def __init__(self, keras_model, rc_name):
# Example of rc_name: FeatureSqueezing?squeezer=bit_depth_1
self.model_predict = lambda x: keras_model.predict(x)
subject, params = parse_params(rc_name)
assert subject == 'MagNet'
if FLAGS.dataset_name == "MNIST":
self.reformer = SimpleReformer(mnist_autoencoder_fpath)
elif FLAGS.dataset_name == "CIFAR-10":
self.reformer = SimpleReformer(cifar10_autoencoder_fpath)
self.filter = lambda x: self.reformer.heal(x)
def __init__(self, model, detector_name):
classifier = ClassifierWrapper(model)
autoencoder_model_fpath = os.path.join(
project_dir, "downloads/MagNet/defensive_models/CIFAR")
reformer = SimpleReformer(autoencoder_model_fpath)
id_reformer = IdReformer()
# Note: we may swap the two.
reconstructor = id_reformer
prober = reformer
# reconstructor = reformer
# prober = id_reformer
eb_detector = AEDetector(autoencoder_model_fpath, p=1)
db_detector_I = DBDetector(reconstructor, prober, classifier, T=10)
db_detector_II = DBDetector(reconstructor, prober, classifier, T=40)
detector_dict = dict()
detector_dict["db-I"] = db_detector_I
detector_dict["db-II"] = db_detector_II
detector_dict['eb'] = eb_detector
self.operator = Operator(CIFAR(), classifier, detector_dict, reformer)
def __init__(self, model, detector_name):
cur_folder = os.path.dirname(os.path.abspath(__file__))
autoencoder_model_I_fpath = os.path.join(project_dir, "downloads/MagNet/defensive_models/MNIST_I")
autoencoder_model_II_fpath = os.path.join(project_dir, "downloads/MagNet/defensive_models/MNIST_II")
detector_I = AEDetector(autoencoder_model_I_fpath, p=2)
detector_II = AEDetector(autoencoder_model_II_fpath, p=1)
reformer = SimpleReformer(autoencoder_model_I_fpath)
id_reformer = IdReformer()
classifier = ClassifierWrapper(model)
detector_dict = dict()
detector_dict["I"] = detector_I
detector_dict["II"] = detector_II
self.operator = Operator(MNIST(), classifier, detector_dict, reformer)
class MagNetRC:
def __init__(self, keras_model, rc_name):
# Example of rc_name: FeatureSqueezing?squeezer=bit_depth_1
self.model_predict = lambda x: keras_model.predict(x)
subject, params = parse_params(rc_name)
assert subject == 'MagNet'
if tf.flags.FLAGS.dataset_name == "MNIST":
self.reformer = SimpleReformer(mnist_autoencoder_fpath)
elif tf.flags.FLAGS.dataset_name == "CIFAR-10":
self.reformer = SimpleReformer(cifar10_autoencoder_fpath)
self.filter = lambda x: self.reformer.heal(x)
def predict(self, X):
X_filtered = self.filter(X)
Y_pred = self.model_predict(X_filtered)
return Y_pred
def visualize_and_predict(self, X):
X_filtered = self.filter(X)
Y_pred = self.model_predict(X_filtered)
return X_filtered, Y_pred
y = tf.py_func(my_func, [imgs], tf.float32, stateful=False)
return y
none_tf = none_py = lambda x:x
import sys, os
project_path = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(project_path)
from externals.MagNet.worker import SimpleReformer
mnist_autoencoder_fpath = os.path.join(project_path, "downloads/MagNet/defensive_models/MNIST_I")
cifar10_autoencoder_fpath = os.path.join(project_path, "downloads/MagNet/defensive_models/CIFAR")
reformer_mnist = SimpleReformer(mnist_autoencoder_fpath)
reformer_cifar10 = SimpleReformer(cifar10_autoencoder_fpath)
def magnet_mnist_py(imgs):
return reformer_mnist.heal(imgs)
def magnet_cifar10_py(imgs):
return reformer_cifar10.heal(imgs)
# Construct a name search function.
def isfloat(value):
try:
float(value)
return True
except:
return False