def mnist_train(epoch_size, batch_size, lr, momentum):
mnist_path = "./MNIST_unzip/"
ds = generate_mnist_dataset(os.path.join(mnist_path, "train"),
batch_size=batch_size,
repeat_size=1)
network = LeNet5()
net_loss = nn.SoftmaxCrossEntropyWithLogits(is_grad=False,
sparse=True,
reduction="mean")
net_opt = nn.Momentum(network.trainable_params(), lr, momentum)
config_ck = CheckpointConfig(save_checkpoint_steps=1875,
keep_checkpoint_max=10)
ckpoint_cb = ModelCheckpoint(prefix="checkpoint_lenet",
directory="./trained_ckpt_file/",
config=config_ck)
model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
LOGGER.info(TAG, "============== Starting Training ==============")
model.train(epoch_size,
ds,
callbacks=[ckpoint_cb, LossMonitor()],
dataset_sink_mode=False)
LOGGER.info(TAG, "============== Starting Testing ==============")
ckpt_file_name = "trained_ckpt_file/checkpoint_lenet-10_1875.ckpt"
param_dict = load_checkpoint(ckpt_file_name)
load_param_into_net(network, param_dict)
ds_eval = generate_mnist_dataset(os.path.join(mnist_path, "test"),
batch_size=batch_size)
acc = model.eval(ds_eval, dataset_sink_mode=False)
LOGGER.info(TAG, "============== Accuracy: %s ==============", acc)
def test_lenet_mnist_fuzzing():
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
model = Model(net)
# get training data
data_list = "./MNIST_unzip/train"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size, sparse=True)
train_images = []
for data in ds.create_tuple_iterator():
images = data[0].astype(np.float32)
train_images.append(images)
train_images = np.concatenate(train_images, axis=0)
# initialize fuzz test with training dataset
model_coverage_test = ModelCoverageMetrics(model, 1000, 10, train_images)
# fuzz test with original test data
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size, sparse=True)
test_images = []
test_labels = []
for data in ds.create_tuple_iterator():
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
test_images = np.concatenate(test_images, axis=0)
test_labels = np.concatenate(test_labels, axis=0)
initial_seeds = []
# make initial seeds
for img, label in zip(test_images, test_labels):
initial_seeds.append([img, label])
initial_seeds = initial_seeds[:100]
model_coverage_test.test_adequacy_coverage_calculate(
np.array(test_images[:100]).astype(np.float32))
LOGGER.info(TAG, 'KMNC of this test is : %s',
model_coverage_test.get_kmnc())
model_fuzz_test = Fuzzing(initial_seeds, model, train_images, 20)
failed_tests = model_fuzz_test.fuzzing()
if failed_tests:
model_coverage_test.test_adequacy_coverage_calculate(
np.array(failed_tests).astype(np.float32))
LOGGER.info(TAG, 'KMNC of this test is : %s',
model_coverage_test.get_kmnc())
else:
LOGGER.info(TAG, 'Fuzzing test identifies none failed test')
def test_lenet_mnist_coverage():
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
model = Model(net)
# get training data
data_list = "./MNIST_unzip/train"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size, sparse=True)
train_images = []
for data in ds.create_tuple_iterator():
images = data[0].astype(np.float32)
train_images.append(images)
train_images = np.concatenate(train_images, axis=0)
# initialize fuzz test with training dataset
model_fuzz_test = ModelCoverageMetrics(model, 10000, 10, train_images)
# fuzz test with original test data
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size, sparse=True)
test_images = []
test_labels = []
for data in ds.create_tuple_iterator():
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
test_images = np.concatenate(test_images, axis=0)
test_labels = np.concatenate(test_labels, axis=0)
model_fuzz_test.test_adequacy_coverage_calculate(test_images)
LOGGER.info(TAG, 'KMNC of this test is : %s', model_fuzz_test.get_kmnc())
LOGGER.info(TAG, 'NBC of this test is : %s', model_fuzz_test.get_nbc())
LOGGER.info(TAG, 'SNAC of this test is : %s', model_fuzz_test.get_snac())
# generate adv_data
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
attack = FastGradientSignMethod(net, eps=0.3, loss_fn=loss)
adv_data = attack.batch_generate(test_images, test_labels, batch_size=32)
