File size: 9,415 Bytes
4113c4d |
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 |
import torch
from torch import optim
import torch.nn as nn
import numpy as np
import logging
from deeprobust.image.attack.base_attack import BaseAttack
from deeprobust.image.utils import onehot_like
from deeprobust.image.optimizer import AdamOptimizer
class CarliniWagner(BaseAttack):
"""
C&W attack is an effective method to calcuate high-confidence adversarial examples.
References
----------
.. [1] Carlini, N., & Wagner, D. (2017, May). Towards evaluating the robustness of neural networks. https://arxiv.org/pdf/1608.04644.pdf
This reimplementation is based on https://github.com/kkew3/pytorch-cw2
Copyright 2018 Kaiwen Wu
Examples
--------
>>> from deeprobust.image.attack.cw import CarliniWagner
>>> from deeprobust.image.netmodels.CNN import Net
>>> from deeprobust.image.config import attack_params
>>> model = Net()
>>> model.load_state_dict(torch.load("./trained_models/MNIST_CNN_epoch_20.pt", map_location = torch.device('cuda')))
>>> model.eval()
>>> x,y = datasets.MNIST()
>>> attack = CarliniWagner(model, device='cuda')
>>> AdvExArray = attack.generate(x, y, target_label = 1, classnum = 10, **attack_params['CW_MNIST])
"""
def __init__(self, model, device = 'cuda'):
super(CarliniWagner, self).__init__(model, device)
self.model = model
self.device = device
def generate(self, image, label, target_label, **kwargs):
"""
Call this function to generate adversarial examples.
Parameters
----------
image :
original image
label :
target label
kwargs :
user defined paremeters
"""
assert self.check_type_device(image, label)
assert self.parse_params(**kwargs)
self.target = target_label
return self.cw(self.model,
self.image,
self.label,
self.target,
self.confidence,
self.clip_max,
self.clip_min,
self.max_iterations,
self.initial_const,
self.binary_search_steps,
self.learning_rate
)
def parse_params(self,
classnum = 10,
confidence = 1e-4,
clip_max = 1,
clip_min = 0,
max_iterations = 1000,
initial_const = 1e-2,
binary_search_steps = 5,
learning_rate = 0.00001,
abort_early = True):
"""
Parse the user defined parameters.
Parameters
----------
classnum :
number of class
confidence :
confidence
clip_max :
maximum pixel value
clip_min :
minimum pixel value
max_iterations :
maximum number of iterations
initial_const :
initialization of binary search
binary_search_steps :
step number of binary search
learning_rate :
learning rate
abort_early :
Set abort_early = True to allow early stop
"""
self.classnum = classnum
self.confidence = confidence
self.clip_max = clip_max
self.clip_min = clip_min
self.max_iterations = max_iterations
self.initial_const = initial_const
self.binary_search_steps = binary_search_steps
self.learning_rate = learning_rate
self.abort_early = abort_early
return True
def cw(self, model, image, label, target, confidence, clip_max, clip_min, max_iterations, initial_const, binary_search_steps, learning_rate):
#change the input image
img_tanh = self.to_attack_space(image.cpu())
img_ori ,_ = self.to_model_space(img_tanh)
img_ori = img_ori.to(self.device)
#binary search initialization
c = initial_const
c_low = 0
c_high = np.inf
found_adv = False
last_loss = np.inf
for step in range(binary_search_steps):
#initialize w : perturbed image in tanh space
w = torch.from_numpy(img_tanh.numpy())
optimizer = AdamOptimizer(img_tanh.shape)
is_adversarial = False
for iteration in range(max_iterations):
# adversary example
img_adv, adv_grid = self.to_model_space(w)
img_adv = img_adv.to(self.device)
img_adv.requires_grad = True
#output of the layer before softmax
output = model.get_logits(img_adv)
#pending success
is_adversarial = self.pending_f(img_adv)
#calculate loss function and gradient of loss funcition on x
loss, loss_grad = self.loss_function(
img_adv, c, self.target, img_ori, self.confidence, self.clip_min, self.clip_max
)
#calculate gradient of loss function on w
gradient = adv_grid.to(self.device) * loss_grad.to(self.device)
w = w + torch.from_numpy(optimizer(gradient.cpu().detach().numpy(), learning_rate)).float()
if is_adversarial:
found_adv = True
#do binary search on c
if found_adv:
c_high = c
else:
c_low = c
if c_high == np.inf:
c *= 10
else:
c = (c_high + c_low) / 2
if (step % 10 == 0):
print("iteration:{:.0f},loss:{:.4f}".format(step,loss))
# if (step == 50):
# learning_rate = learning_rate/100
#abort early
if(self.abort_early == True and (step % 10) == 0 and step > 100) :
print("early abortion?", loss, last_loss)
if not (loss <= 0.9999 * last_loss):
break
last_loss = loss
return img_adv.detach()
def loss_function(
self, x_p, const, target, reconstructed_original, confidence, min_, max_):
"""Returns the loss and the gradient of the loss w.r.t. x,
assuming that logits = model(x)."""
