forked from RjDuan/AdvDrop
-
Notifications
You must be signed in to change notification settings - Fork 0
/
infod_sample.py
215 lines (183 loc) · 8.56 KB
/
infod_sample.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
import numpy as np
import json
import os
import sys
import time
import math
import io
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import models
import torchvision.datasets as dsets
import torchvision.transforms as transforms
from torchattacks.attack import Attack
from utils import *
from compression import *
from decompression import *
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
class InfoDrop(Attack):
r"""
Distance Measure : l_inf bound on quantization table
Arguments:
model (nn.Module): model to attack.
steps (int): number of steps. (DEFALUT: 40)
batch_size (int): batch size
q_size: bound for quantization table
targeted: True for targeted attack
Shape:
- images: :math:`(N, C, H, W)` where `N = number of batches`, `C = number of channels`, `H = height` and `W = width`. It must have a range [0, 1].
- labels: :math:`(N)` where each value :math:`y_i` is :math:`0 \leq y_i \leq` `number of labels`.
- output: :math:`(N, C, H, W)`.
"""
def __init__(self, model, height = 224, width = 224, steps=40, batch_size = 20, block_size = 8, q_size = 10, targeted = False):
super(InfoDrop, self).__init__("InfoDrop", model)
self.steps = steps
self.targeted = targeted
self.batch_size = batch_size
self.height = height
self.width = width
# Value for quantization range
self.factor_range = [5, q_size]
# Differential quantization
self.alpha_range = [0.1, 1e-20]
self.alpha = torch.tensor(self.alpha_range[0])
self.alpha_interval = torch.tensor((self.alpha_range[1] - self.alpha_range[0])/ self.steps)
block_n = np.ceil(height / block_size) * np.ceil(height / block_size)
q_ini_table = np.empty((batch_size,int(block_n),block_size,block_size), dtype = np.float32)
q_ini_table.fill(q_size)
self.q_tables = {"y": torch.from_numpy(q_ini_table),
"cb": torch.from_numpy(q_ini_table),
"cr": torch.from_numpy(q_ini_table)}
def forward(self, images, labels):
r"""
Overridden.
"""
q_table = None
self.alpha = self.alpha.to(self.device)
self.alpha_interval = self.alpha_interval.to(self.device)
images = images.clone().detach().to(self.device)
labels = labels.clone().detach().to(self.device)
adv_loss = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam([self.q_tables["y"], self.q_tables["cb"], self.q_tables["cr"]], lr= 0.01)
images = images.permute(0, 2, 3, 1)
components = {'y': images[:,:,:,0], 'cb': images[:,:,:,1], 'cr': images[:,:,:,2]}
for i in range(self.steps):
self.q_tables["y"].requires_grad = True
self.q_tables["cb"].requires_grad = True
self.q_tables["cr"].requires_grad = True
upresults = {}
for k in components.keys():
comp = block_splitting(components[k])
comp = dct_8x8(comp)
comp = quantize(comp, self.q_tables[k], self.alpha)
comp = dequantize(comp, self.q_tables[k])
comp = idct_8x8(comp)
merge_comp = block_merging(comp, self.height, self.width)
upresults[k] = merge_comp
rgb_images = torch.cat([upresults['y'].unsqueeze(3), upresults['cb'].unsqueeze(3), upresults['cr'].unsqueeze(3)], dim=3)
rgb_images = rgb_images.permute(0, 3, 1, 2)
outputs = self.model(rgb_images)
_, pre = torch.max(outputs.data, 1)
if self.targeted:
suc_rate = ((pre == labels).sum()/self.batch_size).cpu().detach().numpy()
else:
suc_rate = ((pre != labels).sum()/self.batch_size).cpu().detach().numpy()
adv_cost = adv_loss(outputs, labels)
if not self.targeted:
adv_cost = -1* adv_cost
total_cost = adv_cost
optimizer.zero_grad()
total_cost.backward()
