-
Notifications
You must be signed in to change notification settings - Fork 191
/
ig_attack.py
213 lines (180 loc) · 8.43 KB
/
ig_attack.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
"""
Adversarial Examples on Graph Data: Deep Insights into Attack and Defense
https://arxiv.org/pdf/1903.01610.pdf
"""
import torch
import torch.multiprocessing as mp
from deeprobust.graph.targeted_attack import BaseAttack
from torch.nn.parameter import Parameter
from deeprobust.graph import utils
import torch.nn.functional as F
import numpy as np
import scipy.sparse as sp
from torch import optim
from torch.nn import functional as F
from torch.nn.modules.module import Module
import numpy as np
from tqdm import tqdm
import math
import scipy.sparse as sp
class IGAttack(BaseAttack):
"""IGAttack: IG-FGSM. Adversarial Examples on Graph Data: Deep Insights into Attack and Defense, https://arxiv.org/pdf/1903.01610.pdf.
Parameters
----------
model :
model to attack
nnodes : int
number of nodes in the input graph
feature_shape : tuple
shape of the input node features
attack_structure : bool
whether to attack graph structure
attack_features : bool
whether to attack node features
device: str
'cpu' or 'cuda'
Examples
--------
>>> from deeprobust.graph.data import Dataset
>>> from deeprobust.graph.defense import GCN
>>> from deeprobust.graph.targeted_attack import IGAttack
>>> data = Dataset(root='/tmp/', name='cora')
>>> adj, features, labels = data.adj, data.features, data.labels
>>> idx_train, idx_val, idx_test = data.idx_train, data.idx_val, data.idx_test
>>> # Setup Surrogate model
>>> surrogate = GCN(nfeat=features.shape[1], nclass=labels.max().item()+1,
nhid=16, dropout=0, with_relu=False, with_bias=False, device='cpu').to('cpu')
>>> surrogate.fit(features, adj, labels, idx_train, idx_val, patience=30)
>>> # Setup Attack Model
>>> target_node = 0
>>> model = IGAttack(surrogate, nnodes=adj.shape[0], attack_structure=True, attack_features=True, device='cpu').to('cpu')
>>> # Attack
>>> model.attack(features, adj, labels, idx_train, target_node, n_perturbations=5, steps=10)
>>> modified_adj = model.modified_adj
>>> modified_features = model.modified_features
"""
def __init__(self, model, nnodes=None, feature_shape=None, attack_structure=True, attack_features=True, device='cpu'):
super(IGAttack, self).__init__(model, nnodes, attack_structure, attack_features, device)
assert attack_features or attack_structure, 'attack_features or attack_structure cannot be both False'
self.modified_adj = None
self.modified_features = None
self.target_node = None
def attack(self, ori_features, ori_adj, labels, idx_train, target_node, n_perturbations, steps=10, **kwargs):
"""Generate perturbations on the input graph.
Parameters
----------
ori_features :
Original (unperturbed) node feature matrix
ori_adj :
Original (unperturbed) adjacency matrix
labels :
node labels
idx_train:
training nodes indices
target_node : int
target node index to be attacked
n_perturbations : int
Number of perturbations on the input graph. Perturbations could
be edge removals/additions or feature removals/additions.
