def test_generate(art_warning, fix_get_mnist_subset,
image_dl_estimator_for_attack):
try:
classifier = image_dl_estimator_for_attack(
AutoProjectedGradientDescent)
attack = AutoProjectedGradientDescent(
estimator=classifier,
norm=np.inf,
eps=0.3,
eps_step=0.1,
max_iter=5,
targeted=False,
nb_random_init=1,
batch_size=32,
loss_type="cross_entropy",
verbose=False,
)
(x_train_mnist, y_train_mnist, x_test_mnist,
y_test_mnist) = fix_get_mnist_subset
x_train_mnist_adv = attack.generate(x=x_train_mnist, y=y_train_mnist)
assert np.mean(np.abs(x_train_mnist_adv -
x_train_mnist)) == pytest.approx(0.0329,
abs=0.005)
assert np.max(np.abs(x_train_mnist_adv -
x_train_mnist)) == pytest.approx(0.3, abs=0.01)
except ARTTestException as e:
art_warning(e)
def test_generate(is_tf_version_2, fix_get_mnist_subset, get_image_classifier_list_for_attack):
if is_tf_version_2:
classifier_list = get_image_classifier_list_for_attack(AutoProjectedGradientDescent)
if classifier_list is None:
logging.warning("Couldn't perform this test because no classifier is defined")
return
for classifier in classifier_list:
attack = AutoProjectedGradientDescent(
estimator=classifier,
norm=np.inf,
eps=0.3,
eps_step=0.1,
max_iter=5,
targeted=False,
nb_random_init=1,
batch_size=32,
loss_type="cross_entropy",
)
(x_train_mnist, y_train_mnist, x_test_mnist, y_test_mnist) = fix_get_mnist_subset
x_train_mnist_adv = attack.generate(x=x_train_mnist, y=y_train_mnist)
assert np.mean(np.abs(x_train_mnist_adv - x_train_mnist)) == pytest.approx(0.0329, abs=0.005)
assert np.max(np.abs(x_train_mnist_adv - x_train_mnist)) == pytest.approx(0.3, abs=0.01)
def test_generate(art_warning, fix_get_mnist_subset,
image_dl_estimator_for_attack, framework, loss_type, norm):
print("test_generate")
try:
classifier = image_dl_estimator_for_attack(
AutoProjectedGradientDescent, from_logits=True)
print("framework", framework)
if framework in ["tensorflow1", "tensorflow2v1"
] and loss_type == "difference_logits_ratio":
with pytest.raises(ValueError):
_ = AutoProjectedGradientDescent(
estimator=classifier,
norm=norm,
eps=0.3,
eps_step=0.1,
max_iter=5,
targeted=False,
nb_random_init=1,
batch_size=32,
loss_type=loss_type,
verbose=False,
)
else:
attack = AutoProjectedGradientDescent(
estimator=classifier,
norm=norm,
eps=0.3,
eps_step=0.1,
max_iter=5,
targeted=False,
nb_random_init=1,
batch_size=32,
loss_type=loss_type,
verbose=False,
)
(x_train_mnist, y_train_mnist, x_test_mnist,
y_test_mnist) = fix_get_mnist_subset
x_train_mnist_adv = attack.generate(x=x_train_mnist,
y=y_train_mnist)
assert np.max(np.abs(x_train_mnist_adv - x_train_mnist)) > 0.0
except ARTTestException as e:
art_warning(e)
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 train_adv(data='mnist',
model_name='basic',
n_samples=2000,
eps=2.,
path_output='results',
path_data='data',
is_test=False,
batch_size=128,
device='cpu'):
# Prepare data
transforms = tv.transforms.Compose([tv.transforms.ToTensor()])
if data == 'mnist':
dataset_test = datasets.MNIST(path_data,
train=False,
download=True,
transform=transforms)
elif data == 'cifar10':
dataset_test = datasets.CIFAR10(path_data,
train=False,
download=True,
transform=transforms)
else:
raise NotImplementedError
loader_test = DataLoader(dataset_test,
batch_size=batch_size,
shuffle=False)
# Load model
if data == 'mnist':
model = BaseModel(use_prob=False).to(device)
n_features = (1, 28, 28)
pretrained = 'mnist_200.pt'
elif data == 'cifar10':
n_features = (3, 32, 32)
if model_name == 'resnet':
model = Resnet(use_prob=False).to(device)
pretrained = 'cifar10_resnet_200.pt'
elif model_name == 'vgg':
model = Vgg(use_prob=False).to(device)
pretrained = 'cifar10_vgg_200.pt'
else:
raise NotImplementedError
else:
raise NotImplementedError
pretrained_path = os.path.join(path_output, pretrained)
model.load_state_dict(torch.load(pretrained_path, map_location=device))
optimizer = optim.SGD(model.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=5e-4)