model_fuzz_test.test_adequacy_coverage_calculate(adv_data,
bias_coefficient=0.5)
LOGGER.info(TAG, 'KMNC of this test is : %s', model_fuzz_test.get_kmnc())
LOGGER.info(TAG, 'NBC of this test is : %s', model_fuzz_test.get_nbc())
LOGGER.info(TAG, 'SNAC of this test is : %s', model_fuzz_test.get_snac())
def mnist_train(epoch_size, batch_size, lr, momentum):
context.set_context(mode=context.GRAPH_MODE,
device_target="Ascend",
enable_mem_reuse=False)
lr = lr
momentum = momentum
epoch_size = epoch_size
mnist_path = "./MNIST_unzip/"
ds = generate_mnist_dataset(os.path.join(mnist_path, "train"),
batch_size=batch_size,
repeat_size=1)
network = LeNet5()
network.set_train()
net_loss = CrossEntropyLoss()
net_opt = nn.Momentum(network.trainable_params(), lr, momentum)
config_ck = CheckpointConfig(save_checkpoint_steps=1875,
keep_checkpoint_max=10)
ckpoint_cb = ModelCheckpoint(prefix="checkpoint_lenet",
directory='./trained_ckpt_file/',
config=config_ck)
model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
LOGGER.info(TAG, "============== Starting Training ==============")
model.train(epoch_size,
ds,
callbacks=[ckpoint_cb, LossMonitor()],
dataset_sink_mode=False) # train
LOGGER.info(TAG, "============== Starting Testing ==============")
param_dict = load_checkpoint(
"trained_ckpt_file/checkpoint_lenet-10_1875.ckpt")
load_param_into_net(network, param_dict)
ds_eval = generate_mnist_dataset(os.path.join(mnist_path, "test"),
batch_size=batch_size)
acc = model.eval(ds_eval)
LOGGER.info(TAG, "============== Accuracy: %s ==============", acc)
def test_hsja_mnist_attack():
"""
hsja-Attack test
"""
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
net.set_train(False)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size=batch_size)
# prediction accuracy before attack
model = ModelToBeAttacked(net)
batch_num = 5 # the number of batches of attacking samples
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.concatenate(test_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %s", accuracy)
test_images = np.concatenate(test_images)
# attacking
norm = 'l2'
search = 'grid_search'
target = False
attack = HopSkipJumpAttack(model, constraint=norm, stepsize_search=search)
if target:
target_labels = random_target_labels(true_labels)
target_images = create_target_images(test_images, predict_labels,
target_labels)
attack.set_target_images(target_images)
success_list, adv_data, query_list = attack.generate(
test_images, target_labels)
else:
success_list, adv_data, query_list = attack.generate(test_images, None)
adv_datas = []
gts = []
for success, adv, gt in zip(success_list, adv_data, true_labels):
if success:
adv_datas.append(adv)
gts.append(gt)
if len(gts) > 0:
adv_datas = np.concatenate(np.asarray(adv_datas), axis=0)
gts = np.asarray(gts)
pred_logits_adv = model.predict(adv_datas)
pred_lables_adv = np.argmax(pred_logits_adv, axis=1)
accuracy_adv = np.mean(np.equal(pred_lables_adv, gts))
LOGGER.info(TAG, 'mis-classification rate of adversaries is : %s',
accuracy_adv)
def test_pointwise_attack_on_mnist():
"""
Salt-and-Pepper-Attack test
"""
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size=batch_size)
# prediction accuracy before attack
model = ModelToBeAttacked(net)
batch_num = 3 # the number of batches of attacking samples
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.concatenate(test_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %g", accuracy)
# attacking
is_target = False
attack = PointWiseAttack(model=model, is_targeted=is_target)
if is_target:
targeted_labels = np.random.randint(0, 10, size=len(true_labels))
for i in range(len(true_labels)):
if targeted_labels[i] == true_labels[i]:
targeted_labels[i] = (targeted_labels[i] + 1) % 10
else:
targeted_labels = true_labels
success_list, adv_data, query_list = attack.generate(
np.concatenate(test_images), targeted_labels)
success_list = np.arange(success_list.shape[0])[success_list]
LOGGER.info(TAG, 'success_list: %s', success_list)
LOGGER.info(TAG, 'average of query times is : %s', np.mean(query_list))
adv_preds = []
for ite_data in adv_data:
pred_logits_adv = model.predict(ite_data)
