-
Notifications
You must be signed in to change notification settings - Fork 1
/
FingerSafe.py
383 lines (328 loc) · 16.5 KB
/
FingerSafe.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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
import torch
import torch.nn as nn
import lpips
from L_orientation import ridge_orient
import numpy as np
import cv2
from cal_contrast import cal_contrast
import torch
import pickle
import torch.nn.functional as F
from models.inception_resnet_v1 import ResNet
import torchvision
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
import cv2
import torchvision.transforms as transforms
import torch.nn.functional as functional
import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import confusion_matrix
import itertools
from torchvision import transforms
import os
from criterion import Criterion
from L_orientation import Gray
from utils import draw_convergence
os.environ["CUDA_VISIBLE_DEVICES"]="1"
# steal some code from torchattacks
class FingerSafe(nn.Module):
def __init__(self, model, device, eps=0.3, alpha=2 / 255, steps=40, random_start=False, inc=None, dense=None, lamda=100, gamma=500, gauss_1=7, gauss_sigma=3, draw_convergence=False):
super(FingerSafe, self).__init__()
self.eps = eps
self.alpha = alpha
self.steps = steps
self.random_start = random_start
self.ridge = ridge_orient()
self.contrast = cal_contrast()
self.pad = torch.nn.ZeroPad2d((31, 31, 31, 31))
self.model = model
self.criterion = Criterion()
self.gray = Gray()
self.device = device
self.image_per_class = 6
# self.target_feature = target_feature # target feature
# self.target_image = target_image
self.inception = inc
self.densenet = dense
self.lamda = lamda
self.gamma = gamma
self.gauss_1 = gauss_1
self.gauss_sigma = gauss_sigma
self.draw_convergence = draw_convergence
def forward_untarget(self, images):
print(images.shape)
images = images.clone().squeeze().detach().to(self.device)
adv_images = images.clone().detach()
for i in range(self.steps):
images = images.clone().detach().to(self.device)
adv_images = adv_images.clone().detach()
images.requires_grad = True
adv_images.requires_grad = True
adv_images_rep = self.model(adv_images)
images_rep = self.model(images)
pred = images_rep.data.max(1)[1]
adv_images_rep_dim0 = adv_images_rep.repeat(self.image_per_class, 1, 1)
adv_images_rep_dim1 = adv_images_rep.repeat(self.image_per_class, 1, 1).transpose(0, 1)
images_rep_dim1 = images_rep.repeat(self.image_per_class, 1, 1).transpose(0, 1)
mask = (torch.ones(self.image_per_class, self.image_per_class) - torch.eye(self.image_per_class)).to(self.device)
# loss_rep = torch.sum((adv_images_rep_dim0 - images_rep_dim1).abs(), dim=-1) # L1
# loss_rep_diff = torch.sum((adv_images_rep_dim0 - adv_images_rep_dim1).abs(), dim=-1).to(
# self.device) # L1
loss_rep = torch.norm((adv_images_rep_dim0 - images_rep_dim1), dim=-1, p=2) # L2
loss_rep_diff = torch.norm((adv_images_rep_dim0 - adv_images_rep_dim1), dim=-1, p=2).to(
self.device) # L2
loss_rep_diff = torch.einsum('ij, ij->ij', loss_rep_diff, mask)
# loss_rep_diff = F.pairwise_distance(adv_images_rep_dim0, adv_images_rep_dim1, p=2)
# print(loss_rep.shape)
# print(loss_rep_diff.shape) # expect 6*6
loss_orientation = L_orientation_lsm(self.ridge, images, adv_images)
loss_orientation_diff = L_orientation_lsm(self.ridge, adv_images, adv_images).to(self.device)
loss_orientation_diff = torch.einsum('ij, ij->ij', loss_orientation_diff, mask)
