-
Notifications
You must be signed in to change notification settings - Fork 10
/
main.py
195 lines (159 loc) · 9.95 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
import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler
from attentionControl import AttentionControlEdit
import diff_latent_attack
from PIL import Image
import numpy as np
import os
import glob
from other_attacks import model_transfer
import random
import sys
from natsort import ns, natsorted
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--save_dir', default="output", type=str,
help='Where to save the adversarial examples, and other results')
parser.add_argument('--images_root', default="demo/images", type=str,
help='The clean images root directory')
parser.add_argument('--label_path', default="demo/labels.txt", type=str,
help='The clean images labels.txt')
parser.add_argument('--is_test', default=False, type=bool,
help='Whether to test the robustness of the generated adversarial examples')
parser.add_argument('--pretrained_diffusion_path',
default="stabilityai/stable-diffusion-2-base",
type=str,
help='Change the path to `stabilityai/stable-diffusion-2-base` if want to use the pretrained model')
parser.add_argument('--diffusion_steps', default=20, type=int, help='Total DDIM sampling steps')
parser.add_argument('--start_step', default=15, type=int, help='Which DDIM step to start the attack')
parser.add_argument('--iterations', default=30, type=int, help='Iterations of optimizing the adv_image')
parser.add_argument('--res', default=224, type=int, help='Input image resized resolution')
parser.add_argument('--model_name', default="inception", type=str,
help='The surrogate model from which the adversarial examples are crafted')
parser.add_argument('--dataset_name', default="imagenet_compatible", type=str,
choices=["imagenet_compatible", "cub_200_2011", "standford_car"],
help='The dataset name for generating adversarial examples')
parser.add_argument('--is_apply_mask', default=False, type=bool,
help='Whether to leverage pseudo mask for better imperceptibility (See Appendix D)')
parser.add_argument('--is_hard_mask', default=False, type=bool,
help='Which type of mask to leverage (See Appendix D)')
parser.add_argument('--guidance', default=2.5, type=float, help='guidance scale of diffusion models')
parser.add_argument('--attack_loss_weight', default=10, type=int, help='attack loss weight factor')
parser.add_argument('--cross_attn_loss_weight', default=10000, type=int, help='cross attention loss weight factor')
parser.add_argument('--self_attn_loss_weight', default=100, type=int, help='self attention loss weight factor')
def seed_torch(seed=42):
"""For reproducibility"""
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
seed_torch(42)
def run_diffusion_attack(image, label, diffusion_model, diffusion_steps, guidance=2.5,
self_replace_steps=1., save_dir=r"C:\Users\PC\Desktop\output", res=224,
model_name="inception", start_step=15, iterations=30, args=None):
controller = AttentionControlEdit(diffusion_steps, self_replace_steps, args.res)
adv_image, clean_acc, adv_acc = diff_latent_attack.diffattack(diffusion_model, label, controller,
num_inference_steps=diffusion_steps,
guidance_scale=guidance,
image=image,
save_path=save_dir, res=res, model_name=model_name,
start_step=start_step,
iterations=iterations, args=args)
return adv_image, clean_acc, adv_acc
if __name__ == "__main__":
args = parser.parse_args()
assert args.res % 32 == 0 and args.res >= 96, "Please ensure the input resolution be a multiple of 32 and also >= 96."
guidance = args.guidance
diffusion_steps = args.diffusion_steps # Total DDIM sampling steps.
start_step = args.start_step # Which DDIM step to start the attack.
iterations = args.iterations # Iterations of optimizing the adv_image.
res = args.res # Input image resized resolution.
model_name = args.model_name # The surrogate model from which the adversarial examples are crafted.
if args.dataset_name == "imagenet_compatible":
assert model_name not in ["cubResnet50", "cubSEResnet154", "cubSEResnet101", "carResnet50", "carSEResnet154",
"carSEResnet101"], f"There is no pretrained weight of {model_name} for ImageNet-Compatible dataset."
if args.dataset_name == "cub_200_2011":
assert model_name in ["cubResnet50", "cubSEResnet154",
"cubSEResnet101"], f"There is no pretrained weight of {model_name} for CUB_200_2011 dataset."
