def test_targeted_attack(bn_model, bn_targeted_criterion, bn_images,
bn_labels):
attack = Attack(bn_model, bn_targeted_criterion)
advs = attack(bn_images, bn_labels, unpack=False)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_impossible(bn_model, bn_impossible_criterion, bn_images,
bn_labels):
attack = Attack(bn_model, bn_impossible_criterion)
advs = attack(bn_images, bn_labels, unpack=False)
for adv in advs:
assert adv.perturbed is None
assert adv.distance.value == np.inf
def test_run_sequential(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs1 = attack(bn_images, bn_labels)
advs2 = run_sequential(Attack, bn_model, bn_criterion, bn_images, bn_labels)
advs2 = np.stack([a.perturbed for a in advs2])
assert np.all(advs1 == advs2)
def test_attack_eps(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images,
bn_labels,
unpack=False,
epsilons=np.linspace(0., 1., 100)[1:])
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_eg(request, bn_model, bn_criterion, bn_images, bn_labels):
GradientEstimator = request.param
gradient_estimator = GradientEstimator(epsilon=0.01)
model = ModelWithEstimatedGradients(bn_model, gradient_estimator)
attack = Attack(model, bn_criterion)
advs = attack(bn_images, bn_labels, unpack=False)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_startingpoint(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
np.random.seed(2)
starting_point = np.random.uniform(
0, 1, size=bn_images[0].shape).astype(bn_images.dtype)
advs = attack(bn_images, bn_labels, unpack=False,
starting_point=starting_point)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_notrans_norot(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images,
bn_labels,
unpack=False,
do_translations=False,
do_rotations=False)
for adv in advs:
assert adv.perturbed is None
assert adv.distance.value == np.inf
def test_attack_impossible(bn_model, bn_impossible_criterion, bn_images,
bn_labels):
attack = Attack(bn_model, bn_impossible_criterion)
advs = attack(bn_images,
bn_labels,
unpack=False,
max_iterations=100,
binary_search_steps=20)
for adv in advs:
assert adv.perturbed is None
assert adv.distance.value == np.inf
def test_targeted_attack(bn_model, bn_targeted_criterion, bn_images,
bn_labels):
attack = Attack(bn_model, bn_targeted_criterion)
advs = attack(bn_images,
bn_labels,
unpack=False,
max_iterations=100,
binary_search_steps=20)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_unknown_image(bn_model, bn_criterion, bn_images, bn_labels):
input_images = np.zeros_like(bn_images)
output_images = np.zeros_like(input_images)
attack = Attack(bn_model, bn_criterion)
with pytest.raises(ValueError):
attack(bn_images,
bn_labels,
unpack=False,
candidate_inputs=input_images,
candidate_outputs=output_images)
def test_attack(bn_model, bn_criterion, bn_images, bn_labels):
input_images = bn_images
output_images = np.zeros_like(input_images)
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images,
bn_labels,
unpack=False,
candidate_inputs=input_images,
candidate_outputs=output_images)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_rnd(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images, bn_labels, unpack=False, random_sampling=True)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack(bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images, bn_labels, unpack=False, epsilons=10, steps=5)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_gl(gl_bn_model, bn_criterion, bn_images, bn_labels):
attack = Attack(gl_bn_model, bn_criterion)
advs = attack(bn_images, bn_labels, unpack=False, max_iterations=100)
for adv in advs:
assert adv.perturbed is None
assert adv.distance.value == np.inf
def test_targeted_attack(bn_model, bn_targeted_criterion, bn_images,
bn_labels):
attack = Attack(bn_model, bn_targeted_criterion)
with pytest.raises(AssertionError):
attack(bn_images, bn_labels, unpack=False)
