attack_name, batch_size, dataset_name = args.method, args.batch_size, args.dataset
goal, distance_metric = args.goal, args.distance_metric
kwargs = dict()
for kwarg in ('learning_rate', 'cw_loss_c', 'samples_per_draw',
'init_distortion'):
attr = getattr(args, kwarg)
if attr is not None:
kwargs[kwarg] = attr
if args.dimension_reduction_height is not None and args.dimension_reduction_width is not None:
kwargs['dimension_reduction'] = (args.dimension_reduction_height,
args.dimension_reduction_width)
if attack_name in ('fgsm', 'bim', 'pgd', 'mim'):
kwargs['loss'] = CrossEntropyLoss(model)
elif attack_name in ('nes', 'spsa', 'nattack'):
kwargs['loss'] = CWLoss(model)
benchmark = AttackBenchmark(attack_name, model, batch_size, dataset_name,
goal, distance_metric, session, **kwargs)
print('Configuring attack...')
benchmark.config(**config_kwargs)
print('Running benchmark...')
acc, acc_adv, total, succ, dist = benchmark.run(dataset, logger)
print('n={}, acc={:3f}, adv_acc={:3f}, succ={:3f}, dist_mean={:3f}'.format(
args.count, np.mean(acc.astype(np.float)),
np.mean(acc_adv.astype(np.float)),
np.sum(succ.astype(np.float)) / np.sum(total.astype(np.float)),
np.mean(dist)))
model_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'../example/cifar10/resnet56.py')
rs_model = load_model_from_path(model_path)
model = rs_model.load(session)
_, (xs_test, ys_test) = load_data()
xs_test = (xs_test / 255.0) * (model.x_max - model.x_min) + model.x_min
ys_test = ys_test.reshape(len(ys_test))
xs_ph = tf.placeholder(model.x_dtype, shape=(1, *model.x_shape))
lgs, lbs = model.logits_and_labels(xs_ph)
MAGNITUDE = 0.5 * (model.x_max - model.x_min)
loss = CWLoss(model)
attack = NAttack(
model=model,
loss=loss,
goal='ut',
distance_metric='l_2',
session=session,
samples_per_draw=100,
samples_batch_size=25,
dimension_reduction=(28, 28),
)
logger = tf.get_logger()
logger.setLevel(tf.logging.INFO)
attack.config(
max_queries=20000,
magnitude=MAGNITUDE,
from realsafe import BIM, CWLoss, EnsembleCWLoss, EnsembleRandomnessCWLoss
from realsafe.dataset import cifar10, dataset_to_iterator
from realsafe.model.loader import load_model_from_path
batch_size = 1000
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
model_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'../example/cifar10/resnet56.py')
model = load_model_from_path(model_path).load(session)
loss_op = CWLoss(model)
e_loss_op = EnsembleCWLoss([model, model], [0.5, 0.5])
er_loss_op = EnsembleRandomnessCWLoss(model, 10, session)
ds = cifar10.load_dataset_for_classifier(model).batch(batch_size).take(1)
_, xs, ys = next(dataset_to_iterator(ds, session))
xs_ph = tf.placeholder(model.x_dtype, shape=(batch_size, *model.x_shape))
ys_ph = tf.placeholder(model.y_dtype, shape=batch_size)
loss = loss_op(xs_ph, ys_ph)
e_loss = e_loss_op(xs_ph, ys_ph)
er_loss = er_loss_op(xs_ph, ys_ph)
dloss_dxs = tf.gradients(loss, xs_ph)[0]
de_loss_dxs = tf.gradients(e_loss, xs_ph)[0]