def build(self):
temp = 100
network1 = DistillationWrapper(resnet.ResNet18().to(self.device), temp)
self.target_instance = DefenseInstance(model=network1)
network2 = DistillationWrapper(
ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True), temp)
instance = DefenseInstance(model=network2, detector=None)
return instance
def build(self):
import zoo.saved_instances.Odds.cifar_model as cifar_model
import zoo.saved_instances.Odds.robustify as robustify
classifier = cifar_model.cifar10(128, pretrained=True, map_location=None, trained_adv=False).to(self.device)
detector = robustify.OddsAreOdd(classifier=classifier)
instance = DefenseInstance(model=classifier, detector=detector)
return instance
def build(self):
Jpegdefense = JpegCompression((0, 1), 30)
RSdefense = ReverseSigmoid()
network1 = SimpleNet(in_ch=1, out_ch=10)
classifier = DAGModule([Jpegdefense, network1, RSdefense], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
Jpegdefense = JpegCompression((0, 1), COMPRESSION_QUALITY)
RSdefense = ReverseSigmoid()
classifier = DAGModule([Jpegdefense, self.network, RSdefense],
device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.saved_instances.TurningWeakness.vgg19 import vgg19, TurningWeaknessDetector
classifier = vgg19()
classifier.features = torch.nn.DataParallel(classifier.features)
classifier.to(self.device)
detector = TurningWeaknessDetector(classifier=classifier)
instance = DefenseInstance(model=classifier, detector=detector)
return instance
def build(self):
network1 = SimpleNet1(in_ch=1, out_ch=10)
network2 = SimpleNet2(in_ch=1, out_ch=10)
detector1 = SimpleDetector(network1)
module1 = DAGModule([network1, network2], dependency=[[-1], [0]], output=[1], device=self.device)
module2 = DAGModule([network1, detector1], dependency=[[-1], [0]], output=[1], device=self.device)
instance = DefenseInstance(model=module1, detector=module2)
return instance
def build(self):
ITdefense = InputTransformation(degrees=(-30, 30))
# Varmin = VarianceMinimization(prob=0.3, norm=2)
# Jpegdefense = JpegCompression((0, 1), 30)
# RSdefense = ReverseSigmoid()
network1 = SimpleNet(in_ch=1, out_ch=10)
classifier = DAGModule([ITdefense, network1], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.saved_instances.JEM.CCF_model import CCF
# nsteps = 10
classifier = CCF(28, 10, None, dropout_rate=0, n_classes=10).to(self.device)
# classifier = WrapperModel(f, nsteps).to(self.device)
# classifier = gradient_attack_wrapper(f)
# classifier = f.to(self.device)
# classifier = nn.DataParallel(classifier).to(self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.saved_instances.JEM.CCF_model import CCF
# nsteps = 10
ITdefense = InputTransformation(degrees=(-ROTATION_DEGREES, ROTATION_DEGREES))
classifier = DAGModule([ITdefense, self.network], device=self.device)
# classifier = WrapperModel(f, nsteps).to(self.device)
# classifier = gradient_attack_wrapper(f)
# classifier = f.to(self.device)
# classifier = nn.DataParallel(classifier).to(self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.saved_instances.JEM.CCF_model import CCF
Jpegdefense = JpegCompression((0, 1), COMPRESSION_QUALITY)
RSdefense = ReverseSigmoid()
classifier = DAGModule([Jpegdefense, self.network, RSdefense], device=self.device)
# classifier = WrapperModel(f, nsteps).to(self.device)
# classifier = gradient_attack_wrapper(f)
# classifier = f.to(self.device)
# classifier = nn.DataParallel(classifier).to(self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
network1 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True)
classifier = DAGModule([network1], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
TEdefense = ThermometerEncoding((0, 1), num_space=10)
network1 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=2,
linear_size=200,
input_channel=30,
with_normalization=False)
classifier = DAGModule([TEdefense, network1], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
import zoo.saved_instances.CCAT.common.state as state
from zoo.saved_instances.CCAT.detector import CCATDetector
import sys
import os
dir_path = os.path.dirname(os.path.realpath(__file__))
sys.path.insert(1, dir_path + '/saved_instances/CCAT')
import zoo.saved_instances.CCAT.models as models
model_file = osp.join(self.path, self.name, "classifier.model")
state = state.State.load(model_file)
classifier = state.model.to(self.device)
detector = CCATDetector(classifier, 0.4)
instance = DefenseInstance(model=classifier, detector=detector) # , detector=detector)
return instance
def build(self):
Jpegdefense = JpegCompression((0, 1), 60)
RSdefense = ReverseSigmoid()
network1 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True)
classifier = DAGModule([Jpegdefense, network1, RSdefense],
device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.wideresnet import WideResNetNormalize
from zoo.saved_instances.FeatureScatter.attack_methods import Attack_FeaScatter
config_feature_scatter = {
'train': True,
'epsilon': 8.0 / 255 * 2,
'num_steps': 1,
'step_size': 8.0 / 255 * 2,
'random_start': True,
'ls_factor': 0.5,
}
basic_net = WideResNetNormalize(28, 10, widen_factor=10, dropRate=0.0)