loss_contrast = L_contrast(self.contrast, images, adv_images)
L_V = torch.mean(loss_rep) + torch.sum(loss_rep_diff) / (self.image_per_class * (self.image_per_class-1))
L_O = torch.mean(loss_orientation) + torch.sum(loss_orientation_diff) / (self.image_per_class * (self.image_per_class-1))
L_C = loss_contrast
print(L_V, L_O, L_C)
# cost = -10 * L_V - 1e2 * L_O + 5e-2 * L_C # 0912B
# cost = -1 * L_V - 10 * L_O + 5e-3 * L_C # 0920a
# cost = -1 * L_V - 10 * L_O + 5e-2 * L_C # 0920b
# cost = -1 * L_V - 10 * L_O + 5e-1 * L_C # 0920c
# cost = -1 * L_V - 1 * L_O + 5e-1 * L_C # 0920d
# cost = -1 * L_V - 10 * L_O + 1 * L_C # 0920e
# cost = -1 * L_V - 1 * L_O + 5e-2 * L_C # 0923f
# cost = -1 * L_V # 0923g
# cost = -1 * L_V - 1 * L_O + 1e-2 * L_C # 0923h
# cost = -1 * L_V - 1 * L_O + 5e-3 * L_C # 0923i
# cost = -1 * L_V - 0.1 * L_O + 1e-3 * L_C # 0923j
# cost = -1 * L_V - 10 * L_O + 0.1 * L_C # 0923k
# cost = -1 * L_V - 1 * L_O + 5e-5 * L_C # color
# cost = -1 * L_V - 10 * L_O + 0 * L_C # noLc
# cost = -1 * L_V - 10 * L_O + 0.05 * L_C # 1001m
cost = -1 * L_V - 10 * L_O + 0.1 * L_C # 1001n now
# cost = -1 * L_V - 0 * L_O + 0.1 * L_C # noLF
grad = torch.autograd.grad(cost, adv_images,
retain_graph=False, create_graph=False)
grad = grad[0]
adv_images = adv_images.detach() - self.alpha * grad.sign()
delta = torch.clamp(adv_images - images, min=-self.eps, max=self.eps)
adv_images = torch.clamp(images + delta, min=0, max=1).clone().detach()
return adv_images
def forward(self, images, masks=None):
images = images.clone().detach().to(self.device)
original_embs = self.model(images)
adv_images = images.clone().detach()
costs = []
if self.random_start:
# Starting at a uniformly random point
adv_images = adv_images + torch.empty_like(adv_images).uniform_(-self.eps, self.eps)
adv_images = torch.clamp(adv_images, min=0, max=1).detach()
for i in range(self.steps):
adv_images.requires_grad = True
outputs = self.model(adv_images)
loss_rep = L_representation(outputs, original_embs) # far from itself
loss_orientation = L_orientation_target(self.ridge, images, adv_images, masks=masks)
loss_contrast = L_contrast(self.contrast, images, adv_images)
cost = -1 * loss_rep - self.lamda * loss_orientation + self.gamma * loss_contrast
costs.append(cost.detach().cpu())
grad = torch.autograd.grad(cost, adv_images, retain_graph=False, create_graph=False)[0]
# todo here to mask grad, only for physical world
# masked_grad = grad.masked_fill(masks.to(self.device), value=0)
adv_images = adv_images.detach() - self.alpha * grad.sign()
delta = torch.clamp(adv_images - images, min=-self.eps, max=self.eps)
adv_images = torch.clamp(images + delta, min=0, max=1).detach()
if self.draw_convergence:
costs = np.array(costs)
np.save('./convergence/fingersafe.npy', costs)
exit(0)
return adv_images
def forward_lowkey(self, images, masks=None, lowkey_lpips = 5.0):
"""
LowKey
"""
images = images.clone().detach().to(self.device)
original_resnet = self.model(images)
original_inception = self.inception(images)
original_densenet = self.densenet(images)
adv_images = images.clone().detach().to(self.device)
gauss = torchvision.transforms.GaussianBlur(self.gauss_1, sigma=self.gauss_sigma)
loss_fn_resnet = lpips.LPIPS(net='alex').to(self.device)
costs = []
if self.random_start:
# Starting at a uniformly random point
adv_images = adv_images + torch.empty_like(adv_images).uniform_(-self.eps, self.eps)
adv_images = torch.clamp(adv_images, min=0, max=1).detach()
