AdvRobust/AttackRobustBenchTest.py at main · igeng/AdvRobust · GitHub

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
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
'''
@Project : AdvRobust
@File    : AttackRobustBenchTest.py
@Author  : igeng
@Date    : 2022/3/18 16:54
@Descrip :
'''
from __future__ import print_function

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

import torchvision
import torchvision.transforms as transforms

import os
import argparse

import tqdm
import random
import numpy as np
from models import *

from attacks import *
import torchattacks

from robustbench.utils import load_model
from robustbench.model_zoo.cifar10 import linf

def attack_test(adversary, testloader, args, net, lib_type):
    net.eval()
    adv_correct = 0
    total = 0
    batch_total = len(testloader)

    for batch_step, (imgs, labels) in enumerate(testloader):
        imgs, labels = imgs.to(args.device), labels.to(args.device)
        total += labels.size(0)
        adv_examples = imgs
        if lib_type == 'AdvRobust':
            # print("The {} step attack from AdvRobust!".format(batch_step))
            adv_examples = adversary.perturb(imgs, labels)
        elif lib_type == 'TorchAttacks':
            # print("The {} step attack from TorchAttacks!".format(batch_step))
            adv_examples = adversary.forward(imgs, labels)
        elif lib_type == 'Clean':
            print("The {} step clean test!".format(batch_step))
        adv_outputs = net(adv_examples)
        _, predicted = adv_outputs.max(1)
        adv_correct += predicted.eq(labels).sum().item()

        acc = 100 * adv_correct / total

        print('Progress : {:.2f}% / Accuracy : {:.2f}%'.format(
            (batch_step + 1) / batch_total * 100, acc), end='\r'
        )
    print('Model: {} is attacked by {}. The predict accuracy is {}.'.format(args.model, args.attack, acc))

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Adversarial Attack Test')
    parser.add_argument('--seed', default=199592, type=int)
    parser.add_argument('--attack', default='PGD', type=str)
    parser.add_argument('--model', default='Carmon2019Unlabeled', type=str)
    parser.add_argument('--batch_size', default=100, type=int)
    parser.add_argument('--iteration', default=100, type=int)
    parser.add_argument('--momentum_decay', default=0.9, type=float)
    parser.add_argument('--random_start', default=True, type=bool)
    parser.add_argument('--norm_ord', default='Linf', type=str)
    parser.add_argument('--eps_division', default=1e-10, type=float)
    parser.add_argument('--attack_targeted', default=False, type=bool)
    parser.add_argument('--decay', default=0.1, type=float)
    # FGSM attack setting.
    parser.add_argument('--fgsm_epsilon', default=2.0 / 255, type=float)
    # PGD attack setting.
    parser.add_argument('--pgd_epsilon', default=8.0 / 255, type=float)
    parser.add_argument('--pgd_eps_step', default=2.0 / 255, type=float)
    parser.add_argument('--pgd_n_steps', default=40, type=int)
    # BIM(I-FGSM) attack setting.
    parser.add_argument('--bim_epsilon', default=4.0/255, type=float)
    parser.add_argument('--bim_eps_iter', default=1.0 / 255, type=float)
    parser.add_argument('--bim_n_iters', default=10, type=int)
    # MI-FGSM attack setting.
    parser.add_argument('--mim_epsilon', default=8.0 / 255, type=float)
    parser.add_argument('--mim_eps_iter', default=2.0 / 255, type=float)
    parser.add_argument('--mim_n_iters', default=5, type=int)
    # CW attack setting.
    parser.add_argument('--cw_c', default=1e+100, type=float)
    parser.add_argument('--cw_k', default=-10000.0, type=float)
    parser.add_argument('--cw_n_iters', default=1000, type=int)
    parser.add_argument('--cw_lr', default=0.0001, type=float)
    parser.add_argument('--cw_binary_search_steps', default=9, type=int)

    args = parser.parse_args()
    args.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
    print('Arguments for attack:')
    print(args)

    # Set random seed.
    seed = args.seed
    random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    np.random.seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

    print('Preparing attack data!')
    transforms_test = transforms.Compose([transforms.ToTensor()])
    test_set = torchvision.datasets.CIFAR10(root='./datasets', train=False, download=True, transform=transforms_test)
    test_loader = torch.utils.data.DataLoader(test_set, batch_size=args.batch_size, shuffle=False, num_workers=4)

    # model_list contains all cifar10-Linf models.
    model_list = []
    for model, _ in linf.items():
        print('Loading {} from pre_models in RobustBench!'.format(model))
        net = load_model(model_name = model, model_dir='./per_comparison/models',dataset='cifar10', threat_model='Linf').to("cuda")
        net = torch.nn.DataParallel(net)
        args.model = model

        for i in range(1, 2):
            advesary = None
            if i == 0:
                print("####### AdvRobust FGSM attack #######")
                advesary = FGSM(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks FGSM attack #######")
                advesary_com = torchattacks.FGSM(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 1:
                print("####### AdvRobust PGD attack #######")
                advesary = PGD(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks PGD attack #######")
                advesary_com = torchattacks.PGD(net, eps=8.0 / 255)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 2:
                print("####### AdvRobust PGDL2 attack #######")
                advesary = PGDL2(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks PGDL2 attack #######")
                advesary_com = torchattacks.PGDL2(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 3:
                print("####### AdvRobust BIM attack #######")
                advesary = BIM(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks BIM attack #######")
                advesary_com = torchattacks.BIM(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 4:
                print("####### AdvRobust MIM attack #######")
                advesary = MIM(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks MIM attack #######")
                advesary_com = torchattacks.MIFGSM(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 5:
                print("####### AdvRobust CW attack #######")
                advesary = CW(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks CW attack #######")
                advesary_com = torchattacks.CW(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 6:
                print("####### AdvRobust CWL2 attack #######")
                advesary = CWL2(net, args)
                attack_test(advesary, test_loader, args, net, 'AdvRobust')
                print("####### TorchAttacks CW attack #######")
                advesary_com = torchattacks.CW(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            elif i == 7:
                # print("####### AdvRobust CWL2 attack #######")
                # advesary = CWL2(net, args)
                # attack_test(advesary, test_loader, args, net)
                print("####### TorchAttacks APGD attack #######")
                advesary_com = torchattacks.APGD(net)
                attack_test(advesary_com, test_loader, args, net, 'TorchAttacks')
            else:
                print("No attack is running, bye!")
                break