Boundary attack in targeted=True setting

[aliotopal]

When I test boundary attack in target=True settings and provide the target label like this:
boundary = BoundaryAttack(estimator= kclassifier,
batch_size= 64,
targeted= True,
delta= 0.01,
epsilon = 0.01,
step_adapt = 0.667,
max_iter = 5000,
num_trial = 25,
sample_size = 20,
init_size = 100,
min_epsilon= 0.0,
verbose = True)
img = load_img("acorn.JPEG", target_size=(224, 224), interpolation='lanczos')
img = img_to_array(img)
img = (np.expand_dims(img, axis=0))
ct = np.array([306])
adv_img = boundary.generate(img, y)

It does not iterate like in target=False setting. This is the only verbose I got:

Boundary attack: 100%|██████████| 1/1 [00:09<00:00, 9.96s/it]

then the attack finishes.

Am I doing something wrong.

[beat-buesser]

Hi @aliotopal Please apologies my delayed response. What is the true label of the attacked image and what is the classification prediction of the model for that image?

[aliotopal]

Hi, true label of the image is acorn, and it is classified as acorn with the classification prediction of the model.

[aliotopal]

When it is untargeted we can see that it iterates till 5000 iteration, but when it is targeted it stops in first iterations as shown in verbose: Boundary attack: 100%|██████████| 1/1 [00:09<00:00, 9.96s/it]

[aliotopal]

and the generated adversarial image is not adversarial, it is still in true label for the model we attacked.

[beat-buesser]

@aliotopal Sorry for the delay. Are the true label and the target label identical for these samples?