def test_lenet_mnist_coverage_ascend():
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
# load network
net = Net()
model = Model(net)
# initialize fuzz test with training dataset
training_data = (np.random.random((10000, 10))*20).astype(np.float32)
model_fuzz_test = ModelCoverageMetrics(model, 10, 1000, training_data)
# fuzz test with original test data
# get test data
test_data = (np.random.random((2000, 10))*20).astype(np.float32)
test_labels = np.random.randint(0, 10, 2000)
test_labels = (np.eye(10)[test_labels]).astype(np.float32)
model_fuzz_test.calculate_coverage(test_data)
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
attack = FastGradientSignMethod(net, eps=0.3)
adv_data = attack.batch_generate(test_data, test_labels, batch_size=32)
model_fuzz_test.calculate_coverage(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 test_lenet_mnist_coverage_cpu():
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
# load network
net = Net()
model = Model(net)
# initialize fuzz test with training dataset
neuron_num = 10
segmented_num = 1000
training_data = (np.random.random((10000, 10)) * 20).astype(np.float32)
model_fuzz_test = ModelCoverageMetrics(model, neuron_num, segmented_num,
training_data)
# fuzz test with original test data
# get test data
test_data = (np.random.random((2000, 10)) * 20).astype(np.float32)
test_labels = np.random.randint(0, 10, 2000).astype(np.int32)
model_fuzz_test.calculate_coverage(test_data)
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(sparse=True)
attack = FastGradientSignMethod(net, eps=0.3, loss_fn=loss)
adv_data = attack.batch_generate(test_data, test_labels, batch_size=32)
model_fuzz_test.calculate_coverage(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 test_fast_gradient_sign_method():
"""
Fast gradient sign method unit test.
"""
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
input_np = np.asarray([[0.1, 0.2, 0.7]], np.float32)
label = np.asarray([2], np.int32)
label = np.eye(3)[label].astype(np.float32)
attack = FastGradientSignMethod(Net())
ms_adv_x = attack.generate(input_np, label)
assert np.any(ms_adv_x != input_np), 'Fast gradient sign method: generate' \
' value must not be equal to' \
' original value.'
def test_lenet_mnist_coverage():
# upload trained network
ckpt_path = '../common/networks/lenet5/trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_path)
load_param_into_net(net, load_dict)
model = Model(net)
# get training data
data_list = "../common/dataset/MNIST/train"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size, sparse=True)
train_images = []
for data in ds.create_tuple_iterator(output_numpy=True):
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, 10, 1000, train_images)
# fuzz test with original test data
# get test data
data_list = "../common/dataset/MNIST/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(output_numpy=True):
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.calculate_coverage(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(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.calculate_coverage(adv_data, bias_coefficient=0.5)
LOGGER.info(TAG, 'KMNC of this adv data is : %s', model_fuzz_test.get_kmnc())
LOGGER.info(TAG, 'NBC of this adv data is : %s', model_fuzz_test.get_nbc())
LOGGER.info(TAG, 'SNAC of this adv data is : %s', model_fuzz_test.get_snac())
def test_fast_gradient_sign_method():
"""
FGSM-Attack test
"""
context.set_context(mode=context.GRAPH_MODE)
# get network
net = resnet50_cifar10(10)
# create test data
test_images = np.random.rand(64, 3, 224, 224).astype(np.float32)
test_labels = np.random.randint(10, size=64).astype(np.int32)
# attacking
loss_fn = CrossEntropyLoss()
attack = FastGradientSignMethod(net, eps=0.1, loss_fn=loss_fn)
adv_data = attack.batch_generate(test_images, test_labels, batch_size=32)
assert np.any(adv_data != test_images)
def test_ead():
"""UT for ensemble adversarial defense."""
num_classes = 10
batch_size = 64
sparse = False
context.set_context(mode=context.GRAPH_MODE)
context.set_context(device_target='Ascend')
# create test data
inputs = np.random.rand(batch_size, 1, 32, 32).astype(np.float32)
labels = np.random.randint(num_classes, size=batch_size).astype(np.int32)
if not sparse:
labels = np.eye(num_classes)[labels].astype(np.float32)
net = Net()
loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=sparse)
optimizer = Momentum(net.trainable_params(), 0.001, 0.9)
net = Net()
fgsm = FastGradientSignMethod(net, loss_fn=loss_fn)
pgd = ProjectedGradientDescent(net, loss_fn=loss_fn)
ead = EnsembleAdversarialDefense(net, [fgsm, pgd],
loss_fn=loss_fn,
optimizer=optimizer)
LOGGER.set_level(logging.DEBUG)
LOGGER.debug(TAG, '---start ensemble adversarial defense--')
loss = ead.defense(inputs, labels)
LOGGER.debug(TAG, '---end ensemble adversarial defense--')
assert np.any(loss >= 0.0)
def test_nad_method():
"""
NAD-Defense test.
