def main():
with open('data.json') as data_json:
data_params = json.load(data_json)
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str)
parser.add_argument('--data_path', type=str, default='data')
parser.add_argument('--output_path', type=str, default='results')
parser.add_argument('--pretrained', type=str, required=True)
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--attack', type=str, required=True, choices=data_params['attacks'])
parser.add_argument('--eps', type=float, default=0.3)
# NOTE: In CW_L2 attack, eps is the upper bound of c.
parser.add_argument('--n_samples', type=int, default=2000)
parser.add_argument('--random_state', type=int, default=1234)
args = parser.parse_args()
print(args)
set_seeds(args.random_state)
if not os.path.exists(args.output_path):
print('Output folder does not exist. Create:', args.output_path)
os.mkdir(args.output_path)
print('Dataset:', args.data)
print('Pretrained model:', args.pretrained)
print('Running attack: {}'.format(args.attack))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: {}'.format(device))
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if args.data == 'mnist':
dataset_train = datasets.MNIST(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.MNIST(args.data_path, train=False, download=True, transform=transforms)
elif args.data == 'cifar10':
dataset_train = datasets.CIFAR10(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.CIFAR10(args.data_path, train=False, download=True, transform=transforms)
else:
data_path = os.path.join(args.data_path, data_params['data'][args.data]['file_name'])
print('Read file:', data_path)
X, y = load_csv(data_path)
X_train, X_test, y_train, y_test = train_test_split(
X, y,
test_size=data_params['data'][args.data]['n_test'],
random_state=args.random_state)
scaler = MinMaxScaler().fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
dataset_train = TensorDataset(torch.from_numpy(X_train).type(torch.float32), torch.from_numpy(y_train).type(torch.long))
dataset_test = TensorDataset(torch.from_numpy(X_test).type(torch.float32), torch.from_numpy(y_test).type(torch.long))
dataloader_train = DataLoader(dataset_train, 256, shuffle=False)
dataloader_test = DataLoader(dataset_test, 256, shuffle=False)
shape_train = get_shape(dataloader_train.dataset)
shape_test = get_shape(dataloader_test.dataset)
print('Train set:', shape_train)
print('Test set:', shape_test)
# Load model
use_prob = args.attack not in ['apgd', 'apgd1', 'apgd2', 'cw2', 'cwinf']
print('Attack:', args.attack)
print('Using softmax layer:', use_prob)
if args.data == 'mnist':
model = BaseModel(use_prob=use_prob).to(device)
model_name = 'basic'
elif args.data == 'cifar10':
model_name = args.pretrained.split('_')[1]
if model_name == 'resnet':
model = Resnet(use_prob=use_prob).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=use_prob).to(device)
else:
raise ValueError('Unknown model: {}'.format(model_name))
else:
n_features = data_params['data'][args.data]['n_features']
n_classes = data_params['data'][args.data]['n_classes']
model = NumericModel(
n_features,
n_hidden=n_features * 4,
n_classes=n_classes,
use_prob=use_prob).to(device)
model_name = 'basic' + str(n_features * 4)
optimizer = optim.SGD(model.parameters(), lr=0.01,
momentum=0.9, weight_decay=5e-4)
loss = nn.CrossEntropyLoss()
pretrained_path = os.path.join(args.output_path, args.pretrained)
model.load_state_dict(torch.load(pretrained_path, map_location=device))
_, acc_train = validate(model, dataloader_train, loss, device)
_, acc_test = validate(model, dataloader_test, loss, device)
print('Accuracy on train set: {:.4f}%'.format(acc_train * 100))
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
# Create a subset which only contains recognisable samples.
tensor_test_X, tensor_test_y = get_correct_examples(
model, dataset_test, device=device, return_tensor=True)
dataset_perfect = TensorDataset(tensor_test_X, tensor_test_y)
loader_perfect = DataLoader(dataset_perfect, batch_size=512, shuffle=False)
_, acc_perfect = validate(model, loader_perfect, loss, device)
print('Accuracy on {} filtered test examples: {:.4f}%'.format(
len(dataset_perfect), acc_perfect * 100))
# Generate adversarial examples
n_features = data_params['data'][args.data]['n_features']
n_classes = data_params['data'][args.data]['n_classes']
if isinstance(n_features, int):
n_features = (n_features,)
classifier = PyTorchClassifier(
model=model,
loss=loss,
input_shape=n_features,
optimizer=optimizer,
nb_classes=n_classes,
clip_values=(0.0, 1.0),
device_type='gpu')
if args.attack == 'apgd':
eps_step = args.eps / 10.0 if args.eps <= 0.1 else 0.1
attack = AutoProjectedGradientDescent(
estimator=classifier,
eps=args.eps,
eps_step=eps_step,
max_iter=1000,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'apgd1':
attack = AutoProjectedGradientDescent(
estimator=classifier,
norm=1,
eps=args.eps,
eps_step=0.1,
max_iter=1000,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'apgd2':
attack = AutoProjectedGradientDescent(
estimator=classifier,
norm=2,
eps=args.eps,
eps_step=0.1,
max_iter=1000,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'bim':
eps_step = args.eps / 10.0
attack = BasicIterativeMethod(
estimator=classifier,
eps=args.eps,
eps_step=eps_step,
max_iter=1000,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'boundary':
attack = BoundaryAttack(
estimator=classifier,
max_iter=1000,
sample_size=args.batch_size,
targeted=False)
elif args.attack == 'cw2':