# rescale predict confidences into (0, 1).
pred_logits_adv = softmax(pred_logits_adv, axis=1)
adv_preds.extend(pred_logits_adv)
accuracy_adv = np.mean(np.equal(np.max(adv_preds, axis=1), true_labels))
LOGGER.info(TAG, "prediction accuracy after attacking is : %g",
accuracy_adv)
test_labels_onehot = np.eye(10)[true_labels]
attack_evaluate = AttackEvaluate(np.concatenate(test_images),
test_labels_onehot,
adv_data,
adv_preds,
targeted=is_target,
target_label=targeted_labels)
LOGGER.info(TAG, 'mis-classification rate of adversaries is : %s',
attack_evaluate.mis_classification_rate())
LOGGER.info(TAG, 'The average confidence of adversarial class is : %s',
attack_evaluate.avg_conf_adv_class())
LOGGER.info(TAG, 'The average confidence of true class is : %s',
attack_evaluate.avg_conf_true_class())
LOGGER.info(
TAG, 'The average distance (l0, l2, linf) between original '
'samples and adversarial samples are: %s',
attack_evaluate.avg_lp_distance())
def test_lbfgs_attack():
"""
LBFGS-Attack test
"""
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size=batch_size, sparse=False)
# prediction accuracy before attack
model = Model(net)
batch_num = 3 # the number of batches of attacking samples
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(Tensor(images)).asnumpy(),
axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.argmax(np.concatenate(test_labels), axis=1)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %s", accuracy)
# attacking
is_targeted = True
if is_targeted:
targeted_labels = np.random.randint(
0, 10, size=len(true_labels)).astype(np.int32)
for i in range(len(true_labels)):
if targeted_labels[i] == true_labels[i]:
targeted_labels[i] = (targeted_labels[i] + 1) % 10
else:
targeted_labels = true_labels.astype(np.int32)
targeted_labels = np.eye(10)[targeted_labels].astype(np.float32)
attack = LBFGS(net, is_targeted=is_targeted)
start_time = time.clock()
adv_data = attack.batch_generate(np.concatenate(test_images),
targeted_labels,
batch_size=batch_size)
stop_time = time.clock()
pred_logits_adv = model.predict(Tensor(adv_data)).asnumpy()
# rescale predict confidences into (0, 1).
pred_logits_adv = softmax(pred_logits_adv, axis=1)
pred_labels_adv = np.argmax(pred_logits_adv, axis=1)
accuracy_adv = np.mean(np.equal(pred_labels_adv, true_labels))
LOGGER.info(TAG, "prediction accuracy after attacking is : %s",
accuracy_adv)
attack_evaluate = AttackEvaluate(np.concatenate(test_images).transpose(
0, 2, 3, 1),
np.concatenate(test_labels),
adv_data.transpose(0, 2, 3, 1),
pred_logits_adv,
targeted=is_targeted,
target_label=np.argmax(targeted_labels,
axis=1))
LOGGER.info(TAG, 'mis-classification rate of adversaries is : %s',
attack_evaluate.mis_classification_rate())
LOGGER.info(TAG, 'The average confidence of adversarial class is : %s',
attack_evaluate.avg_conf_adv_class())
LOGGER.info(TAG, 'The average confidence of true class is : %s',
attack_evaluate.avg_conf_true_class())
LOGGER.info(
TAG, 'The average distance (l0, l2, linf) between original '
'samples and adversarial samples are: %s',
attack_evaluate.avg_lp_distance())
LOGGER.info(
TAG, 'The average structural similarity between original '
'samples and adversarial samples are: %s', attack_evaluate.avg_ssim())
LOGGER.info(TAG, 'The average costing time is %s',
(stop_time - start_time) / (batch_num * batch_size))
def test_genetic_attack_on_mnist():
"""
Genetic-Attack test
"""
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size=batch_size)
# prediction accuracy before attack
model = ModelToBeAttacked(net)
batch_num = 3 # the number of batches of attacking samples
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.concatenate(test_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %g", accuracy)
# attacking
attack = GeneticAttack(model=model,
pop_size=6,
mutation_rate=0.05,
per_bounds=0.1,
step_size=0.25,
temp=0.1,
sparse=True)
targeted_labels = np.random.randint(0, 10, size=len(true_labels))
for i, true_l in enumerate(true_labels):
if targeted_labels[i] == true_l:
targeted_labels[i] = (targeted_labels[i] + 1) % 10
start_time = time.clock()
success_list, adv_data, query_list = attack.generate(
np.concatenate(test_images), targeted_labels)
stop_time = time.clock()
LOGGER.info(TAG, 'success_list: %s', success_list)
LOGGER.info(TAG, 'average of query times is : %s', np.mean(query_list))
pred_logits_adv = model.predict(adv_data)
