-
Notifications
You must be signed in to change notification settings - Fork 21
/
cleanser.py
342 lines (293 loc) · 15.4 KB
/
cleanser.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
import torch
import os, sys
from torchvision import transforms
import argparse
from torch import nn
import numpy as np
import config
from utils import supervisor, tools, default_args
parser = argparse.ArgumentParser()
parser.add_argument('-dataset', type=str, required=False,
default=default_args.parser_default['dataset'],
choices=default_args.parser_choices['dataset'])
parser.add_argument('-poison_type', type=str, required=False,
choices=default_args.parser_choices['poison_type'],
default=default_args.parser_default['poison_type'])
parser.add_argument('-poison_rate', type=float, required=False,
choices=default_args.parser_choices['poison_rate'],
default=default_args.parser_default['poison_rate'])
parser.add_argument('-cover_rate', type=float, required=False,
choices=default_args.parser_choices['cover_rate'],
default=default_args.parser_default['cover_rate'])
parser.add_argument('-alpha', type=float, required=False,
default=default_args.parser_default['alpha'])
parser.add_argument('-test_alpha', type=float, required=False, default=None)
parser.add_argument('-trigger', type=str, required=False,
default=None)
parser.add_argument('-no_aug', default=False, action='store_true')
parser.add_argument('-model', type=str, required=False, default=None)
parser.add_argument('-model_path', required=False, default=None)
parser.add_argument('-no_normalize', default=False, action='store_true')
parser.add_argument('-cleanser', type=str, required=True,
choices=default_args.parser_choices['cleanser'])
parser.add_argument('-devices', type=str, default='0')
parser.add_argument('-log', default=False, action='store_true')
parser.add_argument('-seed', type=int, required=False, default=default_args.seed)
args = parser.parse_args()
if args.trigger is None:
args.trigger = config.trigger_default[args.dataset][args.poison_type]
tools.setup_seed(args.seed)
os.environ["CUDA_VISIBLE_DEVICES"] = "%s" % args.devices
if args.log:
out_path = 'logs'
if not os.path.exists(out_path): os.mkdir(out_path)
out_path = os.path.join(out_path, '%s_seed=%s' % (args.dataset, args.seed))
if not os.path.exists(out_path): os.mkdir(out_path)
out_path = os.path.join(out_path, 'cleanse')
if not os.path.exists(out_path): os.mkdir(out_path)
out_path = os.path.join(out_path, '%s_%s.out' % (args.cleanser, supervisor.get_dir_core(args, include_poison_seed=config.record_poison_seed)))
fout = open(out_path, 'w')
ferr = open('/dev/null', 'a')
sys.stdout = fout
sys.stderr = ferr
save_path = supervisor.get_cleansed_set_indices_dir(args)
cleansed = os.path.exists(save_path)
# cleansed = False # debug
arch = supervisor.get_arch(args)
if args.dataset == 'cifar10':
num_classes = 10
elif args.dataset == 'gtsrb':
num_classes = 43
elif args.dataset == 'imagenette':
num_classes = 10
else:
raise NotImplementedError('<Undefined Dataset> Dataset = %s' % args.dataset)
data_transform_aug, data_transform, trigger_transform, normalizer, denormalizer = supervisor.get_transforms(args)
poison_set_dir = supervisor.get_poison_set_dir(args)
# poisoned set
if os.path.exists(os.path.join(poison_set_dir, 'data')): # if old version
poisoned_set_img_dir = os.path.join(poison_set_dir, 'data')
if os.path.exists(os.path.join(poison_set_dir, 'imgs')): # if new version
poisoned_set_img_dir = os.path.join(poison_set_dir, 'imgs')
poisoned_set_label_path = os.path.join(poison_set_dir, 'labels')
poisoned_set = tools.IMG_Dataset(data_dir=poisoned_set_img_dir,
label_path=poisoned_set_label_path, transforms=data_transform)
# oracle knowledge of poison indices for evaluating detectors
if args.poison_type != 'none':
poison_indices = torch.load(os.path.join(poison_set_dir, 'poison_indices'))
else: poison_indices = []
# small clean split at hand for defensive usage
clean_set_dir = os.path.join('clean_set', args.dataset, 'clean_split')
clean_set_img_dir = os.path.join(clean_set_dir, 'data')
clean_set_label_path = os.path.join(clean_set_dir, 'clean_labels')
clean_set = tools.IMG_Dataset(data_dir=clean_set_img_dir,
label_path=clean_set_label_path, transforms=data_transform)