steps : int
steps for computing integrated gradients
"""
self.surrogate.eval()
self.target_node = target_node
modified_adj = ori_adj.todense()
modified_features = ori_features.todense()
adj, features, labels = utils.to_tensor(modified_adj, modified_features, labels, device=self.device)
adj_norm = utils.normalize_adj_tensor(adj)
pseudo_labels = self.surrogate.predict().detach().argmax(1)
pseudo_labels[idx_train] = labels[idx_train]
self.pseudo_labels = pseudo_labels
s_e = np.zeros(adj.shape[1])
s_f = np.zeros(features.shape[1])
if self.attack_structure:
s_e = self.calc_importance_edge(features, adj_norm, labels, steps)
if self.attack_features:
s_f = self.calc_importance_feature(features, adj_norm, labels, steps)
for t in (range(n_perturbations)):
s_e_max = np.argmax(s_e)
s_f_max = np.argmax(s_f)
if s_e[s_e_max] >= s_f[s_f_max]:
value = np.abs(1 - modified_adj[target_node, s_e_max])
modified_adj[target_node, s_e_max] = value
modified_adj[s_e_max, target_node] = value
s_e[s_e_max] = 0
else:
modified_features[target_node, s_f_max] = np.abs(1 - modified_features[target_node, s_f_max])
s_f[s_f_max] = 0
self.modified_adj = sp.csr_matrix(modified_adj)
self.modified_features = sp.csr_matrix(modified_features)
self.check_adj(modified_adj)
def calc_importance_edge(self, features, adj_norm, labels, steps):
"""Calculate integrated gradient for edges. Although I think the the gradient should be
with respect to adj instead of adj_norm, but the calculation is too time-consuming. So I
finally decided to calculate the gradient of loss with respect to adj_norm
"""
baseline_add = adj_norm.clone()
baseline_remove = adj_norm.clone()
baseline_add.data[self.target_node] = 1
baseline_remove.data[self.target_node] = 0
adj_norm.requires_grad = True
integrated_grad_list = []
i = self.target_node
for j in tqdm(range(adj_norm.shape[1])):
if adj_norm[i][j]:
scaled_inputs = [baseline_remove + (float(k)/ steps) * (adj_norm - baseline_remove) for k in range(0, steps + 1)]
else:
scaled_inputs = [baseline_add - (float(k)/ steps) * (baseline_add - adj_norm) for k in range(0, steps + 1)]
_sum = 0
for new_adj in scaled_inputs:
output = self.surrogate(features, new_adj)
loss = F.nll_loss(output[[self.target_node]],
self.pseudo_labels[[self.target_node]])
adj_grad = torch.autograd.grad(loss, adj_norm)[0]
adj_grad = adj_grad[i][j]
_sum += adj_grad
if adj_norm[i][j]:
avg_grad = (adj_norm[i][j] - 0) * _sum.mean()
else:
avg_grad = (1 - adj_norm[i][j]) * _sum.mean()
integrated_grad_list.append(avg_grad.detach().item())
integrated_grad_list[i] = 0
# make impossible perturbation to be negative
integrated_grad_list = np.array(integrated_grad_list)
adj = (adj_norm > 0).cpu().numpy()
integrated_grad_list = (-2 * adj[self.target_node] + 1) * integrated_grad_list
integrated_grad_list[self.target_node] = -10
return integrated_grad_list
def calc_importance_feature(self, features, adj_norm, labels, steps):
"""Calculate integrated gradient for features
"""
baseline_add = features.clone()
baseline_remove = features.clone()
baseline_add.data[self.target_node] = 1
baseline_remove.data[self.target_node] = 0
features.requires_grad = True
integrated_grad_list = []
i = self.target_node
for j in tqdm(range(features.shape[1])):
if features[i][j]:
scaled_inputs = [baseline_add + (float(k)/ steps) * (features - baseline_add) for k in range(0, steps + 1)]
else:
scaled_inputs = [baseline_remove - (float(k)/ steps) * (baseline_remove - features) for k in range(0, steps + 1)]
_sum = 0
for new_features in scaled_inputs:
output = self.surrogate(new_features, adj_norm)
loss = F.nll_loss(output[[self.target_node]],
self.pseudo_labels[[self.target_node]])
feature_grad = torch.autograd.grad(loss, features)[0]
feature_grad = feature_grad[i][j]
_sum += feature_grad
if features[i][j]:
avg_grad = (features[i][j] - 0) * _sum.mean()
else:
avg_grad = (1 - features[i][j]) * _sum.mean()
integrated_grad_list.append(avg_grad.detach().item())
# make impossible perturbation to be negative
features = (features > 0).cpu().numpy()
integrated_grad_list = np.array(integrated_grad_list)
integrated_grad_list = (-2 * features[self.target_node] + 1) * integrated_grad_list
return integrated_grad_list