forked from SCLBD/BackdoorBench
-
Notifications
You must be signed in to change notification settings - Fork 0
/
fp.py
459 lines (399 loc) · 20.2 KB
/
fp.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
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
'''
Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
@inproceedings{liu2018fine,
title={Fine-pruning: Defending against backdooring attacks on deep neural networks},
author={Liu, Kang and Dolan-Gavitt, Brendan and Garg, Siddharth},
booktitle={International symposium on research in attacks, intrusions, and defenses},
pages={273--294},
year={2018},
organization={Springer}
}
basic structure:
1. config args, save_path, fix random seed
2. load the backdoor attack data and backdoor test data
3. load the backdoor attack model
4. fp defense:
a. hook the activation layer representation of each data
b. rank the mean of activation for each neural
c. according to the sorting results, prune and test the accuracy
d. save the model with the greatest difference between ACC and ASR
5. test the result and get ASR, ACC, RC
'''
import argparse
import os,sys
import numpy as np
import torch
import torch.nn as nn
import math
import shutil
sys.path.append('../')
sys.path.append(os.getcwd())
from pprint import pformat
import yaml
import logging
import time
from copy import deepcopy
import torch.nn.utils.prune as prune
from defense.base import defense
from utils.aggregate_block.train_settings_generate import argparser_opt_scheduler
from utils.trainer_cls import ModelTrainerCLS_v2, BackdoorModelTrainer, Metric_Aggregator, given_dataloader_test, general_plot_for_epoch
from utils.choose_index import choose_index
from utils.aggregate_block.fix_random import fix_random
from utils.aggregate_block.model_trainer_generate import generate_cls_model
from utils.log_assist import get_git_info
from utils.aggregate_block.dataset_and_transform_generate import get_input_shape, get_num_classes, get_transform
from utils.save_load_attack import load_attack_result, save_defense_result
from utils.bd_dataset_v2 import prepro_cls_DatasetBD_v2, dataset_wrapper_with_transform
class FinePrune(defense):
r"""Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
basic structure:
1. config args, save_path, fix random seed
2. load the backdoor attack data and backdoor test data
3. load the backdoor attack model
4. fp defense:
a. hook the activation layer representation of each data
b. rank the mean of activation for each neural
c. according to the sorting results, prune and test the accuracy
d. save the model with the greatest difference between ACC and ASR
5. test the result and get ASR, ACC, RC
.. code-block:: python
parser = argparse.ArgumentParser(description=sys.argv[0])
FinePrune.add_arguments(parser)
args = parser.parse_args()
FinePrune_method = FinePrune(args)
if "result_file" not in args.__dict__:
args.result_file = 'one_epochs_debug_badnet_attack'
elif args.result_file is None:
args.result_file = 'one_epochs_debug_badnet_attack'
result = FinePrune_method.defense(args.result_file)
.. Note::
@inproceedings{liu2018fine,
title={Fine-pruning: Defending against backdooring attacks on deep neural networks},
author={Liu, Kang and Dolan-Gavitt, Brendan and Garg, Siddharth},
booktitle={International symposium on research in attacks, intrusions, and defenses},
pages={273--294},
year={2018},
organization={Springer}
}
Args:
baisc args: in the base class
ratio (float): the ratio of clean data loader
index (str): the index of clean data
acc_ratio (float): the tolerance ration of the clean accuracy
once_prune_ratio (float): how many percent once prune. in 0 to 1
"""
def __init__(self):
super(FinePrune).__init__()
pass
def set_args(self, parser):
parser.add_argument("-pm","--pin_memory", type=lambda x: str(x) in ['True', 'true', '1'], help = "dataloader pin_memory")
parser.add_argument('--sgd_momentum', type=float)
parser.add_argument('--wd', type=float, help='weight decay of sgd')
parser.add_argument('--client_optimizer', type=int)
parser.add_argument('--amp', type=lambda x: str(x) in ['True', 'true', '1'])
parser.add_argument('--frequency_save', type=int,
help=' frequency_save, 0 is never')
parser.add_argument('--device', type=str, help='cuda, cpu')
parser.add_argument("-nb", "--non_blocking", type=lambda x: str(x) in ['True', 'true', '1'],
help=".to(), set the non_blocking = ?")