if args.dataset_name == "standford_car":
assert model_name in ["carResnet50", "carSEResnet154",
"carSEResnet101"], f"There is no pretrained weight of {model_name} for Standford Cars dataset."
save_dir = args.save_dir # Where to save the adversarial examples, and other results.
os.makedirs(save_dir, exist_ok=True)
"If you set 'is_test' to True, please turn 'images_root' to the path of the output results' path."
images_root = args.images_root # The clean images' root directory.
label_path = args.label_path # The clean images' labels.txt.
with open(label_path, "r") as f:
label = []
for i in f.readlines():
label.append(int(i.rstrip()) - 1) # The label number of the imagenet-compatible dataset starts from 1.
label = np.array(label)
is_test = args.is_test # Whether to test the robustness of the generated adversarial examples.
print(f"\n******Attack based on Diffusion, Attacked Dataset: {args.dataset_name}*********")
# Change the path to "stabilityai/stable-diffusion-2-base" if you want to use the pretrained model.
pretrained_diffusion_path = args.pretrained_diffusion_path
ldm_stable = StableDiffusionPipeline.from_pretrained(pretrained_diffusion_path).to('cuda:0')
ldm_stable.scheduler = DDIMScheduler.from_config(ldm_stable.scheduler.config)
"Attack a subset images"
all_images = glob.glob(os.path.join(images_root, "*"))
all_images = natsorted(all_images, alg=ns.PATH)
adv_images = []
images = []
clean_all_acc = 0
adv_all_acc = 0
if is_test:
all_clean_images = glob.glob(os.path.join(images_root, "*originImage*"))
all_clean_images = natsorted(all_clean_images, alg=ns.PATH)
all_adv_images = glob.glob(os.path.join(images_root, "*adv_image*"))
all_adv_images = natsorted(all_adv_images, alg=ns.PATH)
for image_path, adv_image_path in zip(all_clean_images, all_adv_images):
tmp_image = Image.open(image_path).convert('RGB')
tmp_image = tmp_image.resize((res, res), resample=Image.LANCZOS)
tmp_image = np.array(tmp_image).astype(np.float32) / 255.0
tmp_image = tmp_image[None].transpose(0, 3, 1, 2)
images.append(tmp_image)
tmp_image = Image.open(adv_image_path).convert('RGB')
tmp_image = tmp_image.resize((res, res), resample=Image.LANCZOS)
tmp_image = np.array(tmp_image).astype(np.float32) / 255.0
tmp_image = tmp_image[None].transpose(0, 3, 1, 2)
adv_images.append(tmp_image)
images = np.concatenate(images)
adv_images = np.concatenate(adv_images)
"""
Test the robustness of the generated adversarial examples across a variety of normally trained models or
adversarially trained models.
"""
model_transfer(images, adv_images, label, res, save_path=save_dir, fid_path=images_root, args=args)
sys.exit()
for ind, image_path in enumerate(all_images):
tmp_image = Image.open(image_path).convert('RGB')
tmp_image.save(os.path.join(save_dir, str(ind).rjust(4, '0') + "_originImage.png"))
adv_image, clean_acc, adv_acc = run_diffusion_attack(tmp_image, label[ind:ind + 1],
ldm_stable,
diffusion_steps, guidance=guidance,
res=res, model_name=model_name,
start_step=start_step,
iterations=iterations,
save_dir=os.path.join(save_dir,
str(ind).rjust(4, '0')), args=args)
adv_image = adv_image.astype(np.float32) / 255.0
adv_images.append(adv_image[None].transpose(0, 3, 1, 2))
tmp_image = tmp_image.resize((res, res), resample=Image.LANCZOS)
tmp_image = np.array(tmp_image).astype(np.float32) / 255.0
tmp_image = tmp_image[None].transpose(0, 3, 1, 2)
images.append(tmp_image)
clean_all_acc += clean_acc
adv_all_acc += adv_acc
print("Clean acc: {}%".format(clean_all_acc / len(all_images) * 100))
print("Adv acc: {}%".format(adv_all_acc / len(all_images) * 100))
images = np.concatenate(images)
adv_images = np.concatenate(adv_images)
"""
Test the robustness of the generated adversarial examples across a variety of normally trained models or
adversarially trained models.
"""
model_transfer(images, adv_images, label, res, save_path=save_dir, args=args)