# rescale predict confidences into (0, 1).
pred_logits_adv = softmax(pred_logits_adv, axis=1)
pred_lables_adv = np.argmax(pred_logits_adv, axis=1)
accuracy_adv = np.mean(np.equal(pred_lables_adv, true_labels))
LOGGER.info(TAG, "prediction accuracy after attacking is : %g",
accuracy_adv)
test_labels_onehot = np.eye(10)[true_labels]
attack_evaluate = AttackEvaluate(np.concatenate(test_images),
test_labels_onehot,
adv_data,
pred_logits_adv,
targeted=True,
target_label=targeted_labels)
LOGGER.info(TAG, 'mis-classification rate of adversaries is : %s',
attack_evaluate.mis_classification_rate())
LOGGER.info(TAG, 'The average confidence of adversarial class is : %s',
attack_evaluate.avg_conf_adv_class())
LOGGER.info(TAG, 'The average confidence of true class is : %s',
attack_evaluate.avg_conf_true_class())
LOGGER.info(
TAG, 'The average distance (l0, l2, linf) between original '
'samples and adversarial samples are: %s',
attack_evaluate.avg_lp_distance())
LOGGER.info(
TAG, 'The average structural similarity between original '
'samples and adversarial samples are: %s', attack_evaluate.avg_ssim())
LOGGER.info(TAG, 'The average costing time is %s',
(stop_time - start_time) / (batch_num * batch_size))
def test_black_defense():
# load trained network
current_dir = os.path.dirname(os.path.abspath(__file__))
ckpt_name = os.path.abspath(
os.path.join(current_dir,
'./trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'))
# ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
wb_net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(wb_net, load_dict)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds_test = generate_mnist_dataset(data_list,
batch_size=batch_size,
sparse=False)
inputs = []
labels = []
for data in ds_test.create_tuple_iterator():
inputs.append(data[0].astype(np.float32))
labels.append(data[1])
inputs = np.concatenate(inputs).astype(np.float32)
labels = np.concatenate(labels).astype(np.float32)
labels_sparse = np.argmax(labels, axis=1)
target_label = np.random.randint(0, 10, size=labels_sparse.shape[0])
for idx in range(labels_sparse.shape[0]):
while target_label[idx] == labels_sparse[idx]:
target_label[idx] = np.random.randint(0, 10)
target_label = np.eye(10)[target_label].astype(np.float32)
attacked_size = 50
benign_size = 500
attacked_sample = inputs[:attacked_size]
attacked_true_label = labels[:attacked_size]
benign_sample = inputs[attacked_size:attacked_size + benign_size]
wb_model = ModelToBeAttacked(wb_net)
# gen white-box adversarial examples of test data
wb_attack = FastGradientSignMethod(wb_net, eps=0.3)
wb_adv_sample = wb_attack.generate(attacked_sample, attacked_true_label)
wb_raw_preds = softmax(wb_model.predict(wb_adv_sample), axis=1)
accuracy_test = np.mean(
np.equal(np.argmax(wb_model.predict(attacked_sample), axis=1),
np.argmax(attacked_true_label, axis=1)))
LOGGER.info(TAG, "prediction accuracy before white-box attack is : %s",
accuracy_test)
accuracy_adv = np.mean(
np.equal(np.argmax(wb_raw_preds, axis=1),
np.argmax(attacked_true_label, axis=1)))
LOGGER.info(TAG, "prediction accuracy after white-box attack is : %s",
accuracy_adv)
# improve the robustness of model with white-box adversarial examples
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=False)
