-
Notifications
You must be signed in to change notification settings - Fork 192
/
test_ugba.py
186 lines (160 loc) · 9.1 KB
/
test_ugba.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
import numpy as np
import argparse
import copy
import torch
import torch_geometric.transforms as T
from torch_geometric.utils import to_undirected
from ogb.nodeproppred import PygNodePropPredDataset
from torch_geometric.datasets import Planetoid
import deeprobust.graph.utils as utils
from deeprobust.graph.targeted_attack import UGBA
from deeprobust.graph.defense_pyg import GCN, SAGE, GAT
parser = argparse.ArgumentParser()
parser.add_argument('--no-cuda', action='store_true', default=False,
help='Disables CUDA training.')
parser.add_argument('--seed', type=int, default=10, help='Random seed.')
parser.add_argument('--model', type=str, default='GCN', help='model',
choices=['GCN','GAT','GraphSage','GIN'])
parser.add_argument('--dataset', type=str, default='ogbn-arxiv',
help='Dataset')
parser.add_argument('--train_lr', type=float, default=0.01,
help='Initial learning rate.')
parser.add_argument('--weight_decay', type=float, default=5e-4,
help='Weight decay (L2 loss on parameters).')
parser.add_argument('--hidden', type=int, default=32,
help='Number of hidden units.')
parser.add_argument('--thrd', type=float, default=0.5)
parser.add_argument('--target_class', type=int, default=0)
parser.add_argument('--dropout', type=float, default=0.5,
help='Dropout rate (1 - keep probability).')
parser.add_argument('--epochs', type=int, default=200, help='Number of epochs to train benign and backdoor model.')
parser.add_argument('--trojan_epochs', type=int, default=400, help='Number of epochs to train trigger generator.')
parser.add_argument('--inner', type=int, default=1, help='Number of inner')
# backdoor setting
parser.add_argument('--lr', type=float, default=0.01,
help='Initial learning rate.')
parser.add_argument('--trigger_size', type=int, default=3,
help='tirgger_size')
parser.add_argument('--use_vs_number', action='store_true', default=False,
help="if use detailed number to decide Vs")
parser.add_argument('--vs_ratio', type=float, default=0,
help="ratio of poisoning nodes relative to the full graph")
parser.add_argument('--vs_number', type=int, default=0,
help="number of poisoning nodes relative to the full graph")
# attack setting
parser.add_argument('--selection_method', type=str, default='none',
choices=['cluster','none'],
help='Method to select idx_attach for training trojan model (none means randomly select)')
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def subgraph(subset, edge_index, edge_attr = None, relabel_nodes: bool = False):
"""Returns the induced subgraph of :obj:`(edge_index, edge_attr)`
containing the nodes in :obj:`subset`.
Args:
subset (LongTensor, BoolTensor or [int]): The nodes to keep.
edge_index (LongTensor): The edge indices.
edge_attr (Tensor, optional): Edge weights or multi-dimensional
edge features. (default: :obj:`None`)
relabel_nodes (bool, optional): If set to :obj:`True`, the resulting
:obj:`edge_index` will be relabeled to hold consecutive indices
starting from zero. (default: :obj:`False`)
num_nodes (int, optional): The number of nodes, *i.e.*
:obj:`max_val + 1` of :attr:`edge_index`. (default: :obj:`None`)
:rtype: (:class:`LongTensor`, :class:`Tensor`)
"""
device = edge_index.device
node_mask = subset
edge_mask = node_mask[edge_index[0]] & node_mask[edge_index[1]]
edge_index = edge_index[:, edge_mask]
edge_attr = edge_attr[edge_mask] if edge_attr is not None else None
return edge_index, edge_attr, edge_mask
def get_split(data, device):
rs = np.random.RandomState(10)
perm = rs.permutation(data.num_nodes)
train_number = int(0.2*len(perm))
idx_train = torch.tensor(sorted(perm[:train_number])).to(device)
data.train_mask = torch.zeros_like(data.train_mask)
data.train_mask[idx_train] = True
val_number = int(0.1*len(perm))
idx_val = torch.tensor(sorted(perm[train_number:train_number+val_number])).to(device)
data.val_mask = torch.zeros_like(data.val_mask)
data.val_mask[idx_val] = True
test_number = int(0.2*len(perm))
idx_test = torch.tensor(sorted(perm[train_number+val_number:train_number+val_number+test_number])).to(device)
data.test_mask = torch.zeros_like(data.test_mask)
data.test_mask[idx_test] = True