# NOTE: Do NOT increase the batch size!
attack = CarliniWagnerAttackL2(
model=model,
n_classes=n_classes,
confidence=args.eps,
verbose=True,
check_prob=False,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'cwinf':
attack = CarliniLInfMethod(
classifier=classifier,
confidence=args.eps,
max_iter=1000,
batch_size=args.batch_size,
targeted=False)
elif args.attack == 'deepfool':
attack = DeepFool(
classifier=classifier,
epsilon=args.eps,
batch_size=args.batch_size)
elif args.attack == 'fgsm':
attack = FastGradientMethod(
estimator=classifier,
eps=args.eps,
batch_size=args.batch_size)
elif args.attack == 'jsma':
attack = SaliencyMapMethod(
classifier=classifier,
gamma=args.eps,
batch_size=args.batch_size)
elif args.attack == 'line':
if args.data == 'mnist':
color = args.eps
elif args.data == 'cifar10':
color = (args.eps, args.eps, args.eps)
else:
raise NotImplementedError
attack = LineAttack(color=color, thickness=1)
elif args.attack == 'shadow':
attack = ShadowAttack(
estimator=classifier,
batch_size=args.batch_size,
targeted=False,
verbose=False)
elif args.attack == 'watermark':
attack = WaterMarkAttack(
eps=args.eps,
n_classes=data_params['data'][args.data]['n_classes'],
x_min=0.0,
x_max=1.0,
targeted=False)
X_train, y_train = get_correct_examples(model, dataset_train, device=device, return_tensor=True)
X_train = X_train.cpu().detach().numpy()
y_train = y_train.cpu().detach().numpy()
attack.fit(X_train, y_train)
else:
raise NotImplementedError
if len(dataset_perfect) > args.n_samples:
n = args.n_samples
else:
n = len(dataset_perfect)
X_benign = tensor_test_X[:n].cpu().detach().numpy()
y = tensor_test_y[:n].cpu().detach().numpy()
print('Creating {} adversarial examples with eps={} (Not all attacks use eps)'.format(n, args.eps))
time_start = time.time()
# Shadow attack only takes single sample!
if args.attack == 'shadow':
adv = np.zeros_like(X_benign)
for i in trange(len(X_benign)):
adv[i] = attack.generate(x=np.expand_dims(X_benign[i], axis=0))
elif args.attack == 'watermark':
# This is untargeted.
adv = attack.generate(X_benign, y)
else:
adv = attack.generate(x=X_benign)
time_elapsed = time.time() - time_start
print('Total time spend: {}'.format(str(datetime.timedelta(seconds=time_elapsed))))
pred_benign = np.argmax(classifier.predict(X_benign), axis=1)
acc_benign = np.sum(pred_benign == y) / n
pred_adv = np.argmax(classifier.predict(adv), axis=1)
acc_adv = np.sum(pred_adv == y) / n
print("Accuracy on benign samples: {:.4f}%".format(acc_benign * 100))
print("Accuracy on adversarial examples: {:.4f}%".format(acc_adv * 100))
# Save results
if args.n_samples < 2000:
output_file = '{}_{}_{}_{}_size{}'.format(args.data, model_name, args.attack, str(args.eps), args.n_samples)
else:
output_file = '{}_{}_{}_{}'.format(args.data, model_name, args.attack, str(args.eps))
path_x = os.path.join(args.output_path, '{}_x.npy'.format(output_file))
path_y = os.path.join(args.output_path, '{}_y.npy'.format(output_file))
path_adv = os.path.join(args.output_path, '{}_adv.npy'.format(output_file))
np.save(path_x, X_benign)
np.save(path_y, y)
np.save(path_adv, adv)
print('Saved to:', '{}_adv.npy'.format(output_file))
print()
def main():
with open('data.json') as data_json:
data_params = json.load(data_json)
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, required=True)
parser.add_argument('--model', type=str, required=True)
parser.add_argument('--pretrained', type=str, required=True)
parser.add_argument('--data_path', type=str, default='data')
parser.add_argument('--output_path', type=str, default='results')
parser.add_argument('--random_state', type=int, default=1234)
args = parser.parse_args()
print(args)
set_seeds(args.random_state)
if not os.path.exists(args.output_path):
print('Output folder does not exist. Create:', args.output_path)
os.mkdir(args.output_path)
print('data:', args.data)
print('model:', args.model)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: {}'.format(device))
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if args.data == 'mnist':
dataset_train = datasets.MNIST(args.data_path,
train=True,
download=True,
transform=transforms)
dataset_test = datasets.MNIST(args.data_path,
train=False,
download=True,
transform=transforms)
elif args.data == 'cifar10':
dataset_train = datasets.CIFAR10(args.data_path,
train=True,
download=True,
transform=transforms)
dataset_test = datasets.CIFAR10(args.data_path,
train=False,
download=True,
transform=transforms)
else:
data_path = os.path.join(args.data_path,
data_params['data'][args.data]['file_name'])
print('Read file:', data_path)
X, y = load_csv(data_path)
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=data_params['data'][args.data]['n_test'],
random_state=args.random_state)
scaler = MinMaxScaler().fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
