def test_fast_gradient_method():
"""
Fast gradient method unit test.
"""
input_np = np.asarray([[0.1, 0.2, 0.7]], np.float32)
label = np.asarray([2], np.int32)
label = np.eye(3)[label].astype(np.float32)
attack = FastGradientMethod(Net())
ms_adv_x = attack.generate(input_np, label)
assert np.any(ms_adv_x != input_np), 'Fast gradient method: generate value' \
' must not be equal to original value.'
def test_batch_generate_attack():
"""
Attack with batch-generate.
"""
input_np = np.random.random((128, 10)).astype(np.float32)
label = np.random.randint(0, 10, 128).astype(np.int32)
label = np.eye(10)[label].astype(np.float32)
attack = FastGradientMethod(Net())
ms_adv_x = attack.batch_generate(input_np, label, batch_size=32)
assert np.any(ms_adv_x != input_np), 'Fast gradient method: generate value' \
' must not be equal to original value.'
def test_fast_gradient_method_gpu():
"""
Fast gradient method unit test.
"""
context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
input_np = np.asarray([[0.1, 0.2, 0.7]], np.float32)
label = np.asarray([2], np.int32)
label = np.eye(3)[label].astype(np.float32)
attack = FastGradientMethod(Net())
ms_adv_x = attack.generate(input_np, label)
assert np.any(ms_adv_x != input_np), 'Fast gradient method: generate value' \
' must not be equal to original value.'
def test_fast_gradient_method_cpu():
"""
Fast gradient method unit test.
"""
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
input_np = np.asarray([[0.1, 0.2, 0.7]], np.float32)
label = np.asarray([2], np.int32)
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
attack = FastGradientMethod(Net(), loss_fn=loss)
ms_adv_x = attack.generate(input_np, label)
assert np.any(ms_adv_x != input_np), 'Fast gradient method: generate value' \
' must not be equal to original value.'