## get the output of model before softmax
x_p.requires_grad = True
logits = self.model.get_logits(x_p).to(self.device)
## find the largest class except the target class
targetlabel_mask = (torch.from_numpy(onehot_like(np.zeros(self.classnum), target))).double()
secondlargest_mask = (torch.from_numpy(np.ones(self.classnum)) - targetlabel_mask).to(self.device)
secondlargest = np.argmax((logits.double() * secondlargest_mask).cpu().detach().numpy(), axis = 1)
is_adv_loss = logits[0][secondlargest] - logits[0][target]
# is_adv is True as soon as the is_adv_loss goes below 0
# but sometimes we want additional confidence
is_adv_loss += confidence
if is_adv_loss == 0:
is_adv_loss_grad = 0
else:
is_adv_loss.backward()
is_adv_loss_grad = x_p.grad
is_adv_loss = max(0, is_adv_loss)
s = max_ - min_
squared_l2_distance = np.sum( ((x_p - reconstructed_original) ** 2).cpu().detach().numpy() ) / s ** 2
total_loss = squared_l2_distance + const * is_adv_loss
squared_l2_distance_grad = (2 / s ** 2) * (x_p - reconstructed_original)
#print(is_adv_loss_grad)
total_loss_grad = squared_l2_distance_grad + const * is_adv_loss_grad
return total_loss, total_loss_grad
def pending_f(self, x_p):
"""Pending is the loss function is less than 0
"""
targetlabel_mask = torch.from_numpy(onehot_like(np.zeros(self.classnum), self.target))
secondlargest_mask = torch.from_numpy(np.ones(self.classnum)) - targetlabel_mask
targetlabel_mask = targetlabel_mask.to(self.device)
secondlargest_mask = secondlargest_mask.to(self.device)
Zx_i = np.max((self.model.get_logits(x_p).double().to(self.device) * secondlargest_mask).cpu().detach().numpy())
Zx_t = np.max((self.model.get_logits(x_p).double().to(self.device) * targetlabel_mask).cpu().detach().numpy())
if ( Zx_i - Zx_t < - self.confidence):
return True
else:
return False
def to_attack_space(self, x):
x = x.detach()
# map from [min_, max_] to [-1, +1]
# x'=(x- 0.5 * (max+min) / 0.5 * (max-min))
a = (self.clip_min + self.clip_max) / 2
b = (self.clip_max - self.clip_min) / 2
x = (x - a) / b
# from [-1, +1] to approx. (-1, +1)
x = x * 0.999999
# from (-1, +1) to (-inf, +inf)
return np.arctanh(x)
def to_model_space(self, x):
"""Transforms an input from the attack space
to the model space. This transformation and
the returned gradient are elementwise."""
# from (-inf, +inf) to (-1, +1)
x = np.tanh(x)
grad = 1 - np.square(x)
# map from (-1, +1) to (min_, max_)
a = (self.clip_min + self.clip_max) / 2
b = (self.clip_max - self.clip_min) / 2
x = x * b + a
grad = grad * b
return x, grad
|