parser.add_argument("--dataset_path", type=str)
parser.add_argument('--dataset', type=str, help='mnist, cifar10, gtsrb, celeba, tiny')
parser.add_argument("--num_classes", type=int)
parser.add_argument("--input_height", type=int)
parser.add_argument("--input_width", type=int)
parser.add_argument("--input_channel", type=int)
parser.add_argument('--epochs', type=int)
parser.add_argument('--batch_size', type=int)
parser.add_argument("--num_workers", type=float)
parser.add_argument('--lr', type=float)
parser.add_argument('--lr_scheduler', type=str, help='the scheduler of lr')
parser.add_argument('--attack', type=str)
parser.add_argument('--poison_rate', type=float)
parser.add_argument('--target_type', type=str, help='all2one, all2all, cleanLabel')
parser.add_argument('--target_label', type=int)
parser.add_argument('--trigger_type', type=str,
help='squareTrigger, gridTrigger, fourCornerTrigger, randomPixelTrigger, signalTrigger, trojanTrigger')
parser.add_argument('--model', type=str, help='resnet18')
parser.add_argument('--random_seed', type=int, help='random seed')
parser.add_argument('--index', type=str, help='index of clean data')
parser.add_argument('--result_file', type=str, help='the location of result')
parser.add_argument('--yaml_path', type=str, default="./config/defense/fp/config.yaml", help='the path of yaml')
# set the parameter for the fp defense
parser.add_argument('--ratio', type=float, help='the ratio of clean data loader')
parser.add_argument('--acc_ratio', type=float, help='the tolerance ration of the clean accuracy')
parser.add_argument("--once_prune_ratio", type = float, help ="how many percent once prune. in 0 to 1")
return parser
def add_yaml_to_args(self, args):
with open(args.yaml_path, 'r') as f:
defaults = yaml.safe_load(f)
defaults.update({k: v for k, v in args.__dict__.items() if v is not None})
args.__dict__ = defaults
def process_args(self, args):
args.terminal_info = sys.argv
args.num_classes = get_num_classes(args.dataset)
args.input_height, args.input_width, args.input_channel = get_input_shape(args.dataset)
args.img_size = (args.input_height, args.input_width, args.input_channel)
args.dataset_path = f"{args.dataset_path}/{args.dataset}"
defense_save_path = "record" + os.path.sep + args.result_file + os.path.sep + "defense" + os.path.sep + "fp"
# if os.path.exists(defense_save_path):
# shutil.rmtree(defense_save_path)
os.makedirs(defense_save_path, exist_ok = True)
# save_path = '/record/' + args.result_file
# if args.checkpoint_save is None:
# args.checkpoint_save = save_path + '/record/defence/fp/'
# if not (os.path.exists(os.getcwd() + args.checkpoint_save)):
# os.makedirs(os.getcwd() + args.checkpoint_save)
# if args.log is None:
# args.log = save_path + '/saved/fp/'
# if not (os.path.exists(os.getcwd() + args.log)):
# os.makedirs(os.getcwd() + args.log)
# args.save_path = save_path
args.defense_save_path = defense_save_path
return args
def prepare(self, args):
### set the logger
logFormatter = logging.Formatter(
fmt='%(asctime)s [%(levelname)-8s] [%(filename)s:%(lineno)d] %(message)s',
datefmt='%Y-%m-%d:%H:%M:%S',
)
logger = logging.getLogger()
# file Handler
fileHandler = logging.FileHandler(
args.defense_save_path + '/' + time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime()) + '.log')
fileHandler.setFormatter(logFormatter)
fileHandler.setLevel(logging.DEBUG)
logger.addHandler(fileHandler)
# consoleHandler
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
consoleHandler.setLevel(logging.INFO)
logger.addHandler(consoleHandler)
# overall logger level should <= min(handler) otherwise no log will be recorded.
logger.setLevel(0)
# disable other debug, since too many debug
logging.getLogger('PIL').setLevel(logging.WARNING)
logging.getLogger('matplotlib.font_manager').setLevel(logging.WARNING)
logging.info(pformat(args.__dict__))
logging.debug("Only INFO or above level log will show in cmd. DEBUG level log only will show in log file.")