dataset_train = TensorDataset(
torch.from_numpy(X_train).type(torch.float32),
torch.from_numpy(y_train).type(torch.long))
dataset_test = TensorDataset(
torch.from_numpy(X_test).type(torch.float32),
torch.from_numpy(y_test).type(torch.long))
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=False)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
shape_train = get_shape(loader_train.dataset)
shape_test = get_shape(loader_test.dataset)
print('Train set:', shape_train)
print('Test set:', shape_test)
use_prob = True
print('Using softmax layer:', use_prob)
# Load model
if args.data == 'mnist':
model = BaseModel(use_prob=use_prob).to(device)
model_name = 'basic'
elif args.data == 'cifar10':
model_name = args.pretrained.split('_')[1]
if model_name == 'resnet':
model = Resnet(use_prob=use_prob).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=use_prob).to(device)
else:
raise NotImplementedError
else:
n_features = data_params['data'][args.data]['n_features']
n_classes = data_params['data'][args.data]['n_classes']
model = NumericModel(n_features,
n_hidden=n_features * 4,
n_classes=n_classes,
use_prob=use_prob).to(device)
model_name = 'basic' + str(n_features * 4)
loss = nn.CrossEntropyLoss()
pretrained_path = os.path.join(args.output_path, args.pretrained)
model.load_state_dict(torch.load(pretrained_path, map_location=device))
_, acc_train = validate(model, loader_train, loss, device)
_, acc_test = validate(model, loader_test, loss, device)
print('Accuracy on train set: {:.4f}%'.format(acc_train * 100))
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
# Create a subset which only contains recognisable samples.
# The original train and test sets are no longer needed.
tensor_train_X, tensor_train_y = get_correct_examples(model,
dataset_train,
device=device,
return_tensor=True)
dataset_train = TensorDataset(tensor_train_X, tensor_train_y)
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=False)
_, acc_perfect = validate(model, loader_train, loss, device)
print('Accuracy on {} filtered train set: {:.4f}%'.format(
len(dataset_train), acc_perfect * 100))
tensor_test_X, tensor_test_y = get_correct_examples(model,
dataset_test,
device=device,
return_tensor=True)
dataset_test = TensorDataset(tensor_test_X, tensor_test_y)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
_, acc_perfect = validate(model, loader_test, loss, device)
print('Accuracy on {} filtered test set: {:.4f}%'.format(
len(dataset_test), acc_perfect * 100))
X_train = tensor_train_X.cpu().detach().numpy()
y_train = tensor_train_y.cpu().detach().numpy()
X_baard = baard_preprocess(args.data,
tensor_train_X).cpu().detach().numpy()
obj = {'X_train': X_baard, 'y_train': y_train}
path_ouput = os.path.join(
args.output_path,
'{}_{}_baard_train.pt'.format(args.data, args.model, args.model))
torch.save(obj, path_ouput)
print('Save to:', path_ouput)
print()
def main():
with open('data.json') as data_json:
data_params = json.load(data_json)
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, required=True)
parser.add_argument('--data_path', type=str, default='data')
parser.add_argument('--output_path', type=str, default='results')
parser.add_argument('--pretrained', type=str, required=True)
parser.add_argument('--adv',
type=str,
required=True,
help="Example: 'mnist_basic_apgd_0.3'")
parser.add_argument('--random_state', type=int, default=1234)
args = parser.parse_args()
print(args)
set_seeds(args.random_state)
print('Dataset:', args.data)
print('Pretrained model:', args.pretrained)
print('Pretrained samples:', args.adv + '_adv.npy')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: {}'.format(device))
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if args.data == 'mnist':
dataset_train = datasets.MNIST(args.data_path,
train=True,
download=True,
transform=transforms)
dataset_test = datasets.MNIST(args.data_path,
train=False,
download=True,
transform=transforms)
elif args.data == 'cifar10':
dataset_train = datasets.CIFAR10(args.data_path,
train=True,
download=True,
transform=transforms)
dataset_test = datasets.CIFAR10(args.data_path,
train=False,
download=True,
transform=transforms)
else:
data_path = os.path.join(args.data_path,
data_params['data'][args.data]['file_name'])
print('Read file:', data_path)
X, y = load_csv(data_path)
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=data_params['data'][args.data]['n_test'],
random_state=args.random_state)
scaler = MinMaxScaler().fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
dataset_train = TensorDataset(
torch.from_numpy(X_train).type(torch.float32),
torch.from_numpy(y_train).type(torch.long))
dataset_test = TensorDataset(
torch.from_numpy(X_test).type(torch.float32),
torch.from_numpy(y_test).type(torch.long))