idx_clean_test = idx_test[:int(len(idx_test)/2)]
idx_atk = idx_test[int(len(idx_test)/2):]
data.test_mask = torch.zeros_like(data.test_mask)
data.test_mask[idx_clean_test] = True
return data, idx_train, idx_val, idx_clean_test, idx_atk
dataset = Planetoid('./', 'cora')
dataset.transform = T.NormalizeFeatures()
data = dataset[0]
data, idx_train, idx_val, idx_clean_test, idx_atk = get_split(data,device)
# decide clean test nodes
data.test_mask = utils.index_to_mask(idx_clean_test,size=data.x.shape[0])
data = data.to(device)
# idx_train = data.train_mask.nonzero().flatten()
# idx_val = data.val_mask.nonzero().flatten()
# idx_test = data.test_mask.nonzero().flatten()
data.edge_index = to_undirected(data.edge_index, num_nodes = data.num_nodes)
train_edge_index, _, edge_mask = subgraph(torch.bitwise_not(data.test_mask),data.edge_index,relabel_nodes=False)
# filter out the unlabeled nodes except from training nodes and testing nodes, nonzero() is to get index, flatten is to get 1-d tensor
unlabeled_idx = (torch.bitwise_not(data.test_mask)&torch.bitwise_not(data.train_mask)).nonzero().flatten()
mask_edge_index = data.edge_index[:,torch.bitwise_not(edge_mask)]
test_model = GCN(nfeat=data.x.shape[1],
nhid=args.hidden,
nclass=data.y.max().item() + 1,
nlayers=2, lr=0.01,
dropout=0.5, device=device).to(device)
test_model.fit(data, train_iters=args.epochs, verbose = True)
'''get clean accuracy before attack'''
test_model.test()
'''Perform backdoor attack'''
agent = UGBA(data, vs_number = 10, device = 'cuda', trigger_size = 3,
homo_loss_weight = 0, homo_boost_thrd = 0.5, train_epochs = 200,
trojan_epochs = 400, dis_weight = -1.0)
# train trigger generator
trigger_generator, idx_attach = agent.train_trigger_generator(idx_train, train_edge_index, edge_weights = None, selection_method = 'cluster')
# update poisoned training graph
poison_data = agent.get_poisoned_graph()
# train backdoored GNN
test_model = GCN(nfeat=data.x.shape[1],
nhid=args.hidden,
nclass=data.y.max().item() + 1,
nlayers=2, lr=0.01,
dropout=0.5, device=device).to(device)
# evaluation: inject trigger to target nodes
induct_data = copy.deepcopy(poison_data)
test_model.fit(induct_data, train_iters=args.epochs, verbose = True)
induct_edge_index = torch.cat([poison_data.edge_index,mask_edge_index],dim=1)
induct_edge_weights = torch.cat([poison_data.edge_weights,torch.ones([mask_edge_index.shape[1]],dtype=torch.float,device=device)])
induct_data.edge_index, induct_data.edge_weights = induct_edge_index, induct_edge_weights
acc_test = test_model.test()
'''
Attach generated trigger with a single target node: UGBA.attack(target_node, features, labels, edge_index, edge_attr)
Example:
x, edge_index, edge_weights, y = agent.attack(idx_atk[0], data.x, data.y, data.edge_index, None)
'''
overall_induct_edge_index, overall_induct_edge_weights = induct_edge_index.clone(),induct_edge_weights.clone()
from torch_geometric.utils import k_hop_subgraph
asr = 0
for i, idx in enumerate(idx_atk):
idx=int(idx)
sub_induct_nodeset, sub_induct_edge_index, sub_mapping, sub_edge_mask = k_hop_subgraph(node_idx = [idx], num_hops = 2, edge_index = overall_induct_edge_index, relabel_nodes=True) # sub_mapping means the index of [idx] in sub)nodeset
ori_node_idx = sub_induct_nodeset[sub_mapping]
relabeled_node_idx = sub_mapping
sub_induct_edge_weights = torch.ones([sub_induct_edge_index.shape[1]]).to(device)
with torch.no_grad():
# inject trigger on attack test nodes (idx_atk)'''
induct_x, induct_edge_index, induct_edge_weights, induct_y = agent.inject_trigger(idx_attach = relabeled_node_idx,x = poison_data.x[sub_induct_nodeset],y = poison_data.y[sub_induct_nodeset],edge_index = sub_induct_edge_index,edge_weights = sub_induct_edge_weights)
induct_x, induct_edge_index, induct_edge_weights = induct_x.clone().detach(), induct_edge_index.clone().detach(),induct_edge_weights.clone().detach()
# attack evaluation
output = test_model(induct_x,induct_edge_index,induct_edge_weights)
train_attach_rate = (output.argmax(dim=1)[relabeled_node_idx]==args.target_class).float().mean()
asr += train_attach_rate
induct_x, induct_edge_index,induct_edge_weights = induct_x.cpu(), induct_edge_index.cpu(),induct_edge_weights.cpu()
output = output.cpu()
asr = asr/(idx_atk.shape[0])
print("Overall ASR: {:.4f}".format(asr))