opt = nn.Momentum(wb_net.trainable_params(), 0.01, 0.09)
nad = NaturalAdversarialDefense(wb_net,
loss_fn=loss,
optimizer=opt,
bounds=(0.0, 1.0),
eps=0.3)
wb_net.set_train(False)
nad.batch_defense(inputs[:5000], labels[:5000], batch_size=32, epochs=10)
wb_def_preds = wb_net(Tensor(wb_adv_sample)).asnumpy()
wb_def_preds = softmax(wb_def_preds, axis=1)
accuracy_def = np.mean(
np.equal(np.argmax(wb_def_preds, axis=1),
np.argmax(attacked_true_label, axis=1)))
LOGGER.info(TAG, "prediction accuracy after defense is : %s", accuracy_def)
# calculate defense evaluation metrics for defense against white-box attack
wb_def_evaluate = DefenseEvaluate(wb_raw_preds, wb_def_preds,
np.argmax(attacked_true_label, axis=1))
LOGGER.info(TAG, 'defense evaluation for white-box adversarial attack')
LOGGER.info(
TAG, 'classification accuracy variance (CAV) is : {:.2f}'.format(
wb_def_evaluate.cav()))
LOGGER.info(
TAG, 'classification rectify ratio (CRR) is : {:.2f}'.format(
wb_def_evaluate.crr()))
LOGGER.info(
TAG, 'classification sacrifice ratio (CSR) is : {:.2f}'.format(
wb_def_evaluate.csr()))
LOGGER.info(
TAG, 'classification confidence variance (CCV) is : {:.2f}'.format(
wb_def_evaluate.ccv()))
LOGGER.info(
TAG, 'classification output stability is : {:.2f}'.format(
wb_def_evaluate.cos()))
# calculate defense evaluation metrics for defense against black-box attack
LOGGER.info(TAG, 'defense evaluation for black-box adversarial attack')
bb_raw_preds = []
bb_def_preds = []
raw_query_counts = []
raw_query_time = []
def_query_counts = []
def_query_time = []
def_detection_counts = []
# gen black-box adversarial examples of test data
bb_net = LeNet5()
load_param_into_net(bb_net, load_dict)
bb_model = ModelToBeAttacked(bb_net, defense=False)
attack_rm = GeneticAttack(model=bb_model,
pop_size=6,
mutation_rate=0.05,
per_bounds=0.1,
step_size=0.25,
temp=0.1,
sparse=False)
attack_target_label = target_label[:attacked_size]
true_label = labels_sparse[:attacked_size + benign_size]
# evaluate robustness of original model
# gen black-box adversarial examples of test data
for idx in range(attacked_size):
raw_st = time.time()
raw_sl, raw_a, raw_qc = attack_rm.generate(
np.expand_dims(attacked_sample[idx], axis=0),
np.expand_dims(attack_target_label[idx], axis=0))
raw_t = time.time() - raw_st
bb_raw_preds.extend(softmax(bb_model.predict(raw_a), axis=1))
raw_query_counts.extend(raw_qc)
raw_query_time.append(raw_t)
for idx in range(benign_size):
raw_st = time.time()
bb_raw_pred = softmax(bb_model.predict(
np.expand_dims(benign_sample[idx], axis=0)),
axis=1)
raw_t = time.time() - raw_st
bb_raw_preds.extend(bb_raw_pred)
raw_query_counts.extend([0])
raw_query_time.append(raw_t)
accuracy_test = np.mean(
np.equal(np.argmax(bb_raw_preds[0:len(attack_target_label)], axis=1),
np.argmax(attack_target_label, axis=1)))
LOGGER.info(TAG, "attack success before adv defense is : %s",
accuracy_test)
# improve the robustness of model with similarity-based detector
bb_def_model = ModelToBeAttacked(bb_net,
defense=True,
train_images=inputs[0:6000])
# attack defensed model
attack_dm = GeneticAttack(model=bb_def_model,
pop_size=6,