# Note: Train set alway shuffle!
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=True)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
shape_train = get_shape(loader_train.dataset)
shape_test = get_shape(loader_test.dataset)
print('Train set:', shape_train)
print('Test set:', shape_test)
use_prob = True
print('Using softmax layer:', use_prob)
n_classes = data_params['data'][args.data]['n_classes']
# Load model
if args.data == 'mnist':
model = BaseModel(use_prob=use_prob).to(device)
model_name = 'basic'
elif args.data == 'cifar10':
model_name = args.pretrained.split('_')[1]
if model_name == 'resnet':
model = Resnet(use_prob=use_prob).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=use_prob).to(device)
else:
raise ValueError('Unknown model: {}'.format(model_name))
else:
n_features = data_params['data'][args.data]['n_features']
model = NumericModel(n_features,
n_hidden=n_features * 4,
n_classes=n_classes,
use_prob=use_prob).to(device)
model_name = 'basic' + str(n_features * 4)
loss = nn.CrossEntropyLoss()
pretrained_path = os.path.join(args.output_path, args.pretrained)
model.load_state_dict(torch.load(pretrained_path))
_, acc_train = validate(model, loader_train, loss, device)
_, acc_test = validate(model, loader_test, loss, device)
print('Accuracy on train set: {:.4f}%'.format(acc_train * 100))
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
# Create a subset which only contains recognisable samples.
# The original train and test sets are no longer needed.
tensor_train_X, tensor_train_y = get_correct_examples(model,
dataset_train,
device=device,
return_tensor=True)
dataset_train = TensorDataset(tensor_train_X, tensor_train_y)
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=True)
_, acc_perfect = validate(model, loader_train, loss, device)
print('Accuracy on {} filtered train set: {:.4f}%'.format(
len(dataset_train), acc_perfect * 100))
tensor_test_X, tensor_test_y = get_correct_examples(model,
dataset_test,
device=device,
return_tensor=True)
dataset_test = TensorDataset(tensor_test_X, tensor_test_y)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
_, acc_perfect = validate(model, loader_test, loss, device)
print('Accuracy on {} filtered test set: {:.4f}%'.format(
len(dataset_test), acc_perfect * 100))
# Load pre-trained adversarial examples
path_benign = os.path.join(args.output_path, args.adv + '_x.npy')
path_adv = os.path.join(args.output_path, args.adv + '_adv.npy')
path_y = os.path.join(args.output_path, args.adv + '_y.npy')
X_benign = np.load(path_benign)
adv = np.load(path_adv)
y_true = np.load(path_y)
dataset = TensorDataset(torch.from_numpy(X_benign),
torch.from_numpy(y_true))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
_, acc = validate(model, loader, loss, device)
print('Accuracy on {} benign samples: {:.4f}%'.format(
len(dataset), acc * 100))
dataset = TensorDataset(torch.from_numpy(adv), torch.from_numpy(y_true))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
_, acc = validate(model, loader, loss, device)
print('Accuracy on {} adversarial examples: {:.4f}%'.format(
len(dataset), acc * 100))
# Do NOT shuffle the indices, so different defences can use the same test set.
dataset = TensorDataset(torch.from_numpy(adv))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
pred_adv = predict(model, loader, device).cpu().detach().numpy()
