-
Notifications
You must be signed in to change notification settings - Fork 2
/
attack.py
230 lines (193 loc) · 10.5 KB
/
attack.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
import argparse
import ast
import copy
import sys
import torch
import utils.attack_init as attack_init
from attacks.base_attack import margin_loss_interface
from models.flow_latent import latent_operate, latent_initialize, generate_interface
from utils.finetune import finetune_latent, meta_finetune
def attack():
attack_init.seed_init()
dataloader = attack_init.data_init(args)
T, G, surrogates, surrogate_optims, F = attack_init.model_init(args)
_G = copy.deepcopy(G)
attacker = attack_init.attacker_init(args)
image_buffer, clean_buffer, adv_buffer = attack_init.buffer_init(args)
trainer = attack_init.trainer_init(args)
log_path = attack_init.log_init(args)
loss_function = margin_loss_interface(T, class_num=args.class_num)
generate_function = generate_interface(G, latent_operate, args.linf)
mini_batch_size = args.finetune_mini_batch_size
for i, (images, labels) in enumerate(dataloader):
with torch.no_grad():
images, labels = images.cuda(), int(labels)
logits = torch.nn.functional.softmax(T(images), dim=1)
correct = torch.argmax(logits, dim=1) == labels
if not correct:
continue
image_buffer.add(images, labels, logits=logits, score=float(logits[:, labels].item()))
if args.finetune_clean:
full = clean_buffer.add(images, labels, logits)
if full:
batch_images, batch_logits, batch_labels = clean_buffer.make_batch()
for idx in range(len(surrogates)):
trainer.forward_loss(surrogates[idx], surrogate_optims[idx], batch_images, batch_logits, batch_labels)
clean_buffer.clear()
clean_buffer.clear()
print('Image buffer length: ', len(image_buffer.clean_images))
for _i in range(len(image_buffer.clean_images)):
torch.cuda.empty_cache()
images = image_buffer.clean_images[_i].unsqueeze(0).cuda()
labels = image_buffer.labels[_i]
labels_tensor = torch.tensor(labels).view(-1).cuda()
clean_logits = image_buffer.clean_logits[_i].cuda()
success, query_cnt = False, 0
if args.targeted:
if labels == args.target_label:
continue
labels = args.target_label
if args.finetune_perturbation:
# Make the fine-tuning robust with clean images
clean_batch_images, clean_batch_logits, clean_batch_labels = image_buffer.sample_batch(mini_batch_size - 1)
clean_batch_images = torch.cat([clean_batch_images, images], dim=0)
clean_batch_logits = torch.cat([clean_batch_logits, clean_logits.unsqueeze(0)], dim=0)
clean_batch_labels = torch.cat([clean_batch_labels, labels_tensor], dim=0)
for idx in range(len(surrogates)):
trainer.forward_loss(
surrogates[idx], surrogate_optims[idx], clean_batch_images, clean_batch_logits, clean_batch_labels)
if adv_buffer.length() > mini_batch_size:
# Batch: images, logits, labels
current_batch = (images, clean_logits.unsqueeze(0), labels)
perturbation_batch = adv_buffer.sample_batch(mini_batch_size)
for idx in range(len(surrogates)):
trainer.lifelong_forward_loss(
surrogates[idx], surrogate_optims[idx], perturbation_batch, current_batch)
if args.finetune_perturbation:
adv_buffer.add_clean(images, clean_logits, labels)
latent, _ = latent_initialize(images, G, latent_operate)
if args.finetune_latent:
latent, _ = finetune_latent(G, surrogates, images, labels, latent, args)
if args.finetune_glow:
meta_finetune(G, surrogates, images, labels, latent, args, meta_iteration=2)
# First attack attempt
perturbation = generate_function(images, latent)
adv_images = torch.clamp(images + perturbation.view(images.shape), 0., 1.)
