def load_classifier(model, data):
"""
Load ART PyTorch classifier from pytorch model
:param model: pytorch model instance
:param data: data class
:return: ART classifier
"""
# not used but mandatory for ART
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
classifier = PyTorchClassifier(
model=model,
clip_values=(data.min_pixel_value, data.max_pixel_value),
loss=criterion,
optimizer=optimizer,
input_shape=tuple(data.trainset.data.shape[1:]),
nb_classes=data.num_classes,
device_type="gpu" if USE_CUDA else "cpu")
classifier.set_learning_phase(False)
return classifier
# Step 2a: Define the loss function and the optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.01)
# Step 3: Create the ART classifier
classifier = PyTorchClassifier(
model=model,
clip_values=(-1, 1),
loss=criterion,
optimizer=optimizer,
input_shape=(1, 28, 28),
nb_classes=10,
)
classifier.set_learning_phase(False)
# Step 4: Evaluate the ART classifier on benign test examples
count = 0
ben_sum = 0
att_sum = 0
for i, (imgs, labels) in enumerate(testloader):
predictions = classifier.predict(imgs)
accuracy = np.sum(np.argmax(predictions, axis=1) == labels.numpy()) / 128
#print("Accuracy on benign test examples: {}%".format(accuracy * 100))
ben_sum += accuracy
# Step 5: Generate adversarial test examples
attack = FastGradientMethod(classifier=classifier)
x_test_adv = attack.generate(x=imgs)