mutation_rate=0.05,
per_bounds=0.1,
step_size=0.25,
temp=0.1,
sparse=False)
for idx in range(attacked_size):
def_st = time.time()
def_sl, def_a, def_qc = attack_dm.generate(
np.expand_dims(attacked_sample[idx], axis=0),
np.expand_dims(attack_target_label[idx], axis=0))
def_t = time.time() - def_st
det_res = bb_def_model.get_detected_result()
def_detection_counts.append(np.sum(det_res[-def_qc[0]:]))
bb_def_preds.extend(softmax(bb_def_model.predict(def_a), axis=1))
def_query_counts.extend(def_qc)
def_query_time.append(def_t)
for idx in range(benign_size):
def_st = time.time()
bb_def_pred = softmax(bb_def_model.predict(
np.expand_dims(benign_sample[idx], axis=0)),
axis=1)
def_t = time.time() - def_st
det_res = bb_def_model.get_detected_result()
def_detection_counts.append(np.sum(det_res[-1]))
bb_def_preds.extend(bb_def_pred)
def_query_counts.extend([0])
def_query_time.append(def_t)
accuracy_adv = np.mean(
np.equal(np.argmax(bb_def_preds[0:len(attack_target_label)], axis=1),
np.argmax(attack_target_label, axis=1)))
LOGGER.info(TAG, "attack success rate after adv defense is : %s",
accuracy_adv)
bb_raw_preds = np.array(bb_raw_preds).astype(np.float32)
bb_def_preds = np.array(bb_def_preds).astype(np.float32)
# check evaluate data
max_queries = 6000
def_evaluate = BlackDefenseEvaluate(bb_raw_preds, bb_def_preds,
np.array(raw_query_counts),
np.array(def_query_counts),
np.array(raw_query_time),
np.array(def_query_time),
np.array(def_detection_counts),
true_label, max_queries)
LOGGER.info(
TAG, 'query count variance of adversaries is : {:.2f}'.format(
def_evaluate.qcv()))
LOGGER.info(
TAG, 'attack success rate variance of adversaries '
'is : {:.2f}'.format(def_evaluate.asv()))
LOGGER.info(
TAG, 'false positive rate (FPR) of the query-based detector '
'is : {:.2f}'.format(def_evaluate.fpr()))
LOGGER.info(
TAG, 'the benign query response time variance (QRV) '
'is : {:.2f}'.format(def_evaluate.qrv()))
def test_similarity_detector():
"""
Similarity Detector test.
"""
# load trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
# get mnist data
data_list = "./MNIST_unzip/test"
batch_size = 1000
ds = generate_mnist_dataset(data_list, batch_size=batch_size)
model = ModelToBeAttacked(net)
batch_num = 10 # the number of batches of input samples
all_images = []
true_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
all_images.append(images)
true_labels.append(labels)
pred_labels = np.argmax(model.predict(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
all_images = np.concatenate(all_images)
true_labels = np.concatenate(true_labels)
predict_labels = np.concatenate(predict_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %s", accuracy)
train_images = all_images[0:6000, :, :, :]
attacked_images = all_images[0:10, :, :, :]
attacked_labels = true_labels[0:10]
# generate malicious query sequence of black attack
attack = PSOAttack(model,
bounds=(0.0, 1.0),
pm=0.5,
sparse=True,
t_max=1000)
success_list, adv_data, query_list = attack.generate(
attacked_images, attacked_labels)
LOGGER.info(TAG, 'pso attack success_list: %s', success_list)
LOGGER.info(TAG, 'average of query counts is : %s', np.mean(query_list))
pred_logits_adv = model.predict(adv_data)