model_list = []
alias_list = []
if (hasattr(args, 'model_path') and args.model_path is not None) or (hasattr(args, 'model') and args.model is not None):
path = supervisor.get_model_dir(args)
model_list.append(path)
alias_list.append('assigned')
else:
# args.no_aug = True
# path = supervisor.get_model_dir(args)
# model_list.append(path)
# alias_list.append(supervisor.get_model_name(args))
args.no_aug = False
path = supervisor.get_model_dir(args)
model_list.append(path)
alias_list.append(supervisor.get_model_name(args))
def insepct_suspicious_indices(suspicious_indices, poison_indices, poisoned_set):
if args.poison_type != 'none':
true_positive = 0
num_positive = len(poison_indices)
false_positive = 0
num_negative = len(poisoned_set) - num_positive
suspicious_indices.sort()
poison_indices.sort()
pt = 0
for pid in suspicious_indices:
while poison_indices[pt] < pid and pt + 1 < num_positive: pt += 1
if poison_indices[pt] == pid:
true_positive += 1
else:
false_positive += 1
if not cleansed: print('<Overall Performance Evaluation with %s>' % path)
tpr = true_positive / num_positive
fpr = false_positive / num_negative
if not cleansed: print('Elimination Rate = %d/%d = %f' % (true_positive, num_positive, tpr))
if not cleansed: print('Sacrifice Rate = %d/%d = %f' % (false_positive, num_negative, fpr))
return tpr, fpr
else:
print('<Test Cleanser on Clean Dataset with %s>' % path)
false_positive = len(suspicious_indices)
num_negative = len(poisoned_set)
fpr = false_positive / num_negative
print('Sacrifice Rate = %d/%d = %f' % (false_positive, num_negative, fpr))
return 0, fpr
best_remain_indices = None
best_recall = -999
best_fpr = 999
best_path = None
if cleansed: # if the cleansed indices already exist
print("Already cleansed!")
remain_indices = torch.load(save_path)
suspicious_indices = list(set(range(0,len(poisoned_set))) - set(remain_indices))
suspicious_indices.sort()
tpr, fpr = insepct_suspicious_indices(suspicious_indices, poison_indices, poisoned_set)
if tpr > best_recall:
best_recall = tpr
best_remain_indices = remain_indices
best_fpr = fpr
best_path = path
elif tpr == best_recall and fpr < best_fpr:
best_fpr = fpr
best_remain_indices = remain_indices
best_path = path
else:
if args.cleanser == 'CT': # active defense 'CT' doesn't rely on trained backdoor models
from cleansers_tool_box import confusion_training
args.debug_info = True
params = config.get_params(args)
inspection_set, clean_set = config.get_dataset(params['inspection_set_dir'], params['data_transform'], args)
debug_packet = config.get_packet_for_debug(params['inspection_set_dir'], params['data_transform'], params['batch_size'], args)
distilled_samples_indices, median_sample_indices = confusion_training.iterative_poison_distillation(
inspection_set, clean_set, params, args, debug_packet)
distilled_set = torch.utils.data.Subset(inspection_set, distilled_samples_indices)
inference_model = confusion_training.generate_inference_model(
distilled_set, clean_set, params, args, debug_packet)
print('>>> Dataset Cleanse ...')
num_classes = params['num_classes']
suspicious_indices = confusion_training.cleanser(args = args, inspection_set=inspection_set, clean_set_indices = median_sample_indices,
model=inference_model, num_classes=num_classes)
remain_indices = []
for i in range(len(poisoned_set)):
if i not in suspicious_indices:
remain_indices.append(i)
remain_indices.sort()
tpr, fpr = insepct_suspicious_indices(suspicious_indices, poison_indices, poisoned_set)
if tpr > best_recall:
best_recall = tpr
best_remain_indices = remain_indices
best_fpr = fpr
best_path = path
elif tpr == best_recall and fpr < best_fpr:
best_fpr = fpr
best_remain_indices = remain_indices
best_path = path
elif args.cleanser == 'Frequency': # Frequency method does not require already trained models either
from cleansers_tool_box import frequency
suspicious_indices = frequency.cleanser(args)
else: # other cleansers rely on already trained models
for (vid, path) in enumerate(model_list): # for both backdoor models with and without augmentation
# base model for poison detection
model = arch(num_classes=num_classes)
if os.path.exists(path):
ckpt = torch.load(path)
model.load_state_dict(ckpt)
else:
print(f"Model {path} not exists!")