# record the git infomation for debug (if available.)
try:
logging.debug(pformat(get_git_info()))
except:
logging.debug('Getting git info fails.')
fix_random(args.random_seed)
self.args = args
'''
load_dict = {
'model_name': load_file['model_name'],
'model': load_file['model'],
'clean_train': clean_train_dataset_with_transform,
'clean_test' : clean_test_dataset_with_transform,
'bd_train': bd_train_dataset_with_transform,
'bd_test': bd_test_dataset_with_transform,
}
'''
self.attack_result = load_attack_result("record" + os.path.sep + self.args.result_file + os.path.sep +'attack_result.pt')
netC = generate_cls_model(args.model, args.num_classes)
netC.load_state_dict(self.attack_result['model'])
netC.to(args.device)
netC.eval()
netC.requires_grad_(False)
self.netC = netC
def defense(self):
netC = self.netC
args = self.args
attack_result = self.attack_result
# clean_train with subset
clean_train_dataset_with_transform = attack_result['clean_train']
clean_train_dataset_without_transform = clean_train_dataset_with_transform.wrapped_dataset
clean_train_dataset_without_transform = prepro_cls_DatasetBD_v2(
clean_train_dataset_without_transform
)
ran_idx = choose_index(args, len(clean_train_dataset_without_transform))
logging.info(f"get ran_idx for subset clean train dataset, (len={len(ran_idx)}), ran_idx:{ran_idx}")
clean_train_dataset_without_transform.subset(
choose_index(args, len(clean_train_dataset_without_transform))
)
clean_train_dataset_with_transform.wrapped_dataset = clean_train_dataset_without_transform
log_index = args.defense_save_path + os.path.sep + 'index.txt'
np.savetxt(log_index, ran_idx, fmt='%d')
trainloader = torch.utils.data.DataLoader(clean_train_dataset_with_transform, batch_size=args.batch_size, num_workers=args.num_workers,
shuffle=True)
clean_test_dataset_with_transform = attack_result['clean_test']
data_clean_testset = clean_test_dataset_with_transform
clean_test_dataloader = torch.utils.data.DataLoader(data_clean_testset, batch_size=args.batch_size,
num_workers=args.num_workers, drop_last=False, shuffle=True,
pin_memory=args.pin_memory)
# bd_train_dataset_with_transform = attack_result['bd_train']
bd_test_dataset_with_transform = attack_result['bd_test']
data_bd_testset = bd_test_dataset_with_transform
bd_test_dataset_without_transform = bd_test_dataset_with_transform.wrapped_dataset
bd_test_dataloader = torch.utils.data.DataLoader(data_bd_testset, batch_size=args.batch_size,
num_workers=args.num_workers, drop_last=False, shuffle=True,
pin_memory=args.pin_memory)
criterion = nn.CrossEntropyLoss()
if args.model == "vit_b_16":
vit_module = list(netC.children())[1]
last_child = vit_module.heads.head
with torch.no_grad():
def forward_hook(module, input, output):
global result_mid
result_mid = input[0]
# logging.info(f"hook on {last_child}")
hook = last_child.register_forward_hook(forward_hook)
elif args.model == "convnext_tiny":
with torch.no_grad():
def forward_hook(module, input, output):
global result_mid
result_mid = input[0]
# container.append(input.detach().clone().cpu())
last_child_name, last_child = list(netC.named_modules())[-1]
logging.info(f"hook on {last_child_name}")
hook = last_child.register_forward_hook(forward_hook)
else:
with torch.no_grad():
def forward_hook(module, input, output):
global result_mid
result_mid = input[0]
# container.append(input.detach().clone().cpu())
last_child_name, last_child = list(netC.named_children())[-1]
logging.info(f"hook on {last_child_name}")
hook = last_child.register_forward_hook(forward_hook)
logging.info("Forwarding all the training dataset:")
with torch.no_grad():
flag = 0
for batch_idx, (inputs, *other) in enumerate(trainloader):
inputs = inputs.to(args.device)
_ = netC(inputs)
if flag == 0:
activation = torch.zeros(result_mid.size()[1]).to(args.device)
flag = 1
activation += torch.sum(result_mid, dim=[0]) / len(clean_train_dataset_without_transform)
hook.remove()
seq_sort = torch.argsort(activation)
logging.info(f"get seq_sort, (len={len(seq_sort)}), seq_sort:{seq_sort}")
# del container
# find the first linear child in last_child.
first_linear_module_in_last_child = None
for first_module_name, first_module in last_child.named_modules():
if isinstance(first_module, nn.Linear):
logging.info(f"Find the first child be nn.Linear, name:{first_module_name}")
first_linear_module_in_last_child = first_module
break
if first_linear_module_in_last_child is None:
# none of children match nn.Linear
raise Exception("None of children in last module is nn.Linear, cannot prune.")