net = Attack_FeaScatter(basic_net, config_feature_scatter)
net = torch.nn.DataParallel(net)
instance = DefenseInstance(model=net, detector=None)
return instance
def build(self):
ITdefense = InputTransformation(degrees=(-30, 30))
TEdefense = ThermometerEncoding((0, 1), num_space=10)
Jpegdefense = JpegCompression((0, 1), 30)
RSdefense = ReverseSigmoid()
network1 = SimpleNet(in_ch=1, out_ch=10)
network2 = SimpleNet(in_ch=10, out_ch=10)
network3 = SimpleNet(in_ch=1, out_ch=10)
module1 = DAGModule([TEdefense, network2], device=self.device)
module2 = DAGModule([Jpegdefense, network1, RSdefense], device=self.device)
module3 = DAGModule([ITdefense, network3], device=self.device)
losscomb = LossCombine()
classifier = DAGModule([module1, module2, module3, losscomb], dependency=[[-1], [-1], [-1], [0, 1, 2]],
output=[3], device=self.device)
instance = DefenseInstance(model=classifier)
return instance
class Model8(CIFARInstance):
def __init__(self, load=False, device='cuda'):
name = 'model8'
self.target_instance = None
super(Model8, self).__init__(name, load, device)
def build(self):
temp = 100
network1 = DistillationWrapper(resnet.ResNet18().to(self.device), temp)
self.target_instance = DefenseInstance(model=network1)
network2 = DistillationWrapper(
ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True), temp)
instance = DefenseInstance(model=network2, detector=None)
return instance
def train(self, train_loader, eps, save=True):
train_loss = nn.CrossEntropyLoss()
optimizer = optim.Adam(
self.target_instance.model.train_model.parameters(), lr=0.001)
nb_epoches = 25
attack = None
kwargs = {}
self.target_instance.train()
self.target_instance.advfit(train_loader,
loss_fcn=train_loss,
optimizer=optimizer,
epsilon=eps,
attack=attack,
nb_epochs=nb_epoches,
ratio=0.6,
**kwargs)
self.target_instance.eval()
nb_epoches = 25
optimizer = optim.Adam(self.instance.model.train_model.parameters(),
lr=0.001)
defensive_distillation = DefensiveDistillation(self.target_instance)
defensive_distillation.fit(self.instance,
loader=train_loader,
optimizer=optimizer,
nb_epochs=nb_epoches,
temp=5)
if save:
self._save()
return
def build(self):
ITdefense = InputTransformation(degrees=(-30, 30))
TEdefense = ThermometerEncoding((0, 1), num_space=10)
Jpegdefense = JpegCompression((0, 1), 30)
RSdefense = ReverseSigmoid()
network1 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True)
network2 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=2,
linear_size=200,
input_channel=30,
with_normalization=False)
network3 = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True)
module1 = DAGModule([TEdefense, network2], device=self.device)
module2 = DAGModule([Jpegdefense, network1, RSdefense],
device=self.device)
module3 = DAGModule([ITdefense, network3], device=self.device)
losscomb = LossCombine()
classifier = DAGModule([module1, module2, module3, losscomb],
dependency=[[-1], [-1], [-1], [0, 1, 2]],
output=[3],
device=self.device)
instance = DefenseInstance(model=classifier)
return instance
def build(self):
network = ConvMedBig(device=self.device,
dataset='cifar10',
width1=4,
width2=4,
width3=4,
linear_size=200,
input_channel=3,
with_normalization=True)
network1 = ConvMedBig1(network)
network2 = ConvMedBig2(network)
detector1 = SimpleDetector(network1, 200)
module1 = DAGModule([network1, network2],
dependency=[[-1], [0]],
output=[1],
device=self.device)
module2 = DAGModule([network1, detector1],
dependency=[[-1], [0]],
output=[1],
device=self.device)
classifier = module1
instance = DefenseInstance(model=classifier, detector=module2)
# return instance
return instance
def build(self):
TEdefense = ThermometerEncoding((0, 1), num_space=10)
network1 = SimpleNet(in_ch=10, out_ch=10)
classifier = DAGModule([TEdefense, network1], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
network1 = SimpleNet(in_ch=1, out_ch=10)
classifier = DAGModule([network1], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.saved_instances.kWTA_0p1_adv import resnet
classifier = resnet.SparseResNet18(sparsities=[0.1,0.1,0.1,0.1], sparse_func='vol').to(self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
import zoo.saved_instances.EnResNet.enresnet_wideresnet as EWRes
classifier = EWRes.WideResNet(noise_coef=0.0).cuda()
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
network = resnet.SparseResNet18(sparsities=[0.1, 0.1, 0.1, 0.1],
sparse_func='reg').to(self.device)
classifier = DAGModule([network], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
model = load_model(model_name=self.name, norm='Linf')
instance = DefenseInstance(model=model.to(self.device), detector=None)
return instance
def build(self):
ITdefense = InputTransformation(degrees=(-ROTATION_DEGREES, ROTATION_DEGREES))
classifier = DAGModule([ITdefense, self.network], device=self.device)
instance = DefenseInstance(model=classifier, detector=None)
return instance
def build(self):
from zoo.wideresnet import WideResNet
classifier = WideResNet(28, 10, widen_factor=10, dropRate=0.0)
classifier = nn.DataParallel(classifier).cuda()
instance = DefenseInstance(model=classifier, detector=None)
return instance