def main():
parser = argparse.ArgumentParser()
parser.add_argument('--cuda', type=int, default=0, help='cuda')
parser.add_argument('--seed',
type=int,
default=123,
help='Random seed for model')
parser.add_argument('--data_seed',
type=int,
default=123,
help='Random seed for data split')
parser.add_argument('--dataset', type=str, default='cora', help='dataset')
parser.add_argument('--gnn_path',
type=str,
required=True,
help='Path of saved model')
parser.add_argument(
'--model', type=str, default='minmax', help='model variant'
) # ['minmax', 'Meta-Self', 'A-Meta-Self', 'Meta-Train', 'A-Meta-Train', 'random']
parser.add_argument('--loss_type',
type=str,
default='CE',
help='loss type')
parser.add_argument('--att_lr',
type=float,
default=200,
help='Initial learning rate')
parser.add_argument('--perturb_epochs',
type=int,
default=200,
help='Number of epochs to poisoning loop')
parser.add_argument('--ptb_rate',
type=float,
default=0.05,
help='pertubation rate')
parser.add_argument('--loss_weight',
type=float,
default=1.0,
help='loss weight')
parser.add_argument('--reg_weight',
type=float,
default=0.0,
help='regularization weight')
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('--dropout',
type=float,
default=0.0,
help='Dropout rate (1 - keep probability)')
parser.add_argument('--data_dir',
type=str,
default='./tmp/',
help='Directory to download dataset')
parser.add_argument('--target_node',
type=str,
default='train',
help='target node set')
parser.add_argument('--sanitycheck',
type=str,
default='no',
help='whether store the intermediate results')
parser.add_argument('--distance_type',
type=str,
default='l2',
help='distance type')
parser.add_argument('--opt_type',
type=str,
default='max',
help='optimization type')
parser.add_argument('--sample_type',
type=str,
default='sample',
help='sample type')
args = parser.parse_args()
args.device = torch.device(
f'cuda:{args.cuda}' if torch.cuda.is_available() else 'cpu')
torch.set_num_threads(1) # limit cpu use
set_random_seed(args.seed, args.device)
if not os.path.exists(args.data_dir):
os.mkdir(args.data_dir)
if not os.path.exists(args.gnn_path):
raise AssertionError(f'No trained model found under {args.gnn_path}!')
print('==== Environment ====')
print(f'torch version: {torch.__version__}')
print(f'device: {args.device}')
print(f'torch seed: {args.seed}')
#########################################################
# Load data for node classification task
data = Dataset(root=args.data_dir,
name=args.dataset,
setting='gcn',
seed=args.data_seed)
adj, features, labels = data.process(process_adj=False,
process_feature=False,
device=args.device)
idx_train, idx_val, idx_test = data.idx_train, data.idx_val, data.idx_test
idx_unlabeled = np.union1d(idx_val, idx_test)
print('==== Dataset ====')
print(f'density: {nx.density(nx.from_numpy_array(adj.cpu().numpy()))}')
print(f'adj shape: {adj.shape}')
print(f'feature shape: {features.shape}')
print(f'label number: {labels.max().item()+1}')
print(f'split seed: {args.data_seed}')
print(
f'train|valid|test set: {idx_train.shape}|{idx_val.shape}|{idx_test.shape}'
)
#########################################################
# Load victim model and test it on clean training nodes
weight_decay = 0 if args.dataset == 'polblogs' else args.weight_decay
victim_model = GCN(nfeat=features.shape[1],
nclass=labels.max().item() + 1,
nhid=args.hidden,
dropout=args.dropout,
weight_decay=weight_decay,
device=args.device)
victim_model = victim_model.to(args.device)
victim_model.load_state_dict(torch.load(args.gnn_path))
surrogate_model = GCN(nfeat=features.shape[1],
nclass=labels.max().item() + 1,
nhid=args.hidden,
dropout=args.dropout,
weight_decay=weight_decay,
device=args.device)
surrogate_model = surrogate_model.to(args.device)
surrogate_model.load_state_dict(torch.load(args.gnn_path))
print('==== Initial Surrogate Model on Clean Graph ====')
surrogate_model.eval()
check_victim_model_performance(surrogate_model, features, adj, labels,
idx_test, idx_train)
#########################################################
# Setup attack model
if args.model == 'minmax':
model = MinMax(model=surrogate_model,
nnodes=adj.shape[0],
loss_type=args.loss_type,
loss_weight=args.loss_weight,
regularization_weight=args.reg_weight,
device=args.device)
model = model.to(args.device)
elif 'Meta' in args.model: # 'Meta-Self', 'A-Meta-Self', 'Meta-Train', 'A-Meta-Train'
if 'Self' in args.model:
lambda_ = 0
if 'Train' in args.model:
lambda_ = 1
if 'Both' in args.model:
lambda_ = 0.5
if 'A' in args.model:
model = MetaApprox(model=surrogate_model,
nnodes=adj.shape[0],
attack_structure=True,
attack_features=False,
regularization_weight=args.reg_weight,
device=args.device,
lambda_=lambda_)
else:
model = Metattack(model=surrogate_model,
nnodes=adj.shape[0],
attack_structure=True,
attack_features=False,
regularization_weight=args.reg_weight,
device=args.device,
lambda_=lambda_)
model = model.to(args.device)
elif args.model == 'random':
model = Random()
else:
raise AssertionError(f'Attack {args.model} not found!')