# Find the thresholds using the 2nd half
n = len(X_benign) // 2
# Merge benign samples and adversarial examples into one set.
# This labels indicate a sample is an adversarial example or not.
X_val, labels_val = merge_and_generate_labels(adv[n:],
X_benign[n:],
flatten=False)
# The predictions for benign samples are exactly same as the true labels.
pred_val = np.concatenate((pred_adv[n:], y_true[n:]))
X_train = tensor_train_X.cpu().detach().numpy()
y_train = tensor_train_y.cpu().detach().numpy()
# Train defence
time_start = time.time()
detector = RegionBasedClassifier(model=model,
r=0.2,
sample_size=1000,
n_classes=n_classes,
x_min=0.0,
x_max=1.0,
batch_size=512,
r0=0.0,
step_size=0.02,
stop_value=0.4,
device=device)
r_best = detector.search_thresholds(X_val, pred_val, labels_val, verbose=0)
time_elapsed = time.time() - time_start
print('Total training time:',
str(datetime.timedelta(seconds=time_elapsed)))
param = {
"r": r_best,
"sample_size": 1000,
"batch_size": 512,
"r0": 0,
"step_size": 0.02,
"stop_value": 0.40
}
path_json = os.path.join(
'params', 'rc_param_{}_{}.json'.format(args.data, args.model))
with open(path_json, 'w') as f:
json.dump(param, f)
print('Save to:', path_json)
print()
def main():
with open('data.json') as data_json:
data_params = json.load(data_json)
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, required=True)
parser.add_argument('--data_path', type=str, default='data')
parser.add_argument('--output_path', type=str, default='results')
parser.add_argument('--pretrained', type=str, required=True)
parser.add_argument('--adv', type=str, required=True, help="Example: 'mnist_basic_apgd_0.3'")
parser.add_argument('--defence', type=str, required=True, choices=data_params['defences'])
parser.add_argument('--param', type=str, required=True)
parser.add_argument('--suffix', type=str)
parser.add_argument('--random_state', type=int, default=1234)
parser.add_argument('--save', type=int, default=1, choices=[0, 1])
args = parser.parse_args()
print(args)
set_seeds(args.random_state)
if not os.path.exists(args.output_path):
print('Output folder does not exist. Create:', args.output_path)
os.mkdir(args.output_path)
print('Dataset:', args.data)
print('Pretrained model:', args.pretrained)
print('Pretrained samples:', args.adv + '_adv.npy')
print('Defence:', args.defence)
with open(args.param) as param_json:
param = json.load(param_json)
param['n_classes'] = data_params['data'][args.data]['n_classes']
print('Param:', param)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: {}'.format(device))
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if args.data == 'mnist':
dataset_train = datasets.MNIST(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.MNIST(args.data_path, train=False, download=True, transform=transforms)
elif args.data == 'cifar10':
dataset_train = datasets.CIFAR10(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.CIFAR10(args.data_path, train=False, download=True, transform=transforms)
else:
data_path = os.path.join(args.data_path, data_params['data'][args.data]['file_name'])
print('Read file:', data_path)
X, y = load_csv(data_path)
X_train, X_test, y_train, y_test = train_test_split(
X, y,
test_size=data_params['data'][args.data]['n_test'],
random_state=args.random_state)
scaler = MinMaxScaler().fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
dataset_train = TensorDataset(torch.from_numpy(X_train).type(torch.float32), torch.from_numpy(y_train).type(torch.long))
dataset_test = TensorDataset(torch.from_numpy(X_test).type(torch.float32), torch.from_numpy(y_test).type(torch.long))
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=False)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
shape_train = get_shape(loader_train.dataset)
shape_test = get_shape(loader_test.dataset)
print('Train set:', shape_train)
print('Test set:', shape_test)
use_prob = True
print('Using softmax layer:', use_prob)
# Load model
if args.data == 'mnist':
model = BaseModel(use_prob=use_prob).to(device)
model_name = 'basic'
elif args.data == 'cifar10':
model_name = args.pretrained.split('_')[1]
if model_name == 'resnet':
model = Resnet(use_prob=use_prob).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=use_prob).to(device)
else:
raise ValueError('Unknown model: {}'.format(model_name))
else:
n_features = data_params['data'][args.data]['n_features']
n_classes = data_params['data'][args.data]['n_classes']
model = NumericModel(n_features, n_hidden=n_features * 4, n_classes=n_classes, use_prob=use_prob).to(device)
model_name = 'basic' + str(n_features * 4)
loss = nn.CrossEntropyLoss()
pretrained_path = os.path.join(args.output_path, args.pretrained)
model.load_state_dict(torch.load(pretrained_path, map_location=device))
_, acc_train = validate(model, loader_train, loss, device)
_, acc_test = validate(model, loader_test, loss, device)
print('Accuracy on train set: {:.4f}%'.format(acc_train * 100))