self.alpha += self.alpha_interval
for k in self.q_tables.keys():
self.q_tables[k] = self.q_tables[k].detach() - torch.sign(self.q_tables[k].grad)
self.q_tables[k] = torch.clamp(self.q_tables[k], self.factor_range[0], self.factor_range[1]).detach()
if i%10 == 0:
print('Step: ', i, " Loss: ", total_cost.item(), " Current Suc rate: ", suc_rate )
if suc_rate >= 1:
print('End at step {} with suc. rate {}'.format(i, suc_rate))
q_images = torch.clamp(rgb_images, min=0, max=255.0).detach()
return q_images, pre, i
q_images = torch.clamp(rgb_images, min=0, max=255.0).detach()
return q_images, pre, q_table
class Normalize(nn.Module) :
def __init__(self, mean, std) :
super(Normalize, self).__init__()
self.register_buffer('mean', torch.Tensor(mean))
self.register_buffer('std', torch.Tensor(std))
def forward(self, input):
# Broadcasting
input = input/255.0
mean = self.mean.reshape(1, 3, 1, 1)
std = self.std.reshape(1, 3, 1, 1)
return (input - mean) / std
def save_img(img, img_name, save_dir):
create_dir(save_dir)
img_path = os.path.join(save_dir, img_name)
img_pil = Image.fromarray(img.astype(np.uint8))
img_pil.save(img_path)
def pred_label_and_confidence(model, input_batch, labels_to_class):
input_batch = input_batch.cuda()
with torch.no_grad():
out = model(input_batch)
_, index = torch.max(out, 1)
percentage = torch.nn.functional.softmax(out, dim=1) * 100
# print(percentage.shape)
pred_list = []
for i in range(index.shape[0]):
pred_class = labels_to_class[index[i]]
pred_conf = str(round(percentage[i][index[i]].item(),2))
pred_list.append([pred_class, pred_conf])
return pred_list
if __name__ == "__main__":
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class_idx = json.load(open("./imagenet_class_index.json"))
idx2label = [class_idx[str(k)][1] for k in range(len(class_idx))]
class2label = [class_idx[str(k)][0] for k in range(len(class_idx))]
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),])
norm_layer = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
resnet_model = nn.Sequential(
norm_layer,
models.resnet50(pretrained=True)
).to(device)
resnet_model = resnet_model.eval()
# Uncomment if you want save results
# save_dir = "./results"
# create_dir(save_dir)
batch_size = 20
tar_cnt = 1000
q_size = 40
cur_cnt = 0
suc_cnt = 0
data_dir = "./test-data"
data_clean(data_dir)
normal_data = image_folder_custom_label(root=data_dir, transform=transform, idx2label=class2label)
normal_loader = torch.utils.data.DataLoader(normal_data, batch_size=batch_size, shuffle=False)
normal_iter = iter(normal_loader)
for i in range(tar_cnt//batch_size):
print("Iter: ", i)
images, labels = normal_iter.next()
# For target attack: set random target.
# Comment if you set untargeted attack.
labels = torch.from_numpy(np.random.randint(0, 1000, size = batch_size))
images = images * 255.0
attack = InfoDrop(resnet_model, batch_size=batch_size, q_size =q_size, steps=150, targeted = True)
at_images, at_labels, suc_step = attack(images, labels)
# Uncomment following codes if you wang to save the adv imgs
# at_images_np = at_images.detach().cpu().numpy()
# adv_img = at_images_np[0]
# adv_img = np.moveaxis(adv_img, 0, 2)
# adv_dir = os.path.join(save_dir, str(q_size))
# img_name = "adv_{}.jpg".format(i)
# save_img(adv_img, img_name, adv_dir)
labels = labels.to(device)
suc_cnt += (at_labels == labels).sum().item()
print("Current suc. rate: ", suc_cnt/((i+1)*batch_size))
score_list = np.zeros(tar_cnt)
score_list[:suc_cnt] = 1.0
stderr_dist = np.std(np.array(score_list))/np.sqrt(len(score_list))
print('Avg suc rate: %.5f +/- %.5f'%(suc_cnt/tar_cnt,stderr_dist))