-
Notifications
You must be signed in to change notification settings - Fork 0
/
func.py
67 lines (56 loc) · 2.54 KB
/
func.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
def cw_loss(logits, y, confidence=0):
onehot_y = torch.nn.functional.one_hot(y, num_classes=logits.shape[1]).float()
self_loss = F.nll_loss(-logits, y, reduction='none')
other_loss = torch.max((1 - onehot_y) * logits, dim=1)[0]
return -torch.mean(torch.clamp(self_loss - other_loss + confidence, 0))
class PGDAttack:
def __init__(self, epsilon=0.031, num_steps=100, step_size=0.0078, image_constraints=(0, 1)):
self.boxmin = image_constraints[0]
self.boxmax = image_constraints[1]
self.epsilon = epsilon
self.num_steps = num_steps
self.step_size = step_size
def grad_proj(self, data, order):
if order == 'inf':
return data.sign()
elif order == '2':
norm = torch.norm(data.view(len(data), -1), 2, 1, keepdim=True)
return data / (norm + 1e-8).unsqueeze(2).unsqueeze(3).expand_as(data)
def eta_proj(self, eta, order, epsilon):
if order == 'inf':
return torch.clamp(eta, -epsilon, epsilon)
elif order == '2':
norm_eta = torch.norm(eta.view(len(eta), -1), p=2, dim=1, keepdim=True)
norm_eta = torch.clamp(norm_eta, epsilon, np.inf).unsqueeze(2).unsqueeze(3).expand_as(eta)
return eta * epsilon / norm_eta
def attack(self, model, X, y, loss_type='ce'):
out = model(X)
err = (out.data.max(1)[1] != y.data).float().sum()
X_pgd = X.clone()
for i in range(self.num_steps):
X_pgd = Variable(X_pgd, requires_grad=True)
if loss_type == 'ce':
loss = nn.CrossEntropyLoss()(model(X_pgd), y)
else:
loss = cw_loss(model(X_pgd), y)
loss.backward()
X_pgd = X_pgd + self.step_size * self.grad_proj(X_pgd.grad.data, 'inf')
eta = self.eta_proj(X_pgd - X, 'inf', self.epsilon)
X_pgd = X + eta
X_pgd = torch.clamp(X_pgd, self.boxmin, self.boxmax)
err_pgd = (model(X_pgd).data.max(1)[1] != y.data).float().sum()
return err.item(), err_pgd.item()
class DataAugmentModel(nn.Module):
def __init__(self, model, im_mean=None, im_std=None):
super(DataAugmentModel, self).__init__()
self.model = model
self.im_mean = im_mean
self.im_std = im_std
def forward(self, image):
image = (image - self.im_mean.expand_as(image)) / self.im_std.expand_as(image)
return self.model(image)