for i in range(self.steps):
adv_images.requires_grad = True
# resnet
outputs_resnet = self.model(adv_images)
guass_outputs_resnet = self.model(gauss(adv_images))
# inception
outputs_inception = self.inception(adv_images)
guass_outputs_inception = self.inception(gauss(adv_images))
# densenet
outputs_densenet = self.densenet(adv_images)
guass_outputs_densenet = self.densenet(gauss(adv_images))
# todo: FingerSafe: to make adv representation close to target representation
loss_resnet = L_representation(outputs_resnet, original_resnet) + L_representation(guass_outputs_resnet, original_resnet)
loss_resnet = loss_resnet / torch.sqrt(L_representation(original_resnet, 0))
loss_inception = L_representation(outputs_inception, original_inception) + L_representation(guass_outputs_inception,
original_inception)
loss_inception = loss_inception / torch.sqrt(L_representation(original_inception, 0))
loss_densenet = L_representation(outputs_densenet, original_densenet) + L_representation(guass_outputs_densenet,
original_densenet)
loss_densenet = loss_densenet / torch.sqrt(L_representation(original_densenet, 0))
loss_rep = (loss_resnet + loss_inception + loss_densenet) / 6
loss_lpips = loss_fn_resnet(images, adv_images)
cost = -1 * loss_rep + lowkey_lpips * torch.mean(loss_lpips) # orient: A=1, B=10, C=100, D=1000
costs.append(cost.detach().cpu())
# print(loss_rep, torch.mean(loss_lpips))
grad = torch.autograd.grad(cost, adv_images,
retain_graph=False, create_graph=False)[0]
adv_images = adv_images.detach() - self.alpha * grad.sign()
# delta = torch.clamp(adv_images - images, min=-self.eps, max=self.eps)
delta = adv_images - images
adv_images = torch.clamp(images + delta, min=0, max=1).detach()
if self.draw_convergence:
costs = np.array(costs)
np.save('./convergence/lowkey.npy', costs)
exit(0)
return adv_images
class FeatureExtractor(nn.Module):
def __init__(self, submodule, extracted_layers):
super(FeatureExtractor, self).__init__()
self.submodule = submodule
self.extracted_layers = extracted_layers
def forward(self, x):
outputs = []
for name, module in self.submodule._modules.items():
if name is "fc": x = x.view(x.size(0), -1)
x = module(x)
if name in self.extracted_layers:
outputs.append(x)
return outputs
class ContentLoss(nn.Module):
def __init__(self, target, ):
super(ContentLoss, self).__init__()
# we 'detach' the target content from the tree used
# to dynamically compute the gradient: this is a stated value,
# not a variable. Otherwise the forward method of the criterion
# will throw an error.
self.target = target.detach()
def forward(self, input):
self.loss = F.mse_loss(input, self.target)
return input
def L_color(images, adv_images):
# pairwise distance, 1 if similiar, 0 if dissimiliar
distance = 0
for i in range(len(images)):
delta = adv_images[i] - images[i]
# noise in BGR
R_bar = (adv_images[i, 2, ...] + images[i, 2, ...]) / 2
color_distance = torch.sqrt((2 + R_bar) * torch.pow(delta[2, ...] * 255, 2) + 4 * torch.pow(delta[1, ...] * 255, 2) + (2 + (1-R_bar) * torch.pow(delta[0, ...] * 255, 2)))
distance += torch.norm(color_distance, p='fro')
return distance / len(images)
def L_representation(features, target):
# d = F.pairwise_distance(features, target, p=2)
diff = features - target
scale_factor = 1 # torch.sqrt(torch.sum(torch.pow(target, 2), dim=1))
distance = torch.sum(torch.pow(diff, 2), dim=1)
distance = distance / scale_factor
return torch.mean(distance)
def L_orientation_lsm(ridge, images, adv_images):