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
# Create a subset which only contains recognisable samples.
# The original train and test sets are no longer needed.
tensor_train_X, tensor_train_y = get_correct_examples(model, dataset_train, device=device, return_tensor=True)
dataset_train = TensorDataset(tensor_train_X, tensor_train_y)
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=True)
_, acc_perfect = validate(model, loader_train, loss, device)
print('Accuracy on {} filtered train set: {:.4f}%'.format(len(dataset_train), acc_perfect * 100))
tensor_test_X, tensor_test_y = get_correct_examples(model, dataset_test, device=device, return_tensor=True)
dataset_test = TensorDataset(tensor_test_X, tensor_test_y)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
_, acc_perfect = validate(model, loader_test, loss, device)
print('Accuracy on {} filtered test set: {:.4f}%'.format(len(dataset_test), acc_perfect * 100))
# Load pre-trained adversarial examples
path_benign = os.path.join(args.output_path, args.adv + '_x.npy')
path_adv = os.path.join(args.output_path, args.adv + '_adv.npy')
path_y = os.path.join(args.output_path, args.adv + '_y.npy')
X_benign = np.load(path_benign)
adv = np.load(path_adv)
y_true = np.load(path_y)
dataset = TensorDataset(torch.from_numpy(X_benign), torch.from_numpy(y_true))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
_, acc = validate(model, loader, loss, device)
print('Accuracy on {} benign samples: {:.4f}%'.format(len(dataset), acc * 100))
dataset = TensorDataset(torch.from_numpy(adv), torch.from_numpy(y_true))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
_, acc = validate(model, loader, loss, device)
print('Accuracy on {} adversarial examples: {:.4f}%'.format(len(dataset), acc * 100))
# Do NOT shuffle the indices, so different defences can use the same test set.
dataset = TensorDataset(torch.from_numpy(adv))
loader = DataLoader(dataset, batch_size=512, shuffle=False)
pred_adv = predict(model, loader, device).cpu().detach().numpy()
# Find the thresholds using the 2nd half
n = len(X_benign) // 2
# Merge benign samples and adversarial examples into one set.
# This labels indicate a sample is an adversarial example or not.
X_val, labels_val = merge_and_generate_labels(adv[n:], X_benign[n:], flatten=False)
# The predictions for benign samples are exactly same as the true labels.
pred_val = np.concatenate((pred_adv[n:], y_true[n:]))
X_train = tensor_train_X.cpu().detach().numpy()
y_train = tensor_train_y.cpu().detach().numpy()
# Train defence
time_start = time.time()
if args.defence == 'baard':
sequence = param['sequence']
stages = []
if sequence[0]:
stages.append(ApplicabilityStage(n_classes=param['n_classes'], quantile=param['q1']))
if sequence[1]:
stages.append(ReliabilityStage(n_classes=param['n_classes'], k=param['k_re'], quantile=param['q2']))
if sequence[2]:
stages.append(DecidabilityStage(n_classes=param['n_classes'], k=param['k_de'], quantile=param['q3']))
print('BAARD: # of stages:', len(stages))
detector = BAARDOperator(stages=stages)
# Run preprocessing
baard_train_path = os.path.join(args.output_path, '{}_{}_baard_train.pt'.format(args.data, model_name))
obj = torch.load(baard_train_path)
X_baard = obj['X_train']
y_train = obj['y_train']
# Fit the model with the filtered the train set.
detector.stages[0].fit(X_baard, y_train)
detector.stages[1].fit(X_train, y_train)
if len(detector.stages) == 3:
detector.stages[2].fit(X_train, y_train)
detector.search_thresholds(X_val, pred_val, labels_val)
path_baard = os.path.join(args.output_path, 'baard_{}_{}_param.pt'.format(args.data, model_name))
detector.save(path_baard)
elif args.defence == 'fs':
squeezers = []
if args.data == 'mnist':
squeezers.append(DepthSqueezer(x_min=0.0, x_max=1.0, bit_depth=1))
squeezers.append(MedianSqueezer(x_min=0.0, x_max=1.0, kernel_size=2))
elif args.data == 'cifar10':
squeezers.append(DepthSqueezer(x_min=0.0, x_max=1.0, bit_depth=4))
squeezers.append(MedianSqueezer(x_min=0.0, x_max=1.0, kernel_size=2))
squeezers.append(NLMeansColourSqueezer(x_min=0.0, x_max=1.0, h=2, templateWindowsSize=3, searchWindowSize=13))
else:
raise NotImplementedError
print('FS: # of squeezers:', len(squeezers))
detector = FeatureSqueezingTorch(
classifier=model,
lr=0.001,
momentum=0.9,
weight_decay=5e-4,
loss=loss,
batch_size=128,
x_min=0.0,
x_max=1.0,
squeezers=squeezers,
n_classes=param['n_classes'],
device=device)
path_fs = os.path.join(args.output_path, '{}_fs.pt'.format(args.pretrained.split('.')[0]))
detector.load(path_fs)
detector.search_thresholds(X_val, pred_val, labels_val)
elif args.defence == 'lid':