#########################################################
# Attack and evaluate
print('***************** seed {} *****************'.format(args.seed))
print('==== Attacking ====')
perturbations = int(args.ptb_rate * (adj.sum() / 2))
nat_adj = copy.deepcopy(adj)
# switch target node set for minmax attack
if args.target_node == 'test':
idx_target = idx_test
elif args.target_node == 'train':
idx_target = idx_train
else:
idx_target = np.hstack((idx_test, idx_train)).astype(np.int)
# Start attack
if args.model == 'random':
model.attack(nat_adj, perturbations, 'flip')
elif 'Meta' in args.model:
model.attack(features,
nat_adj,
labels,
idx_train,
idx_unlabeled,
perturbations,
ll_constraint=False,
verbose=True)
else:
model.attack(features,
nat_adj,
labels,
idx_target,
perturbations,
att_lr=args.att_lr,
epochs=args.perturb_epochs,
distance_type=args.distance_type,
sample_type=args.sample_type,
opt_type=args.opt_type)
modified_adj = model.modified_adj
# evaluation
#########################################################
# vitim model on clean graph
print('==== Victim Model on clean graph ====')
check_victim_model_performance(victim_model, features, adj, labels,
idx_test, idx_train)
# vitim model on perturbed graph
print('==== Victim Model on perturbed graph ====')
check_victim_model_performance(victim_model, features, modified_adj,
labels, idx_test, idx_train)
# retrain victim model on perturbed graph
print('==== Poisoned Surrogate Model on perturbed graph ====')
surrogate_model.initialize()
surrogate_model.fit(features,
modified_adj,
labels,
idx_train,
idx_val=None,
train_iters=1000,
verbose=False)
check_victim_model_performance(surrogate_model, features, modified_adj,
labels, idx_test, idx_train)
# test poisoned model on clean graph
print('==== Poisoned Surrogate Model on clean graph ====')
check_victim_model_performance(surrogate_model, features, adj, labels,
idx_test, idx_train)
print("==== Parameter ====")
print(f'seed: data {args.data_seed}, attack {args.seed}')
print(
f'dataset: {args.dataset}, attack model {args.model}, target {args.target_node}'
)
print(f'loss type: {args.loss_type}')
print(
f'perturbation rate: {args.ptb_rate}, epoch: {args.perturb_epochs}, lr: {args.att_lr}'
)
print(f'weight: loss {args.loss_weight}, reg {args.reg_weight}')
print(f'distance type: {args.distance_type}, opt type: {args.opt_type}')
# if you want to save the modified adj/features, uncomment the code below
if args.sanitycheck == 'yes':
root = './sanitycheck_evasion/{}_{}_{}_{}_{}_{}lr_{}epoch_{}rate_{}reg_{}target_{}seed'
root = root.format(args.dataset, args.distance_type, args.sample_type,
args.model, args.loss_type, args.att_lr,
args.perturb_epochs, args.ptb_rate, args.reg_weight,
args.target_node, args.seed)
save_all(root, model)
surrogate = surrogate.to(device)
surrogate.fit(features, adj, labels, idx_train)
# Setup Attack Model
if 'Self' in args.model:
lambda_ = 0
if 'Train' in args.model:
lambda_ = 1
if 'Both' in args.model:
lambda_ = 0.5
if 'A' in args.model:
model = MetaApprox(model=surrogate,
nnodes=adj.shape[0],
feature_shape=features.shape,
attack_structure=True,
attack_features=False,
device=device,
lambda_=lambda_)
else:
model = Metattack(model=surrogate,
nnodes=adj.shape[0],
feature_shape=features.shape,
attack_structure=True,
attack_features=False,
device=device,
lambda_=lambda_)
model = model.to(device)