HopskipJump Attack

[fukashi-hatake]

Hello @beat-buesser,
Thank you for your wonderful effort to ART.

I have some issues to use HopskipJump attack with targeted label on CIFAR 10. Can you give me some example how to use it. While generating I am doing as follow:

hopskipjump_attack = HopSkipJump(classifier=classifier,  verbose=False, targeted=True)   
hopskipjump_attack.generate(x=input_batch, y = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 1]) )  

But it failed to succeed.

Thank you beforehand!

[beat-buesser]

Hi @fukashi-hatake

Thank you very much!

How does it fail? Is the attack not able to find an adversarial example or does the code return an error message?

[fukashi-hatake]

Is the label provided to y in the method generate the true or target label?
Answer: it is the target label

Does classifier.predict(x=input_batch)predict the correct class/true label?
Answer: yes

[fukashi-hatake]

I just want to know that is it correct way to define target label?
hopskipjump_attack.generate(x=input_batch, y = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 1]) )

[beat-buesser]

If input_batch contains a single sample it should work, to ensure correct shapes you could make y two-dimensional with np.array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 1]]).

Are you using the default parameters of HopSkipJump?

[fukashi-hatake]

Are you using the default parameters of HopSkipJump?
Answer: yes (only targeted=True)

[beat-buesser]

It could be that the initialisation step of HopSkipJump is not able to sample a random starting perturbation that makes the sample adversarial. Have you tried increasing the argument init_size to allow for more samples in the initialisation step?

Is the attack successful in the untargeted case for this sample (with true label for y in generate)?