# This batch_size is not same as the mini batch size for the neural network.
before_softmax = args.data == 'cifar10'
detector = LidDetector(
model,
k=param['k'],
batch_size=param['batch_size'],
x_min=0.0,
x_max=1.0,
device=device,
before_softmax=before_softmax)
# LID uses different training set
X_train, y_train = detector.get_train_set(X_benign[n:], adv[n:], std_dominator=param['std_dominator'])
detector.fit(X_train, y_train, verbose=1)
elif args.defence == 'magnet':
magnet_detectors = []
# Different datasets require different autoencoders.
if args.data == 'mnist':
# autoencoder1 and autoencoder2
magnet_detectors.append(MagNetDetector(
encoder=Autoencoder1(n_channel=1),
classifier=model,
lr=param['lr'],
batch_size=param['batch_size'],
weight_decay=param['weight_decay'],
x_min=0.0,
x_max=1.0,
noise_strength=param['noise_strength'],
algorithm='error',
p=1,
device=device))
magnet_detectors.append(MagNetDetector(
encoder=Autoencoder2(n_channel=1),
classifier=model,
lr=param['lr'],
batch_size=param['batch_size'],
weight_decay=param['weight_decay'],
x_min=0.0,
x_max=1.0,
noise_strength=param['noise_strength'],
algorithm='error',
p=2,
device=device))
elif args.data == 'cifar10':
autoencoder = Autoencoder2(
n_channel=data_params['data'][args.data]['n_features'][0])
# There are 3 autoencoder based detectors, but they use the same architecture.
magnet_detectors.append(MagNetDetector(
encoder=autoencoder,
classifier=model,
lr=param['lr'],
batch_size=param['batch_size'],
weight_decay=param['weight_decay'],
x_min=0.0,
x_max=1.0,
noise_strength=param['noise_strength'],
algorithm='error',
p=2,
device=device))
magnet_detectors.append(MagNetDetector(
encoder=autoencoder,
classifier=model,
lr=param['lr'],
batch_size=param['batch_size'],
weight_decay=param['weight_decay'],
x_min=0.0,
x_max=1.0,
noise_strength=param['noise_strength'],
algorithm='prob',
temperature=10,
device=device))
magnet_detectors.append(MagNetDetector(
encoder=autoencoder,
classifier=model,
lr=param['lr'],
batch_size=param['batch_size'],
weight_decay=param['weight_decay'],
x_min=0.0,
x_max=1.0,
noise_strength=param['noise_strength'],
algorithm='prob',
temperature=40,
device=device))
else:
raise ValueError('Magnet requires autoencoder.')
for i, ae in enumerate(magnet_detectors, start=1):
ae_path = os.path.join(args.output_path, 'autoencoder_{}_{}_{}.pt'.format(args.data, model_name, i))
ae.load(ae_path)
tensor_X_test, _ = dataset2tensor(dataset_test)
X_test = tensor_X_test.cpu().detach().numpy()
print('Autoencoder {} MSE training set: {:.6f}, test set: {:.6f}'.format(i, ae.score(X_train), ae.score(X_test)))
print('Autoencoder {} threshold: {}'.format(i, ae.threshold))
reformer = MagNetAutoencoderReformer(
encoder=magnet_detectors[0].encoder,
batch_size=param['batch_size'],
device=device)
detector = MagNetOperator(
classifier=model,
detectors=magnet_detectors,
reformer=reformer,
batch_size=param['batch_size'],
device=device)
elif args.defence == 'rc':
detector = RegionBasedClassifier(
model=model,
r=param['r'],
sample_size=param['sample_size'],
n_classes=param['n_classes'],
x_min=0.0,
x_max=1.0,
batch_size=param['batch_size'],
r0=param['r0'],
step_size=param['step_size'],
stop_value=param['stop_value'],
device=device)
# Region-based classifier only uses benign samples to search threshold.
# The r value is already set to the optimal. We don't need to search it.
# detector.search_thresholds(X_val, pred_val, labels_val, verbose=0)
else:
raise ValueError('{} is not supported!'.format(args.defence))
time_elapsed = time.time() - time_start
print('Total training time:', str(datetime.timedelta(seconds=time_elapsed)))
# Test defence
time_start = time.time()
X_test, labels_test = merge_and_generate_labels(adv[:n], X_benign[:n], flatten=False)
pred_test = np.concatenate((pred_adv[:n], y_true[:n]))
y_test = np.concatenate((y_true[:n], y_true[:n]))
