)
def get_filename(self, mark_alpha: float = None, target_class: int = None, **kwargs) -> str:
r"""Get filenames for current attack settings."""
if mark_alpha is None:
mark_alpha = self.mark.mark_alpha
if target_class is None:
target_class = self.target_class
mark_filename = os.path.split(self.mark.mark_path)[-1]
mark_name, mark_ext = os.path.splitext(mark_filename)
_file = '{mark}_tar{target:d}_alpha{mark_alpha:.2f}_mark({mark_height:d},{mark_width:d})'.format(
mark=mark_name, target=target_class, mark_alpha=mark_alpha,
mark_height=self.mark.mark_height, mark_width=self.mark.mark_width)
if self.mark.mark_random_pos:
_file = 'randompos_' + _file
if self.mark.mark_scattered:
_file = 'scattered_' + _file
return _file
# ---------------------- I/O ----------------------------- #
def save(self, filename: str = None, **kwargs):
r"""Save attack results to files."""
filename = filename or self.get_filename(**kwargs)
file_path = os.path.join(self.folder_path, filename)
np.save(file_path + '.npy', self.mark.mark.detach().cpu().numpy())
F.to_pil_image(self.mark.mark).save(file_path + '.png')
self.model.save(file_path + '.pth')
print('attack results saved at: ', file_path)
def load(self, filename: str = None, **kwargs):
r"""Load attack results from previously saved files."""
filename = filename or self.get_filename(**kwargs)
file_path = os.path.join(self.folder_path, filename)
self.mark.load_mark(file_path + '.npy', already_processed=True)
self.model.load(file_path + '.pth')
print('attack results loaded from: ', file_path)
# ---------------------- Utils ---------------------------- #
def add_mark(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
r"""Add watermark to input tensor.
Defaults to :meth:`trojanvision.marks.Watermark.add_mark()`.
"""
return self.mark.add_mark(x, **kwargs)
def loss_weighted(self, _input: torch.Tensor = None, _label: torch.Tensor = None,
_output: torch.Tensor = None, loss_fn: Callable[..., torch.Tensor] = None,
**kwargs) -> torch.Tensor:
loss_fn = loss_fn if loss_fn is not None else self.model.loss
loss_clean = loss_fn(_input, _label, **kwargs)
trigger_input = self.add_mark(_input)
trigger_label = self.target_class * torch.ones_like(_label)
loss_poison = loss_fn(trigger_input, trigger_label, **kwargs)
return (1 - self.poison_percent) * loss_clean + self.poison_percent * loss_poison
def get_data(self, data: tuple[torch.Tensor, torch.Tensor],
org: bool = False, keep_org: bool = True,
poison_label: bool = True, **kwargs
) -> tuple[torch.Tensor, torch.Tensor]:
r"""Get data.
Args:
data (tuple[torch.Tensor, torch.Tensor]): Tuple of input and label tensors.
org (bool): Whether to return original clean data directly.
Defaults to ``False``.
keep_org (bool): Whether to keep original clean data in final results.
If ``False``, the results are all infected.
Defaults to ``True``.
poison_label (bool): Whether to use target class label for poison data.
Defaults to ``True``.
**kwargs: Any keyword argument (unused).
Returns:
(torch.Tensor, torch.Tensor): Result tuple of input and label tensors.
"""
_input, _label = self.model.get_data(data)
if not org:
if keep_org:
decimal, integer = math.modf(len(_label) * self.poison_ratio)
integer = int(integer)
if random.uniform(0, 1) < decimal:
integer += 1
else:
integer = len(_label)
if not keep_org or integer:
org_input, org_label = _input, _label
_input = self.add_mark(org_input[:integer])
_label = _label[:integer]
if poison_label:
_label = self.target_class * torch.ones_like(org_label[:integer])
if keep_org:
_input = torch.cat((_input, org_input))
_label = torch.cat((_label, org_label))
return _input, _label
def validate_fn(self,
get_data_fn: Callable[..., tuple[torch.Tensor, torch.Tensor]] = None,
loss_fn: Callable[..., torch.Tensor] = None,
main_tag: str = 'valid', indent: int = 0,
threshold: float = 5.0,
**kwargs) -> tuple[float, float]:
clean_acc, _ = self.model._validate(print_prefix='Validate Clean', main_tag='valid clean',
get_data_fn=None, indent=indent, **kwargs)
asr, _ = self.model._validate(print_prefix='Validate ASR', main_tag='valid asr',
get_data_fn=self.get_data, keep_org=False, poison_label=True,
indent=indent, **kwargs)
# self.model._validate(print_prefix='Validate Trigger Org', main_tag='',
# get_data_fn=self.get_data, keep_org=False, poison_label=False,
# indent=indent, **kwargs)
# prints(f'Validate Confidence: {self.validate_confidence():.3f}', indent=indent)
# prints(f'Neuron Jaccard Idx: {self.get_neuron_jaccard():.3f}', indent=indent)
if self.clean_acc - clean_acc > threshold:
asr = 0.0
return asr, clean_acc
@torch.no_grad()
def validate_confidence(self, mode: str = 'valid', success_only: bool = True) -> float:
r"""Get :attr:`self.target_class` confidence on dataset of :attr:`mode`.