# pairwise distance, 1 if similiar, 0 if dissimiliar
# distance = 0
loss_matrix = torch.zeros(6, 6)
for i in range(len(images)):
for j in range(len(images)):
images_orient = ridge(images[i])
adv_images_orient = ridge(adv_images[j])
diff = torch.abs(adv_images_orient - images_orient)
# d = F.pairwise_distance(images_orient, adv_images_orient, p=2)
d = torch.sin(diff)
d = d.abs()
loss_matrix[i, j] = torch.mean(d)
# distance += torch.mean(d) # + torch.std(d)
return loss_matrix
def L_orientation_target(ridge, t_images, adv_images, masks=None):
# let orientation of adv_images close to target images
distance = 0
for i in range(len(t_images)):
if masks is not None: # physical world
mask = masks[i].bool()
images_orient = ridge(t_images[i]).masked_fill(~mask.to('cuda'), value=0)
adv_images_orient = ridge(adv_images[i]).masked_fill(~mask.to('cuda'), value=0)
else:
images_orient = ridge(t_images[i])
adv_images_orient = ridge(adv_images[i])
# images_orient = ridge(t_images[i])
# adv_images_orient = ridge(adv_images[i])
diff = torch.abs(adv_images_orient - images_orient)
d = torch.sin(diff)
d = d.abs()
distance += torch.mean(d)
return distance / (len(t_images))
def L_contrast(contrast, images, adv_images):
# pairwise distance, 1 if similiar, 0 if dissimiliar
distance = 0
for i in range(len(images)):
# delta = (adv_images[i] - images[i]) * 255
# print(delta.shape)
# exit(0)
# noise in BGR
# contrast_img, C_WLF = contrast(delta)
# con1, _, sal1 = contrast(adv_images[i])
# con2, _, sal2 = contrast(images[i])
# contrast_img = con1 - con2
# distance += torch.norm(contrast_img, p='fro')
local1, _, sal1 = contrast(images[i])
local2, _, sal2 = contrast(adv_images[i])
sal1, sal2 = sal1.unsqueeze(0).repeat(3, 1, 1), sal2.unsqueeze(0).repeat(3, 1, 1)
con1 = torch.mul(local1, sal2)
con2 = torch.mul(local2, sal2)
# distance = distance + torch.sum(F.relu(con2 - con1)) # 1116: sum to mean
distance = distance + torch.mean(F.relu(con2 - con1))
distance = distance / len(images)
return distance
def L_color(images, adv_images):
# pairwise distance, 1 if similiar, 0 if dissimiliar
distance = 0
for i in range(len(images)):
delta = adv_images[i] - images[i]
# noise in BGR
R_bar = (adv_images[i, 2, ...] + images[i, 2, ...]) / 2
color_distance = torch.sqrt(
(2 + R_bar) * torch.pow(delta[2, ...] * 255, 2) + 4 * torch.pow(delta[1, ...] * 255, 2) + (
2 + (1 - R_bar) * torch.pow(delta[0, ...] * 255, 2)))
distance += torch.norm(color_distance, p='fro')
return distance / len(images)
def show_img(img, name):
img = img.cpu().detach().numpy()
img = ((img / np.max(img + 1e-6)) * 255).astype('uint8')
img = cv2.resize(img, (224, 224))
cv2.imshow(name, img)
def export(paths, images):
unloader = torchvision.transforms.ToPILImage()
rgb2bgr = [2, 1, 0]
images = images[:, rgb2bgr, :, :]
for idx in range(len(images)):
p, f = os.path.split(paths[0][idx])
if p.find('train') != -1:
new_path = p.replace('train', 'perturb2')
else:
new_path = p.replace('veri_test', 'sample2')
os.makedirs(os.path.dirname(os.path.join(new_path, f)), exist_ok=True)
img = unloader(images[idx].cpu().detach().squeeze(0)).convert('RGB')
print("exporting to:" + str(os.path.join(new_path, f)))
img.save(os.path.join(new_path, f))
class Gray(object):
def __call__(self, tensor): # tensor: 3 * w * h
# TODO: make efficient
_, w, h = tensor.shape
R = tensor[0]
G = tensor[1]
B = tensor[2]
tmp = 0.299 * R + 0.587 * G + 0.114 * B
tensor = tmp
tensor = tensor.view(1, w, h)
return tensor