# Only MegNet uses reformer.
X_reformed = None
if args.defence == 'magnet':
X_reformed, res_test = detector.detect(X_test, pred_test)
y_pred = predict_numpy(model, X_reformed, device)
elif args.defence == 'rc':
y_pred = detector.detect(X_test, pred_test)
res_test = np.zeros_like(y_pred)
else:
res_test = detector.detect(X_test, pred_test)
y_pred = pred_test
acc = acc_on_adv(y_pred[:n], y_test[:n], res_test[:n])
if args.defence == 'rc':
fpr = np.mean(y_pred[n:] != y_test[n:])
else:
fpr = np.mean(res_test[n:])
print('Acc_on_adv:', acc)
print('FPR:', fpr)
time_elapsed = time.time() - time_start
print('Total test time:', str(datetime.timedelta(seconds=time_elapsed)))
# Save results
suffix = '_' + args.suffix if args.suffix is not None else ''
if args.save:
path_result = os.path.join(args.output_path, '{}_{}{}.pt'.format(args.adv, args.defence, suffix))
torch.save({
'X_val': X_val,
'y_val': np.concatenate((y_true[n:], y_true[n:])),
'labels_val': labels_val,
'X_test': X_test,
'y_test': y_test,
'labels_test': labels_test,
'res_test': y_pred if args.defence == 'rc' else res_test,
'X_reformed': X_reformed,
'param': param}, path_result)
print('Saved to:', path_result)
else:
print('No file is save!')
print()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, required=True, choices=['mnist', 'cifar10'])
parser.add_argument('--data_path', type=str, default='data')
parser.add_argument('--output_path', type=str, default='results')
parser.add_argument('--pretrained', type=str, required=True)
parser.add_argument('--param', type=str, required=True)
parser.add_argument('--random_state', type=int, default=1234)
args = parser.parse_args()
print(args)
set_seeds(args.random_state)
if not os.path.exists(args.output_path):
print('Output folder does not exist. Create:', args.output_path)
os.mkdir(args.output_path)
print('Dataset:', args.data)
print('Pretrained model:', args.pretrained)
with open(args.param) as param_json:
param = json.load(param_json)
param['n_classes'] = 10
print('Param:', param)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: {}'.format(device))
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if args.data == 'mnist':
dataset_train = datasets.MNIST(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.MNIST(args.data_path, train=False, download=True, transform=transforms)
elif args.data == 'cifar10':
dataset_train = datasets.CIFAR10(args.data_path, train=True, download=True, transform=transforms)
dataset_test = datasets.CIFAR10(args.data_path, train=False, download=True, transform=transforms)
else:
raise ValueError('{} is not supported.'.format(args.data))
# Note: Train set alway shuffle!
loader_train = DataLoader(dataset_train, batch_size=512, shuffle=True)
loader_test = DataLoader(dataset_test, batch_size=512, shuffle=False)
shape_train = get_shape(loader_train.dataset)
shape_test = get_shape(loader_test.dataset)
print('Train set:', shape_train)
print('Test set:', shape_test)
use_prob = True
print('Using softmax layer:', use_prob)
# Load model
if args.data == 'mnist':
model = BaseModel(use_prob=use_prob).to(device)
model_name = 'basic'
else: # args.data == 'cifar10':
model_name = args.pretrained.split('_')[1]
if model_name == 'resnet':
model = Resnet(use_prob=use_prob).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=use_prob).to(device)
else:
raise ValueError('Unknown model: {}'.format(model_name))
loss = nn.CrossEntropyLoss()
pretrained_path = os.path.join(args.output_path, args.pretrained)
model.load_state_dict(torch.load(pretrained_path))
_, acc_train = validate(model, loader_train, loss, device)
_, acc_test = validate(model, loader_test, loss, device)
print('Accuracy on train set: {:.4f}%'.format(acc_train * 100))
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
tensor_train_X, tensor_train_y = dataset2tensor(dataset_train)
X_train = tensor_train_X.cpu().detach().numpy()
y_train = tensor_train_y.cpu().detach().numpy()
# Train defence
squeezers = []
if args.data == 'mnist':
squeezers.append(DepthSqueezer(x_min=0.0, x_max=1.0, bit_depth=1))
squeezers.append(MedianSqueezer(x_min=0.0, x_max=1.0, kernel_size=2))
else:
# CIFAR10
squeezers.append(DepthSqueezer(x_min=0.0, x_max=1.0, bit_depth=4))
squeezers.append(MedianSqueezer(x_min=0.0, x_max=1.0, kernel_size=2))
squeezers.append(NLMeansColourSqueezer(x_min=0.0, x_max=1.0, h=2, templateWindowsSize=3, searchWindowSize=13))
print('FS: # of squeezers:', len(squeezers))
detector = FeatureSqueezingTorch(
classifier=model,
lr=0.001,
momentum=0.9,
weight_decay=5e-4,
loss=loss,
batch_size=128,
x_min=0.0,
x_max=1.0,
squeezers=squeezers,
n_classes=param['n_classes'],
device=device)
detector.fit(X_train, y_train, epochs=param['epochs'], verbose=1)
path_fs = os.path.join(args.output_path, '{}_fs.pt'.format(args.pretrained.split('.')[0]))
detector.save(path_fs)
print('Saved fs to:', path_fs)
print()
