/
CLGA.py
244 lines (223 loc) · 11.5 KB
/
CLGA.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
from torch.nn.modules.module import Module
import argparse
import os.path as osp
import time
import pickle as pkl
import torch
from torch_geometric.utils import dropout_adj, degree, to_undirected, dense_to_sparse, is_undirected, to_networkx, contains_isolated_nodes
from simple_param.sp import SimpleParam
from pGRACE.functional import drop_feature, drop_edge_weighted, \
degree_drop_weights, evc_drop_weights, pr_drop_weights, \
feature_drop_weights, drop_feature_weighted, feature_drop_weights_dense
from pGRACE.utils import get_activation, compute_pr, eigenvector_centrality
from pGRACE.dataset import get_dataset
from differentiable_models.gcn import GCN
from differentiable_models.model import GRACE
class Metacl(Module):
def __init__(self, args, dataset, param, device):
super(Metacl, self).__init__()
self.model = None
self.optimizer = None
self.param = param
self.args = args
self.device = device
self.dataset = dataset
self.data = dataset.data.to(device)
self.drop_weights = None
self.feature_weights = None
def drop_edge(self, p):
if self.param['drop_scheme'] == 'uniform':
return dropout_adj(self.data.edge_index, p=p)[0]
elif self.param['drop_scheme'] in ['degree', 'evc', 'pr']:
return drop_edge_weighted(self.data.edge_index, self.drop_weights, p=p,
threshold=0.7)
else:
raise Exception(f'undefined drop scheme: {param["drop_scheme"]}')
def train_gcn(self):
self.model.train()
self.optimizer.zero_grad()
edge_index_1 = self.drop_edge(self.param['drop_edge_rate_1'])
edge_index_2 = self.drop_edge(self.param['drop_edge_rate_2'])
x_1 = drop_feature(self.data.x, self.param['drop_feature_rate_1'])
x_2 = drop_feature(self.data.x, self.param['drop_feature_rate_2'])
edge_sp_adj_1 = torch.sparse.FloatTensor(edge_index_1,
torch.ones(edge_index_1.shape[1]).to(self.device),
[self.data.num_nodes, self.data.num_nodes])
edge_sp_adj_2 = torch.sparse.FloatTensor(edge_index_2,
torch.ones(edge_index_2.shape[1]).to(self.device),
[self.data.num_nodes, self.data.num_nodes])
if self.param['drop_scheme'] in ['pr', 'degree', 'evc']:
x_1 = drop_feature_weighted(self.data.x, self.feature_weights, self.param['drop_feature_rate_1'])
x_2 = drop_feature_weighted(self.data.x, self.feature_weights, self.param['drop_feature_rate_2'])
z1 = self.model(x_1, edge_sp_adj_1, sparse=True)
z2 = self.model(x_2, edge_sp_adj_2, sparse=True)
loss = self.model.loss(z1, z2, batch_size=None)
loss.backward()
self.optimizer.step()
return loss.item()
def compute_drop_weights(self):
if self.param['drop_scheme'] == 'degree':
self.drop_weights = degree_drop_weights(self.data.edge_index).to(self.device)
elif self.param['drop_scheme'] == 'pr':
self.drop_weights = pr_drop_weights(self.data.edge_index, aggr='sink', k=200).to(self.device)
elif self.param['drop_scheme'] == 'evc':
self.drop_weights = evc_drop_weights(self.data).to(self.device)
else:
self.drop_weights = None
if self.param['drop_scheme'] == 'degree':
edge_index_ = to_undirected(self.data.edge_index)
node_deg = degree(edge_index_[1])
if self.args.dataset == 'WikiCS':
self.feature_weights = feature_drop_weights_dense(self.data.x, node_c=node_deg).to(self.device)
else:
self.feature_weights = feature_drop_weights(self.data.x, node_c=node_deg).to(self.device)
elif self.param['drop_scheme'] == 'pr':
node_pr = compute_pr(self.data.edge_index)
if self.args.dataset == 'WikiCS':
self.feature_weights = feature_drop_weights_dense(self.data.x, node_c=node_pr).to(self.device)
else:
self.feature_weights = feature_drop_weights(self.data.x, node_c=node_pr).to(self.device)
elif self.param['drop_scheme'] == 'evc':
node_evc = eigenvector_centrality(self.data)
if self.args.dataset == 'WikiCS':
self.feature_weights = feature_drop_weights_dense(self.data.x, node_c=node_evc).to(self.device)
else:
self.feature_weights = feature_drop_weights(self.data.x, node_c=node_evc).to(self.device)
else:
self.feature_weights = torch.ones((self.data.x.size(1),)).to(self.device)
def inner_train(self):
encoder = GCN(self.dataset.num_features, self.param['num_hidden'], get_activation(self.param['activation']))
self.model = GRACE(encoder, self.param['num_hidden'], self.param['num_proj_hidden'], self.param['tau']).to(self.device)
self.optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.param['learning_rate'],
weight_decay=self.param['weight_decay']
)
self.compute_drop_weights()
for epoch in range(1, self.param['num_epochs'] + 1):
loss = self.train_gcn()
def compute_gradient(self, pe1, pe2, pf1, pf2):
self.model.eval()
self.compute_drop_weights()
edge_index_1 = self.drop_edge(pe1)
edge_index_2 = self.drop_edge(pe2)
x_1 = drop_feature(self.data.x, pf1)
x_2 = drop_feature(self.data.x, pf2)
edge_sp_adj_1 = torch.sparse.FloatTensor(edge_index_1,
torch.ones(edge_index_1.shape[1]).to(self.device),
[self.data.num_nodes, self.data.num_nodes])
edge_sp_adj_2 = torch.sparse.FloatTensor(edge_index_2,
torch.ones(edge_index_2.shape[1]).to(self.device),
[self.data.num_nodes, self.data.num_nodes])
if self.param['drop_scheme'] in ['pr', 'degree', 'evc']:
x_1 = drop_feature_weighted(self.data.x, self.feature_weights, pf1)
x_2 = drop_feature_weighted(self.data.x, self.feature_weights, pf2)
edge_adj_1 = edge_sp_adj_1.to_dense()
edge_adj_2 = edge_sp_adj_2.to_dense()
edge_adj_1.requires_grad = True
edge_adj_2.requires_grad = True
z1 = self.model(x_1, edge_adj_1, sparse=False)
z2 = self.model(x_2, edge_adj_2, sparse=False)
loss = self.model.loss(z1, z2, batch_size=None)
loss.backward()
return edge_adj_1.grad, edge_adj_2.grad
def attack(self):
perturbed_edges = []
num_total_edges = self.data.num_edges
adj_sp = torch.sparse.FloatTensor(self.data.edge_index, torch.ones(self.data.edge_index.shape[1]).to(self.device),
[self.data.num_nodes, self.data.num_nodes])
adj = adj_sp.to_dense()
print('Begin perturbing.....')
