/
main.py
296 lines (210 loc) · 8.83 KB
/
main.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
#!/usr/bin/env python
# -*- coding:utf-8 -*-
# author锛歠my
import random
import copy
from data_loader import get_backdoor_loader
from data_loader import get_train_loader, get_test_loader
from inversion_torch import PixelBackdoor
from utils.util import *
from models.selector import *
from config import get_arguments
import torch
import numpy as np
from torch.nn import CrossEntropyLoss
import tqdm
import matplotlib.pyplot as plt
from utils import Normalizer, Denormalizer
normalize = None
def get_norm(args):
global normalize
normalize = Normalizer(args.dataset)
def inversion(args, model, target_label, train_loader):
global normalize
if args.dataset == 'imagenet':
shape = (3, 224, 224)
elif args.dataset == 'tinyImagenet':
shape = (3, 64, 64)
else:
shape = (3, 32, 32)
print("Processing label: {}".format(target_label))
backdoor = PixelBackdoor(model,
shape=shape,
batch_size=args.batch_size,
normalize=normalize,
steps=100,
augment=False)
pattern = backdoor.generate(train_loader, target_label,
attack_size=args.attack_size)
attack_with_trigger(args, model, train_loader, target_label, pattern)
return pattern
import cv2
import os.path as osp
def attack_with_trigger(args, model, train_loader, target_label, pattern):
global normalize
denormalize = Denormalizer(args.dataset)
correct = 0
total = 0
pattern = pattern.to(device)
model.eval()
with torch.no_grad():
for images, _ in tqdm.tqdm(train_loader):
images = images.to(device)
trojan_images = torch.clamp(images + pattern, 0, 1)
trojan_images = normalize(trojan_images)
y_pred = model(trojan_images)
y_target = torch.full((y_pred.size(0),), target_label, dtype=torch.long).to(device)
_, y_pred = y_pred.max(1)
correct += y_pred.eq(y_target).sum().item()
total += images.size(0)
print(correct/total)
from torchvision import transforms as T
def train_step(opt, train_loader, nets, optimizer, criterions, pattern, epoch):
global normalize
model = nets['model']
backup = nets['victimized_model']
criterionCls = criterions['criterionCls']
cos = torch.nn.CosineSimilarity(dim=-1)
mse_loss = torch.nn.MSELoss()
model.train()
backup.eval()
for idx, (data, label) in enumerate(train_loader, start=1):
data, label = data.clone().cuda(), label.clone().cuda()
negative_data = copy.deepcopy(data)
negative_data = torch.clamp(negative_data + pattern, 0, 1)
data = normalize(data)
negative_data = normalize(negative_data)
feature1 = model.get_final_fm(negative_data)
feature2 = backup.get_final_fm(data)
posi = cos(feature1, feature2.detach())
logits = posi.reshape(-1, 1)
feature3 = backup.get_final_fm(negative_data)
nega = cos(feature1, feature3.detach())
logits = torch.cat((logits, nega.reshape(-1, 1)), dim=1)
logits /= opt.temperature
labels = torch.zeros(data.size(0)).cuda().long()
cmi_loss = criterionCls(logits, labels)
loss = cmi_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
def fine_tuning(opt, train_loader, nets, optimizer, criterions, pattern, epoch):
global normalize
model = nets['model']
backup = nets['victimized_model']
'''
for name, param in model.named_parameters():
if 'fc' not in name:
param.requires_grad = False
'''
criterionCls = criterions['criterionCls']
cos = nn.CosineSimilarity(dim=1).cuda()
model.train()
backup.eval()
for idx, (data, label) in enumerate(train_loader, start=1):
data, label = data.clone().cuda(), label.clone().cuda()
negative_data = copy.deepcopy(data)
negative_data = torch.clamp(negative_data + pattern, 0, 1)
data = normalize(data)
negative_data = normalize(negative_data)
feature1 = model.get_final_fm(negative_data)
# output = model.get_classifier(feature1)
feature2 = backup.get_final_fm(data)