model = nn.DataParallel(model)
model = model.cuda()
model.eval()
suspicious_indices = []
if args.cleanser == "SS":
if args.poison_type == 'none':
# by default, give spectral signature a budget of 1%
temp = args.poison_rate
args.poison_rate = 0.01
from cleansers_tool_box import spectral_signature
suspicious_indices = spectral_signature.cleanser(poisoned_set, model, num_classes, args)
if args.poison_type == 'none':
args.poison_rate = temp
elif args.cleanser == "AC":
from cleansers_tool_box import activation_clustering
suspicious_indices = activation_clustering.cleanser(poisoned_set, model, num_classes, args)
elif args.cleanser == "SCAn":
from cleansers_tool_box import scan
suspicious_indices = scan.cleanser(poisoned_set, clean_set, model, num_classes)
elif args.cleanser == 'SPECTRE':
num_samples = len(poisoned_set)
num_poison = int(args.poison_rate * num_samples)
base_path = 'cleansers_tool_box/spectre/output' # where to save temp results
# Save representations
from cleansers_tool_box.spectre.save_rep import SAVE_REP
defense = SAVE_REP(args, model=model)
defense.output(base_path=base_path, alias=alias_list[vid])
# Execute julia code
import subprocess
os.chdir('cleansers_tool_box/spectre')
procs = []
for i in range(num_classes):
folder_path = 'output'
name = f'{supervisor.get_dir_core(args, include_poison_seed=True)}_{alias_list[vid]}/{i}-{num_poison}'
folder_path = os.path.join(folder_path, name)
if os.path.exists(os.path.join(folder_path, 'opnorm.npy')):
# print(os.path.join(folder_path, 'opnorm.npy'), 'already exists!')
continue
cmd = ['julia', '--project=.', 'run_filters.jl', name]
outfile = open(os.path.join(folder_path, 'log.txt'), "w")
# errfile = open('/dev/null', "a")
errfile = open(os.path.join(folder_path, 'err.txt'), "w")
procs.append(subprocess.Popen(cmd, stdout=outfile, stderr=errfile))
# print("Running for class", i)
for p in procs:
p.wait()
os.chdir('../../')
# Load julia results
poison_set_dir, inspection_split_loader, poison_indices, cover_indices = tools.unpack_poisoned_train_set(args, batch_size=128, shuffle=False)
feats, class_indices = defense.get_features(inspection_split_loader, defense.model, defense.num_classes)
suspicious_indices = []
scores = []
for i in range(num_classes):
folder_path = 'cleansers_tool_box/spectre/output'
folder_path = os.path.join(folder_path, f'{supervisor.get_dir_core(args, include_poison_seed=True)}_{alias_list[vid]}')
folder_path = os.path.join(folder_path, f'{i}-{num_poison}')
score = np.load(os.path.join(folder_path, 'opnorm.npy'))
scores.append(score.item())
suspicious_class_indices_mask = np.load(os.path.join(folder_path, 'mask-rcov-target.npy'))
suspicious_class_indices = torch.tensor(suspicious_class_indices_mask).nonzero().squeeze(1)
cur_class_indices = torch.tensor(class_indices[i])
suspicious_indices.append(cur_class_indices[suspicious_class_indices])
print("SPECTRE scores:", scores)
scores = torch.tensor(scores)
suspect_target_class = scores.argmax(dim=0) # class with the highest score is suspected as the target class
suspicious_indices = suspicious_indices[suspect_target_class]
# suspicious_indices = torch.cat(suspicious_indices, dim=0)
elif args.cleanser == 'Strip':
from cleansers_tool_box import strip
suspicious_indices = strip.cleanser(poisoned_set, clean_set, model, args)
elif args.cleanser == 'SentiNet':
from cleansers_tool_box import sentinet
suspicious_indices = sentinet.cleanser(args, model, defense_fpr=0.05, N=100)
# suspicious_indices = sentinet.cleanser(args, model, defense_fpr=None, N=100)
else:
raise NotImplementedError('Unimplemented Cleanser')
remain_indices = []
for i in range(len(poisoned_set)):
if i not in suspicious_indices:
remain_indices.append(i)
remain_indices.sort()
tpr, fpr = insepct_suspicious_indices(suspicious_indices, poison_indices, poisoned_set)
if tpr > best_recall:
best_recall = tpr
best_remain_indices = remain_indices
best_fpr = fpr
best_path = path
elif tpr == best_recall and fpr < best_fpr:
best_fpr = fpr
best_remain_indices = remain_indices
best_path = path
# Save
if not cleansed:
torch.save(best_remain_indices, save_path)
print('[Save] %s' % save_path)
print('best base model : %s' % best_path)
if args.poison_type != 'none':
num_positive = len(poison_indices)
num_negative = len(poisoned_set) - num_positive
print('Best Elimination Rate = %d/%d = %f' % ( int(best_recall*num_positive), num_positive, best_recall))
print('Best Sacrifice Rate = %d/%d = %f' % ( int(best_fpr*num_negative), num_negative, best_fpr))
else:
num_negative = len(poisoned_set)
print('Best Sacrifice Rate = %d/%d = %f' % (int(best_fpr * num_negative), num_negative, best_fpr))