# save three poisoned adj when the perturbation rate reaches 1%, 5%, 10%
while len(perturbed_edges) < int(0.10 * num_total_edges):
if len(perturbed_edges) == int(0.01 * num_total_edges) or \
len(perturbed_edges) == int(0.01 * num_total_edges)-1 or \
len(perturbed_edges) == int(0.01 * num_total_edges)-2:
output_adj = adj.to(device)
pkl.dump(output_adj.to(torch.device('cpu')), open('poisoned_adj/%s_CLGA_0.010000_adj.pkl' % args.dataset, 'wb'))
print('---1% poisoned adjacency matrix saved---')
if len(perturbed_edges) == int(0.05 * num_total_edges) or \
len(perturbed_edges) == int(0.05 * num_total_edges)-1 or \
len(perturbed_edges) == int(0.05 * num_total_edges)-2:
output_adj = adj.to(device)
pkl.dump(output_adj.to(torch.device('cpu')), open('poisoned_adj/%s_CLGA_0.050000_adj.pkl' % args.dataset, 'wb'))
print('---5% perturbed adjacency matrix saved---')
start = time.time()
self.inner_train()
adj_1_grad, adj_2_grad = self.compute_gradient(self.param['drop_edge_rate_1'], self.param['drop_edge_rate_2'], self.param['drop_feature_rate_1'], self.param['drop_feature_rate_2'])
grad_sum = adj_1_grad + adj_2_grad
grad_sum_1d = grad_sum.view(-1)
grad_sum_1d_abs = torch.abs(grad_sum_1d)
values, indices = grad_sum_1d_abs.sort(descending=True)
i = -1
while True:
i += 1
index = int(indices[i])
row = int(index / self.data.num_nodes)
column = index % self.data.num_nodes
if [row, column] in perturbed_edges:
continue
if grad_sum_1d[index] < 0 and adj[row, column] == 1:
adj[row, column] = 0
adj[column, row] = 0
perturbed_edges.append([row, column])
perturbed_edges.append([column, row])
break
elif grad_sum_1d[index] > 0 and adj[row, column] == 0:
adj[row, column] = 1
adj[column, row] = 1
perturbed_edges.append([row, column])
perturbed_edges.append([column, row])
break
self.data.edge_index = dense_to_sparse(adj)[0]
end = time.time()
print('Perturbing edges: %d/%d. Finished in %.2fs' % (len(perturbed_edges)/2, int(0.10 * num_total_edges)/2, end-start))
print('Number of perturbed edges: %d' % (len(perturbed_edges)/2))
output_adj = adj.to(device)
return output_adj
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--device', type=str, default='cuda:0')
parser.add_argument('--dataset', type=str, default='Cora')
parser.add_argument('--param', type=str, default='local:general.json')
default_param = {
'learning_rate': 0.01,
'num_hidden': 256,
'num_proj_hidden': 32,
'activation': 'prelu',
'drop_edge_rate_1': 0.3,
'drop_edge_rate_2': 0.4,
'drop_feature_rate_1': 0.1,
'drop_feature_rate_2': 0.0,
'tau': 0.4,
'num_epochs': 3000,
'weight_decay': 1e-5,
'drop_scheme': 'degree',
}
# add hyper-parameters into parser
param_keys = default_param.keys()
for key in param_keys:
parser.add_argument(f'--{key}', type=type(default_param[key]), nargs='?')
args = parser.parse_args()
# parse param
sp = SimpleParam(default=default_param)
param = sp(source=args.param, preprocess='nni')
# merge cli arguments and parsed param
for key in param_keys:
if getattr(args, key) is not None:
param[key] = getattr(args, key)
use_nni = args.param == 'nni'
if use_nni and args.device != 'cpu':
args.device = 'cuda'
# torch_seed = args.seed
# torch.manual_seed(torch_seed)
# random.seed(12345)
device = torch.device(args.device)
path = osp.expanduser('dataset')
path = osp.join(path, args.dataset)
dataset = get_dataset(path, args.dataset)
data = dataset[0]
model = Metacl(args, dataset, param, device).to(device)
poisoned_adj = model.attack()
pkl.dump(poisoned_adj.to(torch.device('cpu')), open('poisoned_adj/%s_CLGA_0.100000_adj.pkl' % args.dataset, 'wb'))
print('---10% perturbed adjacency matrix saved---')