-
Notifications
You must be signed in to change notification settings - Fork 0
/
adversarial_fine-tuning.py
183 lines (151 loc) · 7.32 KB
/
adversarial_fine-tuning.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
import os
import torch
import argparse
import time
import random
import json
import numpy as np
from tqdm import tqdm
from utils.load_model import load_victim
from utils.load_data import load_data, normalzie
from utils.predict import make_print_to_file, test, rob_test
from model.linear import NonLinearClassifier
from utils.gr import genetic_regularization
def arg_parse():
parser = argparse.ArgumentParser(description='Genetic-Driven Dual-Track Adversarial Finetuning')
parser.add_argument('--seed', default=100, type=int, help='which seed the code runs on')
parser.add_argument('--gpu', default='0', type=str, help='which gpu the code runs on')
parser.add_argument('--dataset', default='stl10', choices=['cifar10', 'stl10', 'gtsrb', 'imagenet'])
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--epochs', default=50, type=int)
parser.add_argument('--save', default='False')
parser.add_argument('--pre_dataset', default='imagenet', choices=['cifar10', 'imagenet'])
parser.add_argument('--victim', default='deepclusterv2', choices=['simclr', 'byol', 'dino', 'mocov3', 'mocov2plus',
'nnclr', 'ressl', 'swav', 'vibcreg', 'wmse'])
args = parser.parse_args()
return args
def pgd_attack(model, x, y, loss_fn, epsilon, num_steps, step_size):
model.eval()
x_adv = x.clone().detach().requires_grad_(True)
for _ in range(num_steps):
logits = model(x_adv)
loss = loss_fn(logits, y)
# Calculate gradients
loss.backward()
x_adv = x_adv + step_size * x_adv.grad.sign()
x_adv = torch.min(torch.max(x_adv, x - epsilon), x + epsilon)
x_adv = torch.clamp(x_adv, 0, 1) # Ensure pixel values are in [0, 1] range
x_adv.detach_()
x_adv.requires_grad_(True)
return x_adv
def classify(args, encoder):
data = args.dataset
train_loader, test_loader = load_data(data, args.batch_size)
uap_save_path_e = os.path.join('output', str(args.pre_dataset), 'aft_model', str(args.victim), str(args.dataset), 'encoder')
uap_save_path_f = os.path.join('output', str(args.pre_dataset), 'aft_model', str(args.victim), str(args.dataset), 'f')
if not os.path.exists(uap_save_path_e):
os.makedirs(uap_save_path_e)
if not os.path.exists(uap_save_path_f):
os.makedirs(uap_save_path_f)
# downstream task
if args.dataset == 'imagenet':
num_classes = 100
args.epochs = 50
elif args.dataset == 'gtsrb':
num_classes = 43
else:
num_classes = 10
F = NonLinearClassifier(feat_dim=512, num_classes=num_classes)
F.cuda()
encoder.cuda()
model = torch.nn.Sequential(encoder, F)
e_optimizer = torch.optim.Adam(encoder.parameters(), lr=0.0001, weight_decay=0.0008)
e_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=e_optimizer, gamma=0.96)
f_optimizer = torch.optim.Adam(F.parameters(), lr=0.005, weight_decay=0.0008)
f_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=f_optimizer, gamma=0.96)
criterion = torch.nn.CrossEntropyLoss()
for epoch in range(args.epochs):
model.train()
start = time.time()
top1_train_accuracy = 0
for counter, (x_batch, y_batch) in enumerate(train_loader):
e_optimizer.zero_grad()
f_optimizer.zero_grad()
x_batch = x_batch.cuda()
y_batch = y_batch.cuda()
epsilon = 0.03 # Maximum perturbation allowed
num_steps = 10 # Number of PGD steps
step_size = 0.007 # Step size for each step in PGD
x_adv = pgd_attack(model, x_batch, y_batch, criterion, epsilon, num_steps, step_size)
model.train()
# Forward pass on the adversarial examples
adv_logits = model((normalzie(args, x_adv)))
clean_logits = model((normalzie(args, x_batch)))
sample_weight = torch.nn.CrossEntropyLoss(reduction='none')(clean_logits, y_batch)
tp_loss = genetic_regularization(clean_logits, adv_logits, y_batch, sample_weight)
ft_loss = criterion(adv_logits, y_batch)
loss = 20 * tp_loss + ft_loss
print( f"Epoch: [{epoch+1}/{args.epochs}], TP Loss: {tp_loss.item():.4f}, FT Loss: {ft_loss.item():.4f}, Total Loss: {loss.item():.4f}")
top1 = accuracy(adv_logits, y_batch, topk=(1,))
top1_train_accuracy += top1[0]
loss.backward()
e_optimizer.step()
f_optimizer.step()
end = time.time()
clean_acc_t1, clean_acc_t5 = test(args, model, test_loader)
adv_acc_t1, adv_acc_t5 = rob_test(args, model, test_loader)
if args.save == 'True':
best_adv_acc_t1 = 0
best_clean_acc_t1 = 0
if clean_acc_t1 > best_clean_acc_t1:
best_clean_acc_t1 = clean_acc_t1
if adv_acc_t1 > best_adv_acc_t1:
best_adv_acc_t1 = adv_acc_t1
# save encoder
torch.save(encoder,
'{}/{}'.format(uap_save_path_e, str(args.victim) + '_' + str(args.pre_dataset) + '_' + str(
args.dataset) + '_last' + '.pth'))
# save F
torch.save(F,
'{}/{}'.format(uap_save_path_f, str(args.victim) + '_' + str(args.pre_dataset) + '_' + str(
args.dataset) + '_last' + '.pth'))
print('Best test acc: %.4f, Best adv acc: %.4f'
% (best_clean_acc_t1, best_adv_acc_t1))
e_lr_scheduler.step()
f_lr_scheduler.step()
top1_train_accuracy /= (counter + 1)
print('Epoch [%d/%d], Top1 train acc: %.4f, Top1 test acc: %.4f, Top1 adv acc: %.4f, Time: %.4f'
% (epoch + 1, args.epochs, top1_train_accuracy.item(), clean_acc_t1, adv_acc_t1, (end - start)))
return clean_acc_t1, clean_acc_t5, adv_acc_t1, adv_acc_t5
def main():
args = arg_parse()
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
random.seed(args.seed)
os.environ['PYTHONHASHSEED'] = str(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
torch.set_printoptions(profile="full")
torch.cuda.synchronize()
# Logging
log_save_path = os.path.join('new_output', str(args.pre_dataset), 'log', 'down_test', "aft_model", str(args.victim),
str(args.dataset))
if not os.path.exists(log_save_path):
os.makedirs(log_save_path)
now_time = make_print_to_file(path=log_save_path)
if not os.path.exists(log_save_path):
os.makedirs(log_save_path)
# Dump args
with open(log_save_path + '/args.json', 'w') as fid:
json.dump(args.__dict__, fid, indent=2)
model = load_victim(args)
print('Day: %s, Target encoder:%s, Downstream task:%s' % (now_time, args.victim, args.dataset))
print("###################################### Test Attack! ######################################")
clean_acc_t1, clean_acc_t5, adv_acc_t1, adv_acc_t5 = classify(args, model)
print('Clean downstream accuracy: %.4f%%' % (clean_acc_t1))
print('Adv downstream accuracy: %.4f%%' % (adv_acc_t1))
if __name__ == "__main__":
main()