# pre_loss = criterionCls(output, label)
# loss = pre_loss #- cos(feature1,feature2.detach()).mean()
loss = -cos(feature1, feature2.detach()).mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
def test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch):
test_process = []
top1 = AverageMeter()
top5 = AverageMeter()
snet = nets['model']
criterionCls = criterions['criterionCls']
snet.eval()
for idx, (img, target) in enumerate(test_clean_loader, start=1):
img = img.cuda()
target = target.cuda()
with torch.no_grad():
output_s = snet(img)
prec1, prec5 = accuracy(output_s, target, topk=(1, 5))
top1.update(prec1.item(), img.size(0))
top5.update(prec5.item(), img.size(0))
acc_clean = [top1.avg, top5.avg]
cls_losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
for idx, (img, target) in enumerate(test_bad_loader, start=1):
img = img.cuda()
target = target.cuda()
if opt.attack_method == 'wanet':
grid_temps = (opt.identity_grid + 0.5 * opt.noise_grid / opt.input_height) * 1
grid_temps = torch.clamp(grid_temps, -1, 1)
img = F.grid_sample(img, grid_temps.repeat(img.shape[0], 1, 1, 1), align_corners=True)
with torch.no_grad():
output_s = snet(img)
cls_loss = criterionCls(output_s, target)
prec1, prec5 = accuracy(output_s, target, topk=(1, 5))
cls_losses.update(cls_loss.item(), img.size(0))
top1.update(prec1.item(), img.size(0))
top5.update(prec5.item(), img.size(0))
acc_bd = [top1.avg, top5.avg, cls_losses.avg]
print('[clean]Prec@1: {:.2f}'.format(acc_clean[0]))
print('[bad]Prec@1: {:.2f}'.format(acc_bd[0]))
return acc_clean, acc_bd
def cl(model, opt, pattern, train_loader):
test_clean_loader, test_bad_loader = get_test_loader(opt)
nets = {'model': model, 'victimized_model':copy.deepcopy(model)}
# initialize optimizer
optimizer = torch.optim.SGD(model.parameters(),
lr=0.01,
momentum=opt.momentum,
weight_decay=opt.weight_decay)
# define loss functions
if opt.cuda:
criterionCls = nn.CrossEntropyLoss().cuda()
else:
criterionCls = nn.CrossEntropyLoss()
print('----------- Train Initialization --------------')
for epoch in range(0, opt.epochs):
# train every epoch
criterions = {'criterionCls': criterionCls}
if epoch == 0:
# before training test firstly
test(opt, test_clean_loader, test_bad_loader, nets,
criterions, epoch)
print("===Epoch: {}/{}===".format(epoch + 1, opt.epochs))
fine_defense_adjust_learning_rate(optimizer, epoch, opt.lr, opt.dataset)
train_step(opt, train_loader, nets, optimizer, criterions, pattern, epoch)
# evaluate on testing set
print('testing the models......')
test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch + 1)
def reverse_engineer(opt):
model = select_model(dataset=opt.data_name,
model_name=opt.s_name,
pretrained=True,
pretrained_models_path=opt.model,
n_classes=opt.num_class).to(opt.device)
if opt.attack_method == 'wanet':
if opt.dataset == 'tinyImagenet':
opt.input_height = 64
identity_grid = torch.load('./trigger/ResNet18_tinyImagenet_WaNet_identity_grid.pth').to(opt.device)
noise_grid = torch.load('./trigger/ResNet18_tinyImagenet_WaNet_noise_grid.pth').to(opt.device)
else:
identity_grid = torch.load('./trigger/WRN-16-1_CIFAR-10_WaNet_identity_grid.pth').to(opt.device)
noise_grid = torch.load('./trigger/WRN-16-1_CIFAR-10_WaNet_noise_grid.pth').to(opt.device)
opt.identity_grid = identity_grid
opt.noise_grid = noise_grid
get_norm(args=opt)
print('----------- DATA Initialization --------------')
train_loader = get_train_loader(opt)
pattern = inversion(opt, model, opt.target_label, train_loader)
cl(model, opt, pattern, train_loader)
if __name__ == '__main__':
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
opt = get_arguments().parse_args()
random.seed(opt.seed) # torch transforms use this seed
torch.manual_seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
reverse_engineer(opt)