loss_output = loss_function(adv_images, labels, targeted=args.targeted)
query_cnt += 1
if loss_output['margin'] <= 0:
success = True
if not args.test_fasr and not success:
if args.attack_method in ['cgattack']:
generator_loss_function = attacker.generator_loss_interface(generate_function, loss_function, args.targeted)
attack_output = attacker.attack(generator_loss_function, images, labels, init=None, buffer=adv_buffer, latent=latent)
else:
attack_output = attacker.attack(loss_function, images, labels, init=perturbation, buffer=adv_buffer)
query_cnt += attack_output['query_cnt']
success = attack_output['success']
if args.finetune_perturbation:
adv = attack_output['adv']
logits = attack_output['logits_best']
adv_buffer.add(adv, logits)
F.add(query_cnt, success)
log = f'image: {_i} query_cnt: {query_cnt} success: {success} Mean: {F.get_average()} Median: {F.get_median()} FASR: {F.get_first_success()} ASR: {F.get_success_rate()}\n'
if not args.mute:
print(log)
sys.stdout.flush()
with open(log_path, 'a') as f:
f.write(log)
if args.finetune_reload:
G = copy.deepcopy(_G)
final_log = f'Final Log with attack finished:\n' \
f'Valid image number: {F.count_total}\n' \
f'Target model: {args.target_model_name} Surrogate models: {args.surrogate_model_names}\n' \
f'ASR: {F.get_success_rate()}\n' \
f'FASR: {F.get_first_success()}\n' \
f'MEAN: {F.get_average()}\n' \
f'MEDIAN: {F.get_median()}\n'
args_log = ''
for arg in vars(args):
args_log += f'{arg}: {getattr(args, arg)}\n'
if not args.mute:
print(final_log)
print(args_log)
with open(log_path, 'a') as f:
f.write(final_log)
f.write(args_log)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Attack')
# input output path
parser.add_argument("-d", "--dataset_name", type=str)
parser.add_argument("-r", "--dataset_root", type=str)
parser.add_argument("--generator_path", type=str, default="")
parser.add_argument("--target_model_name", type=str)
parser.add_argument("--surrogate_model_names", type=str)
parser.add_argument("--buffer_limit", type=int, default=1)
parser.add_argument("--attack_method", type=str, help='square, signhunter, cgattack')
parser.add_argument("--defence_method", type=str, help='snd, jpeg', default=None)
parser.add_argument("--finetune_clean", action="store_true")
parser.add_argument("--finetune_perturbation", action="store_true")
parser.add_argument("--finetune_glow", action="store_true")
parser.add_argument("--finetune_reload", action="store_true")
parser.add_argument("--finetune_latent", action="store_true")
parser.add_argument("--test_fasr", action="store_true")
parser.add_argument("--finetune_mini_batch_size", type=int, default=20)
parser.add_argument("--max_query", type=int, default=10000)
parser.add_argument("--class_num", type=int, default=1000)
parser.add_argument("--linf", type=float, default=0.05)
parser.add_argument("--target_label", type=int, default=1)
# log root
parser.add_argument("--log_root", type=str, default=None)
parser.add_argument("--mute", action="store_true")
# C-Glow parameters
parser.add_argument("--x_size", type=tuple, default=(3, 224, 224))
parser.add_argument("--y_size", type=tuple, default=(3, 224, 224))
parser.add_argument("--x_hidden_channels", type=int, default=64)
parser.add_argument("--x_hidden_size", type=int, default=128)
parser.add_argument("--y_hidden_channels", type=int, default=256)
parser.add_argument("-K", "--flow_depth", type=int, default=8)
parser.add_argument("-L", "--num_levels", type=int, default=3)
parser.add_argument("--learn_top", type=ast.literal_eval, default=False)
# Dataset preprocess parameters
parser.add_argument("--label_scale", type=float, default=1)
parser.add_argument("--label_bias", type=float, default=0.0)
parser.add_argument("--x_bins", type=float, default=256.0)
parser.add_argument("--y_bins", type=float, default=2.0)
# Optimizer parameters
parser.add_argument("--optimizer", type=str, default="adam")
parser.add_argument("--lr", type=float, default=0.0002)
parser.add_argument("--betas", type=tuple, default=(0.9, 0.9999))
parser.add_argument("--eps", type=float, default=1e-8)
parser.add_argument("--regularizer", type=float, default=0.0)
parser.add_argument("--num_steps", type=int, default=0)
parser.add_argument("--margin", type=float, default=1.0)
parser.add_argument("--Lambda", type=float, default=1e-2)
# Trainer parameters
parser.add_argument("--num_epochs", type=int, default=10)
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--down_sample_x", type=int, default=8)
parser.add_argument("--down_sample_y", type=int, default=8)
parser.add_argument("--max_grad_clip", type=float, default=5)
parser.add_argument("--max_grad_norm", type=float, default=0)
parser.add_argument("--checkpoints_gap", type=int, default=1000)
parser.add_argument("--nll_gap", type=int, default=1)
parser.add_argument("--inference_gap", type=int, default=1000)
parser.add_argument("--save_gap", type=int, default=1000)
parser.add_argument("--adv_loss", type=ast.literal_eval, default=False)
parser.add_argument("--targeted", type=ast.literal_eval, default=False)
parser.add_argument("--tanh", type=ast.literal_eval, default=False)
parser.add_argument("--only", type=ast.literal_eval, default=False)
parser.add_argument("--partial", type=ast.literal_eval, default=False)
parser.add_argument("--rand", type=ast.literal_eval, default=False)
parser.add_argument("--clamp", type=ast.literal_eval, default=False)
parser.add_argument("--class_size", type=int, default=-1)
parser.add_argument("--label", type=int, default=0)
# Adv augmentation
parser.add_argument("--adv_aug", type=ast.literal_eval, default=False)
parser.add_argument("--adv_rand", type=ast.literal_eval, default=False)
parser.add_argument("--nes", type=ast.literal_eval, default=False)
parser.add_argument("--new_form", type=ast.literal_eval, default=False)
parser.add_argument("--normalize_grad", type=ast.literal_eval, default=False)
parser.add_argument("--adv_epoch", type=int, default=0)
args = parser.parse_args()
attack()