# rescale predict confidences into (0, 1).
pred_logits_adv = softmax(pred_logits_adv, axis=1)
pred_lables_adv = np.argmax(pred_logits_adv, axis=1)
accuracy_adv = np.mean(np.equal(pred_lables_adv, attacked_labels))
LOGGER.info(TAG, "prediction accuracy after attacking is : %g",
accuracy_adv)
benign_queries = all_images[6000:10000, :, :, :]
suspicious_queries = model.get_queries()
# explicit threshold not provided, calculate threshold for K
encoder = Model(EncoderNet(encode_dim=256))
detector = SimilarityDetector(max_k_neighbor=50, trans_model=encoder)
detector.fit(inputs=train_images)
# test benign queries
detector.detect(benign_queries)
fpr = len(detector.get_detected_queries()) / benign_queries.shape[0]
LOGGER.info(TAG, 'Number of false positive of attack detector is : %s',
len(detector.get_detected_queries()))
LOGGER.info(TAG, 'False positive rate of attack detector is : %s', fpr)
# test attack queries
detector.clear_buffer()
detector.detect(suspicious_queries)
LOGGER.info(TAG, 'Number of detected attack queries is : %s',
len(detector.get_detected_queries()))
LOGGER.info(TAG, 'The detected attack query indexes are : %s',
detector.get_detected_queries())
def test_nes_mnist_attack():
"""
hsja-Attack test
"""
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
net.set_train(False)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size=batch_size)
# prediction accuracy before attack
model = ModelToBeAttacked(net)
# the number of batches of attacking samples
batch_num = 5
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.concatenate(test_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %s", accuracy)
test_images = np.concatenate(test_images)
# attacking
scene = 'Query_Limit'
if scene == 'Query_Limit':
top_k = -1
elif scene == 'Partial_Info':
top_k = 5
elif scene == 'Label_Only':
top_k = 5
success = 0
queries_num = 0
nes_instance = NES(model, scene, top_k=top_k)
test_length = 32
advs = []
for img_index in range(test_length):
# Initial image and class selection
initial_img = test_images[img_index]
orig_class = true_labels[img_index]
initial_img = [initial_img]
target_class = random_target_labels([orig_class], true_labels)
target_image = create_target_images(test_images, true_labels,
target_class)
nes_instance.set_target_images(target_image)
tag, adv, queries = nes_instance.generate(initial_img, target_class)
if tag[0]:
success += 1
queries_num += queries[0]
advs.append(adv)
advs = np.reshape(advs, (len(advs), 1, 32, 32))
adv_pred = np.argmax(model.predict(advs), axis=1)
adv_accuracy = np.mean(np.equal(adv_pred, true_labels[:test_length]))
LOGGER.info(TAG, "prediction accuracy after attacking is : %s",
adv_accuracy)
def test_momentum_diverse_input_iterative_method():
"""
M-DI2-FGSM Attack Test for CPU device.
"""
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
# upload trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
# get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size)
# prediction accuracy before attack
model = Model(net)
batch_num = 32 # the number of batches of attacking samples
test_images = []
test_labels = []
predict_labels = []
i = 0
for data in ds.create_tuple_iterator():
i += 1
images = data[0].astype(np.float32)
labels = data[1]
test_images.append(images)
test_labels.append(labels)
pred_labels = np.argmax(model.predict(Tensor(images)).asnumpy(),
axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
predict_labels = np.concatenate(predict_labels)
true_labels = np.concatenate(test_labels)
accuracy = np.mean(np.equal(predict_labels, true_labels))
LOGGER.info(TAG, "prediction accuracy before attacking is : %s", accuracy)
# attacking
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
attack = MomentumDiverseInputIterativeMethod(net, loss_fn=loss)
start_time = time.clock()
adv_data = attack.batch_generate(np.concatenate(test_images),
true_labels,
batch_size=32)
stop_time = time.clock()
pred_logits_adv = model.predict(Tensor(adv_data)).asnumpy()