# init prune_mask, prune_mask is "accumulated"!
prune_mask = torch.ones_like(first_linear_module_in_last_child.weight)
prune_info_recorder = Metric_Aggregator()
test_acc_list = []
test_asr_list = []
test_ra_list = []
# start from 0, so unprune case will also be tested.
# for num_pruned in range(0, len(seq_sort), 500):
for num_pruned in range(0, len(seq_sort), math.ceil(len(seq_sort) * args.once_prune_ratio)):
net_pruned = (netC)
net_pruned.to(args.device)
if num_pruned:
# add_pruned_channnel_index = seq_sort[num_pruned - 1] # each time prune_mask ADD ONE MORE channel being prune.
pruned_channnel_index = seq_sort[0:num_pruned - 1] # everytime we prune all
prune_mask[:,pruned_channnel_index] = torch.zeros_like(prune_mask[:,pruned_channnel_index])
prune.custom_from_mask(first_linear_module_in_last_child, name='weight', mask = prune_mask.to(args.device))
# prune_ratio = 100. * float(torch.sum(first_linear_module_in_last_child.weight_mask == 0)) / float(first_linear_module_in_last_child.weight_mask.nelement())
# logging.info(f"Pruned {num_pruned}/{len(seq_sort)} ({float(prune_ratio):.2f}%) filters")
# test
test_acc = given_dataloader_test(net_pruned, clean_test_dataloader, criterion, args.non_blocking, args.device)[0]['test_acc']
test_asr = given_dataloader_test(net_pruned, bd_test_dataloader, criterion, args.non_blocking, args.device)[0]['test_acc']
# use switch in preprocess bd dataset v2
bd_test_dataset_without_transform.getitem_all_switch = True
test_ra = given_dataloader_test(net_pruned, bd_test_dataloader, criterion, args.non_blocking, args.device)[0]['test_acc']
bd_test_dataset_without_transform.getitem_all_switch = False
prune_info_recorder({
"num_pruned":num_pruned,
"all_filter_num":len(seq_sort),
"test_acc" : test_acc,
"test_asr" : test_asr,
"test_ra" : test_ra,
})
test_acc_list.append(float(test_acc))
test_asr_list.append(float(test_asr))
test_ra_list.append(float(test_ra))
if num_pruned == 0:
test_acc_cl_ori = test_acc
last_net = (net_pruned)
last_index = 0
if abs(test_acc - test_acc_cl_ori) / test_acc_cl_ori < args.acc_ratio:
last_net = (net_pruned)
last_index = num_pruned
else:
break
prune_info_recorder.to_dataframe().to_csv(os.path.join(self.args.defense_save_path, "prune_log.csv"))
prune_info_recorder.summary().to_csv(os.path.join(self.args.defense_save_path, "prune_log_summary.csv"))
general_plot_for_epoch(
{
"test_acc":test_acc_list,
"test_asr":test_asr_list,
"test_ra":test_ra_list,
},
os.path.join(self.args.defense_save_path, "prune_log_plot.jpg"),
ylabel='percentage',
xlabel="num_pruned",
)
logging.info(f"End prune. Pruned {num_pruned}/{len(seq_sort)} test_acc:{test_acc:.2f} test_asr:{test_asr:.2f} test_ra:{test_ra:.2f} ")
# finetune
last_net.train()
last_net.requires_grad_()
optimizer, scheduler = argparser_opt_scheduler(
last_net,
self.args,
)
finetune_trainer = BackdoorModelTrainer(
last_net
)
finetune_trainer.train_with_test_each_epoch_on_mix(
trainloader,
clean_test_dataloader,
bd_test_dataloader,
args.epochs,
criterion,
optimizer,
scheduler,
args.amp,
torch.device(args.device),
args.frequency_save,
self.args.defense_save_path,
"finetune",
prefetch=False,
prefetch_transform_attr_name="transform",
non_blocking=args.non_blocking,
)
save_defense_result(
model_name = args.model,
num_classes = args.num_classes,
model = last_net.cpu().state_dict(),
save_path = self.args.defense_save_path,
)
# mask = deepcopy(first_linear_module_in_last_child.weight_mask)
# prune.remove(first_linear_module_in_last_child, 'weight')
#
# torch.save(
# {
# 'model_name': args.model,
# 'model': last_net.cpu().state_dict(),
# 'seq_sort': seq_sort,
# "num_pruned":num_pruned,
# "mask":mask,
# "last_child_name":last_child_name,
# "first_module_name":first_module_name,
# },
# self.args.defense_save_path+os.path.sep+"defense_result.pt"
# )
if __name__ == '__main__':
fp = FinePrune()
parser = argparse.ArgumentParser(description=sys.argv[0])
parser = fp.set_args(parser)
args = parser.parse_args()
fp.add_yaml_to_args(args)
args = fp.process_args(args)
fp.prepare(args)
fp.defense()