loss = nn.CrossEntropyLoss()
_, acc_test = validate(model, loader_test, loss, device)
print('Accuracy on test set: {:.4f}%'.format(acc_test * 100))
tensor_test_X, tensor_test_y = get_correct_examples(model,
dataset_test,
device=device,
return_tensor=True)
# Get samples from the tail
if not is_test:
# This is for training the surrogate model
tensor_test_X = tensor_test_X[-n_samples:]
tensor_test_y = tensor_test_y[-n_samples:]
else:
# This is for testing the surrogate model
tensor_test_X = tensor_test_X[-n_samples - 2000:-2000]
tensor_test_y = tensor_test_y[-n_samples - 2000:-2000]
dataset_test = TensorDataset(tensor_test_X, tensor_test_y)
loader_test = DataLoader(dataset_test,
batch_size=batch_size,
shuffle=False)
_, acc_perfect = validate(model, loader_test, loss, device)
print('Accuracy on {} filtered test set: {:.4f}%'.format(
len(dataset_test), acc_perfect * 100))
classifier = PyTorchClassifier(model=model,
loss=loss,
input_shape=n_features,
optimizer=optimizer,
nb_classes=10,
clip_values=(0.0, 1.0),
device_type='gpu')
attack = AutoProjectedGradientDescent(estimator=classifier,
eps=eps,
eps_step=0.1,
max_iter=1000,
batch_size=batch_size,
targeted=False)
X_benign = tensor_test_X.cpu().detach().numpy()
y_true = tensor_test_y.cpu().detach().numpy()
adv = attack.generate(x=X_benign)
pred_adv = np.argmax(classifier.predict(adv), axis=1)
acc_adv = np.mean(pred_adv == y_true)
print("Accuracy on adversarial examples: {:.4f}%".format(acc_adv * 100))
if not is_test:
output_file = '{}_{}_baard_surro_train_eps{}_size{}.pt'.format(
data, model_name, eps, n_samples)
else:
output_file = '{}_{}_baard_surro_test_eps{}_size{}.pt'.format(
data, model_name, eps, n_samples)
file_path = os.path.join(path_output, output_file)
output = {'X': X_benign, 'adv': adv, 'y': y_true}
torch.save(output, file_path)
print('Save to:', file_path)
def run_attack_untargeted(file_model, X, y, att_name, eps, device):
path = file_model.split('/')[0]
file_str = file_model.split('/')[-1]
name_arr = file_str.split('_')
data = name_arr[0]
model_name = name_arr[1]
file_data = os.path.join(
path, '{}_{}_{}_{}.pt'.format(data, model_name, att_name,
round(eps * 1000)))
if os.path.exists(file_data):
print('Found existing file:', file_data)
obj = torch.load(file_data)
return obj['adv'], obj['X'], obj['y']
if data == 'mnist':
n_features = (1, 28, 28)
n_classes = 10
model = BaseModel(use_prob=False).to(device)
elif data == 'cifar10':
n_features = (3, 32, 32)
n_classes = 10
if model_name == 'resnet':
model = Resnet(use_prob=False).to(device)
elif model_name == 'vgg':
model = Vgg(use_prob=False).to(device)
else:
raise NotImplementedError
else:
raise NotImplementedError
model.load_state_dict(torch.load(file_model, map_location=device))
loss = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=5e-4)
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 att_name == 'apgd':
eps_step = eps / 4. if eps <= 0.2 else 0.1
attack = AutoProjectedGradientDescent(estimator=classifier,
eps=eps,
eps_step=eps_step,
max_iter=1000,
batch_size=BATCH_SIZE,
targeted=False)
elif att_name == 'apgd2':
attack = AutoProjectedGradientDescent(estimator=classifier,
norm=2,
eps=eps,
eps_step=0.1,
max_iter=1000,
batch_size=BATCH_SIZE,
targeted=False)
elif att_name == 'cw2':
# Do not increase the batch_size
attack = CarliniWagnerAttackL2(model=model,
n_classes=n_classes,
confidence=eps,
verbose=True,
check_prob=False,
batch_size=32,
targeted=False)
elif att_name == 'deepfool':
# Do not adjust Epsilon
attack = DeepFool(classifier=classifier, batch_size=BATCH_SIZE)
elif att_name == 'fgsm':
attack = FastGradientMethod(estimator=classifier,
eps=eps,
batch_size=BATCH_SIZE)
elif att_name == 'line':
attack = LineAttack(color=1, thickness=2)
else:
raise NotImplementedError
time_start = time.time()
adv = attack.generate(x=X)
time_elapsed = time.time() - time_start
print('Total run time:', str(datetime.timedelta(seconds=time_elapsed)))
obj = {'X': X, 'y': y, 'adv': adv}
