示例#1
0
def test_sub_noise(art_warning, image_dl_estimator, fix_get_mnist_subset):
    try:
        (x_train_mnist, y_train_mnist, x_test_mnist,
         y_test_mnist) = fix_get_mnist_subset

        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=4000,
                       verbose=False)
        c = attack._sub_noise(num_noises=64, basis=np.ones((28 * 28, 25)))

        assert c[0].shape == x_train_mnist[0].shape
    except ARTTestException as e:
        art_warning(e)
示例#2
0
def test_is_adversarial(art_warning, image_dl_estimator, fix_get_mnist_subset):
    try:
        (x_train_mnist, y_train_mnist, x_test_mnist,
         y_test_mnist) = fix_get_mnist_subset

        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=4000,
                       verbose=False)

        assert not attack._is_adversarial(x_adv=x_test_mnist[[0]],
                                          y_true=y_test_mnist[[0]])

    except ARTTestException as e:
        art_warning(e)
示例#3
0
def test_generate(art_warning, fix_get_mnist_subset, image_dl_estimator):
    try:
        (x_train_mnist, y_train_mnist, x_test_mnist,
         y_test_mnist) = fix_get_mnist_subset

        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=400,
                       verbose=False)
        x_train_mnist_adv = attack.generate(x=x_train_mnist, y=y_train_mnist)

        assert np.mean(np.abs(x_train_mnist_adv -
                              x_train_mnist)) == pytest.approx(0.057965796,
                                                               abs=0.01)
    except ARTTestException as e:
        art_warning(e)
示例#4
0
def test_opt_query_iteration(art_warning, image_dl_estimator):
    try:
        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=4000,
                       verbose=False)
        opt_q, var_t = attack._opt_query_iteration(var_nq=3800,
                                                   var_t=8,
                                                   lambda_param=0.6)

        opt_q_expected = [75, 106, 149, 210, 295, 414, 582, 818, 1150]
        var_t_expected = 9

        assert opt_q == opt_q_expected
        assert var_t == var_t_expected
    except ARTTestException as e:
        art_warning(e)
示例#5
0
def test_generate_2d_dct_basis(art_warning, image_dl_estimator):
    try:
        sub_dim = 5
        res = 224

        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=4000,
                       verbose=False)

        dct = attack._generate_2d_dct_basis(sub_dim=sub_dim, res=res)
        assert dct.shape == (res * res, sub_dim * sub_dim)
        dct_50000 = np.array([[
            0.00446429,
            -0.0063133,
            0.00631283,
            -0.00631206,
            0.00631097,
            0.00490833,
            -0.00694126,
            0.00694075,
            -0.00693989,
            0.0069387,
            0.00131841,
            -0.00186447,
            0.00186434,
            -0.00186411,
            0.00186379,
            -0.00285836,
            0.00404223,
            -0.00404193,
            0.00404143,
            -0.00404074,
            -0.00576281,
            0.00814965,
            -0.00814905,
            0.00814805,
            -0.00814664,
        ]])
        np.testing.assert_almost_equal(dct[50000], dct_50000)
    except ARTTestException as e:
        art_warning(e)
示例#6
0
def test_find_random_adversarial(art_warning, image_dl_estimator,
                                 fix_get_mnist_subset):
    try:
        (x_train_mnist, y_train_mnist, x_test_mnist,
         y_test_mnist) = fix_get_mnist_subset

        classifier, _ = image_dl_estimator(from_logits=True)
        attack = GeoDA(estimator=classifier,
                       sub_dim=5,
                       max_iter=4000,
                       verbose=False)

        x_adv = attack._find_random_adversarial(x=x_test_mnist[[0]],
                                                y=y_test_mnist[[0]])

        assert np.argmax(classifier.predict(x_adv)) != np.argmax(
            y_test_mnist[[0]])
    except ARTTestException as e:
        art_warning(e)
示例#7
0
def test_check_params(art_warning, image_dl_estimator_for_attack):
    try:
        classifier = image_dl_estimator_for_attack(GeoDA, from_logits=True)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, batch_size=1.0)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, batch_size=-1)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, norm=0)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, sub_dim=1.0)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, sub_dim=-1)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, max_iter=1.0)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, max_iter=-1)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, max_iter=1.0)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, max_iter=-1)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, bin_search_tol="1")
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, bin_search_tol=-1.0)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, lambda_param=1)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, lambda_param=-1.0)

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, sigma=1)
        with pytest.raises(ValueError):
            _ = GeoDA(classifier, sigma=-1.0)

        # with pytest.raises(ValueError):
        #     _ = GeoDA(classifier, targeted="true")

        with pytest.raises(ValueError):
            _ = GeoDA(classifier, verbose="true")

    except ARTTestException as e:
        art_warning(e)