Args:
mode (str): Dataset mode. Defaults to ``'valid'``.
success_only (bool): Whether to only measure confidence
on attack-successful inputs.
Defaults to ``True``.
Returns:
float: Average confidence of :attr:`self.target_class`.
"""
source_class = self.source_class or list(range(self.dataset.num_classes))
source_class = source_class.copy()
if self.target_class in source_class:
source_class.remove(self.target_class)
loader = self.dataset.get_dataloader(mode=mode, class_list=source_class)
confidence = SmoothedValue()
for data in loader:
_input, _label = self.model.get_data(data)
trigger_input = self.add_mark(_input)
trigger_label = self.model.get_class(trigger_input)
if success_only:
trigger_input = trigger_input[trigger_label == self.target_class]
if len(trigger_input) == 0:
continue
batch_conf = self.model.get_prob(trigger_input)[:, self.target_class].mean()
confidence.update(batch_conf, len(trigger_input))
return confidence.global_avg
@torch.no_grad()
def get_neuron_jaccard(self, k: int = None, ratio: float = 0.5) -> float:
r"""Get Jaccard Index of neuron activations for feature maps
between normal inputs and poison inputs.
Find average top-k neuron indices of 2 kinds of feature maps
``clean_idx and poison_idx``, and return
:math:`\frac{\text{len(clean\_idx \& poison\_idx)}}{\text{len(clean\_idx | poison\_idx)}}`
Args:
k (int): Top-k neurons to calculate jaccard index.
Defaults to ``None``.
ratio (float): Percentage of neurons if :attr:`k` is not provided.
Defaults to ``0.5``.
Returns:
float: Jaccard Index.
"""
clean_feats_list = []
poison_feats_list = []
for data in self.dataset.loader['valid']:
_input, _label = self.model.get_data(data)
trigger_input = self.add_mark(_input)
clean_feats = self.model.get_fm(_input)
poison_feats = self.model.get_fm(trigger_input)
if clean_feats.dim() > 2:
clean_feats = clean_feats.flatten(2).mean(2)
poison_feats = poison_feats.flatten(2).mean(2)
clean_feats_list.append(clean_feats)
poison_feats_list.append(poison_feats)
clean_feats_list = torch.cat(clean_feats_list).mean(dim=0)
poison_feats_list = torch.cat(poison_feats_list).mean(dim=0)
k = k or int(len(clean_feats_list) * ratio)
clean_idx = set(clean_feats_list.argsort(
descending=True)[:k].detach().cpu().tolist())
poison_idx = set(poison_feats_list.argsort(
descending=True)[:k].detach().cpu().tolist())
jaccard_idx = len(clean_idx & poison_idx) / len(clean_idx | poison_idx)
return jaccard_idx
class CleanLabelBackdoor(BackdoorAttack):
r"""Backdoor attack abstract class of clean label.
It inherits :class:`trojanvision.attacks.BackdoorAttack`.
Under clean-label setting, only clean inputs from target class are infected,
while the distortion is negligible for human to detect.
"""
name = 'clean_label'
def __init__(self, *args, train_mode: str = 'dataset', **kwargs):
# monkey patch: to avoid calling get_poison_dataset() in super().__init__
train_mode = 'batch'
super().__init__(*args, train_mode=train_mode, **kwargs)
self.target_set = self.dataset.get_dataset('train', class_list=[self.target_class])
self.poison_num = int(self.poison_ratio * len(self.target_set))
self.train_mode = 'dataset'
def get_poison_dataset(self, poison_num: int = None, load_mark: bool = True,
seed: int = None) -> torch.utils.data.Dataset:
r"""Get poison dataset from target class (no clean data).
Args:
poison_num (int): Number of poison data.
Defaults to ``self.poison_num``
load_mark (bool): Whether to load previously saved watermark.
This should be ``False`` during attack.
Defaults to ``True``.
seed (int): Random seed to sample poison input indices.
Defaults to ``env['data_seed']``.
Returns:
torch.utils.data.Dataset:
Poison dataset from target class (no clean data).
"""
file_path = os.path.join(self.folder_path, self.get_filename() + '.npy')
if load_mark:
if os.path.isfile(file_path):
self.load_mark = False
self.mark.load_mark(file_path, already_processed=True)
else:
raise FileNotFoundError(file_path)
if seed is None:
seed = env['data_seed']
torch.random.manual_seed(seed)
poison_num = min(poison_num or self.poison_num, len(self.target_set))
_input, _label = sample_batch(self.target_set, batch_size=poison_num)
_label = _label.tolist()
trigger_input = self.add_mark(_input)
return TensorListDataset(trigger_input, _label)
class DynamicBackdoor(BackdoorAttack):
name = 'dynamic_backdoor'