torch.save(obj, file_data)
print('Save data to:', file_data)
return adv, X, y
def main():
Ad = np.load(AD_MAT_FILE) # Load adjacency matrix
NUM_TEST = 50
NUM_GRAPH = 200
array_std, array_mean_values, array_overlap_ratio = load_raw_result_csv(
RAW_RESULT_FILE)
NUM_CLASS = array_mean_values.shape[1]
print(array_mean_values.shape)
with open("result_PGD.csv", "w", newline='') as csvfile:
writer = csv.writer(csvfile)
writer.writerow([
"overlap ratio", "acc_test_L", "acc_test_WL", "acc_adv_with_Lip",
"acc_adv_without_Lip"
])
for i in range(NUM_TEST):
x_test, y_test = reconstruct_test_data(array_std[i],
array_mean_values[i], Ad,
NUM_GRAPH)
model_with_Lip_constr = tf.keras.models.load_model(
"saved_model/fit{}_model_with_Lip_constr.h5".format(i))
print(model_with_Lip_constr.summary())
model_without_Lip_constr = tf.keras.models.load_model(
"saved_model/fit{}_model_without_Lip_constr.h5".format(i))
print(
"Evaluation of model WITH Lipschitz constant constraint on TEST data"
)
loss_test_L, acc_test_L = model_with_Lip_constr.evaluate(
x_test, y_test, batch_size=x_test.shape[0], verbose=0)
print("Loss: {:.4f}, accuracy: {:.4f}".format(loss_test_L, acc_test_L))
print(
"Evaluation of model WITHOUT Lipschitz constant constraint on TEST data"
)
loss_test_WL, acc_test_WL = model_without_Lip_constr.evaluate(
x_test, y_test, batch_size=x_test.shape[0], verbose=0)
print("Loss: {:.4f}, accuracy: {:.4f}".format(loss_test_WL,
acc_test_WL))
# Reshape model output
reshape_with_Lip = Reshape(
(x_test.shape[1] * NUM_CLASS, ),
name="added_reshape_layer_L")(model_with_Lip_constr.output)
new_model_with_Lip = Model(inputs=model_with_Lip_constr.input,
outputs=reshape_with_Lip)
reshape_without_Lip = Reshape(
(x_test.shape[1] * NUM_CLASS, ),
name="added_reshape_layer_WL")(model_without_Lip_constr.output)
new_model_without_Lip = Model(inputs=model_without_Lip_constr.input,
outputs=reshape_without_Lip)
new_model_with_Lip.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
new_model_without_Lip.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
min_value = np.min(array_mean_values[i]) - 100 * array_std[i]
max_value = np.max(array_mean_values[i]) + 100 * array_std[i]
classifier_with_Lip = KerasClassifier(model=new_model_with_Lip,
clip_values=(min_value,
max_value),
use_logits=False)
classifier_without_Lip = KerasClassifier(model=new_model_without_Lip,
clip_values=(min_value,
max_value),
use_logits=False)
attack1 = AutoProjectedGradientDescent(estimator=classifier_with_Lip,
norm="inf",
eps=0.6,
eps_step=1,
batch_size=200,
nb_random_init=5,
verbose=True,
targeted=False)
attack2 = AutoProjectedGradientDescent(
estimator=classifier_without_Lip,
norm="inf",
eps=0.6,
eps_step=1,
batch_size=200,
nb_random_init=5,
verbose=True,
targeted=False)
x_test_adv1 = attack1.generate(x=x_test,
mask=np.ones((1, x_test.shape[1],
x_test.shape[2])))
x_test_adv2 = attack2.generate(x=x_test,
mask=np.ones((1, x_test.shape[1],
x_test.shape[2])))
y_predict_adv_with_Lip = classifier_with_Lip.predict(x_test_adv1)
y_predict_adv_without_Lip = classifier_without_Lip.predict(x_test_adv2)
y_predict_adv_with_Lip = y_predict_adv_with_Lip.reshape((y_test.shape))
y_predict_adv_without_Lip = y_predict_adv_without_Lip.reshape(
(y_test.shape))
acc_adv_with_Lip = np.sum(
np.argmax(y_predict_adv_with_Lip, axis=2) == np.argmax(
y_test, axis=2)) / (y_test.shape[0] * y_test.shape[1])
print(
"Accuracy on adversarial test examples with Lipschitz constraint: {:.2f}%"
.format(acc_adv_with_Lip * 100))
acc_adv_without_Lip = np.sum(
np.argmax(y_predict_adv_without_Lip, axis=2) == np.argmax(
y_test, axis=2)) / (y_test.shape[0] * y_test.shape[1])
print(
"Accuracy on adversarial test examples without Lipschitz constraint: {:.2f}%"
.format(acc_adv_without_Lip * 100))
with open("result_PGD.csv", "a", newline='') as csvfile:
writer = csv.writer(csvfile)
writer.writerow([
array_overlap_ratio[i], acc_test_L, acc_test_WL,
acc_adv_with_Lip, acc_adv_without_Lip
])