# rescale predict confidences into (0, 1).
pred_logits_adv = softmax(pred_logits_adv, axis=1)
pred_labels_adv = np.argmax(pred_logits_adv, axis=1)
accuracy_adv = np.mean(np.equal(pred_labels_adv, true_labels))
LOGGER.info(TAG, "prediction accuracy after attacking is : %s",
accuracy_adv)
attack_evaluate = AttackEvaluate(
np.concatenate(test_images).transpose(0, 2, 3, 1),
np.eye(10)[true_labels], adv_data.transpose(0, 2, 3, 1),
pred_logits_adv)
LOGGER.info(TAG, 'mis-classification rate of adversaries is : %s',
attack_evaluate.mis_classification_rate())
LOGGER.info(TAG, 'The average confidence of adversarial class is : %s',
attack_evaluate.avg_conf_adv_class())
LOGGER.info(TAG, 'The average confidence of true class is : %s',
attack_evaluate.avg_conf_true_class())
LOGGER.info(
TAG, 'The average distance (l0, l2, linf) between original '
'samples and adversarial samples are: %s',
attack_evaluate.avg_lp_distance())
LOGGER.info(
TAG, 'The average structural similarity between original '
'samples and adversarial samples are: %s', attack_evaluate.avg_ssim())
LOGGER.info(TAG, 'The average costing time is %s',
(stop_time - start_time) / (batch_num * batch_size))
def test_nad_method():
"""
NAD-Defense test.
"""
# 1. load trained network
ckpt_name = './trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_name)
load_param_into_net(net, load_dict)
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=False)
opt = nn.Momentum(net.trainable_params(), 0.01, 0.09)
nad = NaturalAdversarialDefense(net,
loss_fn=loss,
optimizer=opt,
bounds=(0.0, 1.0),
eps=0.3)
# 2. get test data
data_list = "./MNIST_unzip/test"
batch_size = 32
ds_test = generate_mnist_dataset(data_list,
batch_size=batch_size,
sparse=False)
inputs = []
labels = []
for data in ds_test.create_tuple_iterator():
inputs.append(data[0].astype(np.float32))
labels.append(data[1])
inputs = np.concatenate(inputs)
labels = np.concatenate(labels)
# 3. get accuracy of test data on original model
net.set_train(False)
acc_list = []
batchs = inputs.shape[0] // batch_size
for i in range(batchs):
batch_inputs = inputs[i * batch_size:(i + 1) * batch_size]
batch_labels = np.argmax(labels[i * batch_size:(i + 1) * batch_size],
axis=1)
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.debug(TAG, 'accuracy of TEST data on original model is : %s',
np.mean(acc_list))
# 4. get adv of test data
attack = FastGradientSignMethod(net, eps=0.3)
adv_data = attack.batch_generate(inputs, labels)
LOGGER.debug(TAG, 'adv_data.shape is : %s', adv_data.shape)
# 5. get accuracy of adv data on original model
net.set_train(False)
acc_list = []
batchs = adv_data.shape[0] // batch_size
for i in range(batchs):
batch_inputs = adv_data[i * batch_size:(i + 1) * batch_size]
batch_labels = np.argmax(labels[i * batch_size:(i + 1) * batch_size],
axis=1)
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.debug(TAG, 'accuracy of adv data on original model is : %s',
np.mean(acc_list))
# 6. defense
net.set_train()
nad.batch_defense(inputs, labels, batch_size=32, epochs=10)
# 7. get accuracy of test data on defensed model
net.set_train(False)
acc_list = []
batchs = inputs.shape[0] // batch_size
for i in range(batchs):
batch_inputs = inputs[i * batch_size:(i + 1) * batch_size]
batch_labels = np.argmax(labels[i * batch_size:(i + 1) * batch_size],
axis=1)
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.debug(TAG, 'accuracy of TEST data on defensed model is : %s',
np.mean(acc_list))
# 8. get accuracy of adv data on defensed model
acc_list = []
batchs = adv_data.shape[0] // batch_size
for i in range(batchs):
batch_inputs = adv_data[i * batch_size:(i + 1) * batch_size]
batch_labels = np.argmax(labels[i * batch_size:(i + 1) * batch_size],
axis=1)
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.debug(TAG, 'accuracy of adv data on defensed model is : %s',
np.mean(acc_list))
