def test_attack_eval():
# prepare test data
np.random.seed(1024)
inputs = np.random.normal(size=(3, 512, 512, 3))
labels = np.array([[0.1, 0.1, 0.2, 0.6], [0.1, 0.7, 0.0, 0.2],
[0.8, 0.1, 0.0, 0.1]])
adv_x = inputs + np.ones((3, 512, 512, 3)) * 0.001
adv_y = np.array([[0.1, 0.1, 0.2, 0.6], [0.1, 0.0, 0.8, 0.1],
[0.0, 0.9, 0.1, 0.0]])
# create obj
attack_eval = AttackEvaluate(inputs, labels, adv_x, adv_y)
# run eval
mr = attack_eval.mis_classification_rate()
acac = attack_eval.avg_conf_adv_class()
l_0, l_2, l_inf = attack_eval.avg_lp_distance()
ass = attack_eval.avg_ssim()
nte = attack_eval.nte()
res = [mr, acac, l_0, l_2, l_inf, ass, nte]
# compare
expected_value = [0.6666, 0.8500, 1.0, 0.0009, 0.0001, 0.9999, 0.75]
assert np.allclose(res, expected_value, 0.0001, 0.0001)
def test_lbfgs_attack():
"""
LBFGS-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=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
is_targeted = True
if is_targeted:
targeted_labels = np.random.randint(0, 10, size=len(true_labels)).astype(np.int32)
for i, true_l in enumerate(true_labels):
if targeted_labels[i] == true_l:
targeted_labels[i] = (targeted_labels[i] + 1) % 10
else:
targeted_labels = true_labels.astype(np.int32)
loss = SoftmaxCrossEntropyWithLogits(sparse=True)
attack = LBFGS(net, is_targeted=is_targeted, loss_fn=loss)
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.eye(10)[true_labels],
adv_data.transpose(0, 2, 3, 1),
pred_logits_adv,
targeted=is_targeted,
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_pso_attack_on_mnist():
"""
PSO-Attack test
"""
# 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=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(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(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)
# attacking
attack = PSOAttack(model, bounds=(0.0, 1.0), pm=0.5, sparse=True)
start_time = time.clock()
success_list, adv_data, query_list = attack.generate(
np.concatenate(test_images), np.concatenate(test_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_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)
test_labels_onehot = np.eye(10)[np.concatenate(test_labels)]
attack_evaluate = AttackEvaluate(np.concatenate(test_images),
test_labels_onehot, adv_data,
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_salt_and_pepper_attack_on_mnist():
"""
Salt-and-Pepper-Attack test
"""
# 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=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(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(images), axis=1)
predict_labels.append(pred_labels)
if i >= batch_num:
break
LOGGER.debug(
TAG,
'model input image shape is: {}'.format(np.array(test_images).shape))
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 = SaltAndPepperNoiseAttack(model=model,
is_targeted=is_target,
sparse=True)
if is_target:
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
else:
targeted_labels = true_labels
LOGGER.debug(
TAG, 'input shape is: {}'.format(np.concatenate(test_images).shape))
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())