"""
mnist_path = "../../common/dataset/MNIST"
batch_size = 32
# 1. train original model
ds_train = generate_mnist_dataset(os.path.join(mnist_path, "train"),
batch_size=batch_size,
repeat_size=1)
net = LeNet5()
loss = SoftmaxCrossEntropyWithLogits(sparse=True)
opt = nn.Momentum(net.trainable_params(), 0.01, 0.09)
model = Model(net, loss, opt, metrics=None)
model.train(10,
ds_train,
callbacks=[LossMonitor()],
dataset_sink_mode=False)
# 2. get test data
ds_test = generate_mnist_dataset(os.path.join(mnist_path, "test"),
batch_size=batch_size,
repeat_size=1)
inputs = []
labels = []
for data in ds_test.create_tuple_iterator():
inputs.append(data[0].asnumpy().astype(np.float32))
labels.append(data[1].asnumpy())
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 = labels[i * batch_size:(i + 1) * batch_size]
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.info(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, loss_fn=loss)
adv_data = attack.batch_generate(inputs, labels)
LOGGER.info(TAG, 'adv_data.shape is : %s', adv_data.shape)
# 5. get accuracy of adv data on original 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 = labels[i * batch_size:(i + 1) * batch_size]
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.info(TAG, 'accuracy of adv data on original model is : %s',
np.mean(acc_list))
# 6. defense
ds_train = generate_mnist_dataset(os.path.join(mnist_path, "train"),
batch_size=batch_size,
repeat_size=1)
inputs_train = []
labels_train = []
for data in ds_train.create_tuple_iterator():
inputs_train.append(data[0].asnumpy().astype(np.float32))
labels_train.append(data[1].asnumpy())
inputs_train = np.concatenate(inputs_train)
labels_train = np.concatenate(labels_train)
net.set_train()
nad = NaturalAdversarialDefense(net,
loss_fn=loss,
optimizer=opt,
bounds=(0.0, 1.0),
eps=0.3)
nad.batch_defense(inputs_train, labels_train, 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 = labels[i * batch_size:(i + 1) * batch_size]
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.info(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 = labels[i * batch_size:(i + 1) * batch_size]
logits = net(Tensor(batch_inputs)).asnumpy()
label_pred = np.argmax(logits, axis=1)
acc_list.append(np.mean(batch_labels == label_pred))
LOGGER.info(TAG, 'accuracy of adv data on defensed model is : %s',
np.mean(acc_list))
def test_fast_gradient_sign_method():
"""
FGSM-Attack test for CPU device.
"""
# upload trained network
ckpt_path = '../../../common/networks/lenet5/trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
net = LeNet5()
load_dict = load_checkpoint(ckpt_path)
load_param_into_net(net, load_dict)
# get test data
data_list = "../../../common/dataset/MNIST/test"
batch_size = 32
ds = generate_mnist_dataset(data_list, batch_size)
# 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(output_numpy=True):
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(sparse=True)
attack = FastGradientSignMethod(net, eps=0.3, 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()
np.save('./adv_data', adv_data)
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_defense_evaluation():
# load trained network
current_dir = os.path.dirname(os.path.abspath(__file__))
ckpt_path = os.path.abspath(
os.path.join(
current_dir,
'../../common/networks/lenet5/trained_ckpt_file/checkpoint_lenet-10_1875.ckpt'
))
wb_net = LeNet5()
load_dict = load_checkpoint(ckpt_path)
load_param_into_net(wb_net, load_dict)
# get test data
data_list = "../../common/dataset/MNIST/test"
batch_size = 32
ds_test = generate_mnist_dataset(data_list, batch_size=batch_size)
inputs = []
labels = []
for data in ds_test.create_tuple_iterator(output_numpy=True):
inputs.append(data[0].astype(np.float32))
labels.append(data[1])
inputs = np.concatenate(inputs).astype(np.float32)
labels = np.concatenate(labels).astype(np.int32)
target_label = np.random.randint(0, 10, size=labels.shape[0])
for idx in range(labels.shape[0]):
while target_label[idx] == labels[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
loss = SoftmaxCrossEntropyWithLogits(sparse=True)
wb_attack = FastGradientSignMethod(wb_net, eps=0.3, loss_fn=loss)
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),
attacked_true_label))
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), attacked_true_label))
LOGGER.info(TAG, "prediction accuracy after white-box attack is : %s",
accuracy_adv)
# improve the robustness of model with white-box adversarial examples
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), attacked_true_label))
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,
attacked_true_label)
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=0.5,
step_size=0.25,
temp=0.1,
sparse=False)
attack_target_label = target_label[:attacked_size]
true_label = labels[: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_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=0.5,
step_size=0.25,
temp=0.1,
sparse=False)
for idx in range(attacked_size):
def_st = time.time()
_, 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()))