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_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_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_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))