Perturbation in FastGradientMethod

[umbertogagl97]

Hello,
I am new to ART and have a question about the Fast Gradient Sign attack https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/art/attacks/evasion/fast_gradient.py.
Is there a fast way to obtain the perturbation applied to every image? I'm tring the subtraction adversarial_image - original_image but I have problem with values of the result because they're in [-1,1] and I need to normalize them in [0,1]. Thank you.
I report the code:
attack = FastGradientMethod(estimator=classifier, eps=0.1)
x_test_adv = attack.generate(x=x_test,y=y_test)
perturb=x_test_adv-x_test

[beat-buesser]

Hi @umbertogagl97 Thank you very much for your interest in ART! What is the pixel range of x_test and how do you define classifier?

[umbertogagl97]

I try to explain the problem better. The adversarial images are not the problem, they are in [0,1], but the perturbation (calculated by subtraction) contains value <0. For this reason I would like to know if there is another way to get the perturbation.

[beat-buesser]

Ok, that makes sense. In this case, the adversarial perturbation can be negative, this happens in cases where a pixel had a negative loss gradient and therefore its pixel value decreased.

[umbertogagl97]

Okay. So, is the subtraction the right and only way to calculate the perturbation?

[beat-buesser]

Yes, if you would like to know how much the pixel values have changed. Why do you ask?

[umbertogagl97]

Because I'm just start to working with ART and I don't know yet how to use well the library. Thank you very much.