def test_targeted_attack(bn_targeted_adversarial):
set_seeds(22)
adv = bn_targeted_adversarial
attack = Attack()
attack(adv, d=1)
assert adv.perturbed is None
assert adv.distance.value == np.inf
def test_attack(bn_adversarial):
set_seeds(22)
adv = bn_adversarial
attack = Attack()
attack(adv, d=1, t=10)
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_attack_gl(gl_bn_adversarial):
set_seeds(22)
adv = gl_bn_adversarial
attack = Attack()
attack(adv, d=1, t=10)
assert adv.image is not None
assert adv.distance.value < np.inf
def test_attack(bn_model, bn_criterion, bn_images, bn_labels):
set_seeds(22)
attack = Attack(bn_model, bn_criterion)
advs = attack(bn_images, bn_labels, unpack=False, d=1, t=20, R=250)
for adv in advs:
assert adv.perturbed is not None
assert adv.distance.value < np.inf
def test_estimate_gradient_wrapper(eg_bn_adversarial, bn_image):
p, ia = eg_bn_adversarial.forward_one(bn_image)
set_seeds(22)
g = eg_bn_adversarial.gradient_one(bn_image)
set_seeds(22)
p2, g2, ia2 = eg_bn_adversarial.forward_and_gradient_one(bn_image)
assert np.all(p == p2)
assert np.all(g == g2)
assert ia == ia2
def test_batched_estimate_gradient_wrapper(eg_bn_model, bn_image, bn_label):
set_seeds(22)
g1 = eg_bn_model.gradient(bn_image[np.newaxis], [bn_label])[0]
set_seeds(22)
g2 = eg_bn_model.gradient_one(bn_image, bn_label)
assert np.allclose(g1, g2)
gs = eg_bn_model.gradient(np.array([bn_image] * 2),
np.array([bn_label] * 2))
assert np.allclose(gs[0], gs[1], atol=1e-1)
def test_ensemble_average_wrapper(sn_bn_model, bn_model, avg_sn_bn_adversarial,
bn_image, bn_label):
p, ia = avg_sn_bn_adversarial.forward_one(bn_image)
set_seeds(22)
g = avg_sn_bn_adversarial.gradient_one(bn_image)
set_seeds(22)
p2, g2, ia2 = avg_sn_bn_adversarial.forward_and_gradient_one(bn_image)
set_seeds(22)
p3, g3, ia3 = avg_sn_bn_adversarial.forward_and_gradient(
np.array([bn_image]))
p3, g3, ia3 = p3[0], g3[0], ia3[0]
assert np.all(p == p2) and np.all(p == p3)
assert np.all(g == g2) and np.all(g == g3)
assert ia == ia2 == ia3
num_classes = 10
test_grad = np.random.rand(num_classes).astype(np.float32)
model = EnsembleAveragedModel(sn_bn_model, ensemble_size=200)
with model:
assert sn_bn_model.num_classes() == model.num_classes()
assert np.allclose(bn_model.forward_one(bn_image),
model.forward_one(bn_image),
atol=1e-3)
assert np.allclose(
bn_model.gradient_one(bn_image, bn_label),
model.gradient_one(bn_image, bn_label),
atol=1e-3,
)
assert np.allclose(
bn_model.backward_one(test_grad, bn_image),
model.backward_one(test_grad, bn_image),
atol=1e-3,
)
assert np.allclose(
bn_model.forward_one(bn_image),
model.forward_and_gradient_one(bn_image, bn_label)[0],
atol=1e-3,
)
assert np.allclose(
bn_model.forward_and_gradient_one(bn_image, bn_label)[1],
model.forward_and_gradient_one(bn_image, bn_label)[1],
atol=1e-3,
)
assert np.allclose(
bn_model.forward_and_gradient(np.array([bn_image]), [bn_label])[0],
model.forward_and_gradient(np.array([bn_image]), [bn_label])[0],
atol=1e-3,
)
assert np.allclose(
bn_model.forward_and_gradient(np.array([bn_image]), [bn_label])[1],
model.forward_and_gradient(np.array([bn_image]), [bn_label])[1],
atol=1e-3,
)
def test_estimate_gradient_wrapper(eg_bn_adversarial, bn_image):
p, ia = eg_bn_adversarial.forward_one(bn_image)
set_seeds(22)
g = eg_bn_adversarial.gradient_one(bn_image)
set_seeds(22)
p2, g2, ia2 = eg_bn_adversarial.forward_and_gradient_one(bn_image)
set_seeds(22)
p3, g3, ia3 = eg_bn_adversarial.forward_and_gradient(np.array([bn_image]))
p3, g3, ia3 = p3[0], g3[0], ia3[0]
assert np.all(p == p2) and np.all(p == p3)
assert np.all(g == g2) and np.all(g == g3)
assert ia == ia2 == ia3