white-box attack for Preprocessor class

[le000043]

Hello,
Thank you for creating this wonderful software. I see that in the file examples/get_started_inverse_gan.py, the FGSM attack was conducted in a gray-box setting, which means the attacker happens before the defense is applied. So my question is how can I make the attacker be aware of the preprocessor/defender InverseGAN? I suppose that I need to implement a Wrapper class? If so, what functions do I have to override?

[beat-buesser]

Hi @le000043 That's true. You would have to modify the script to pass inverse_gan to argument preprocessing_defences of a new instance of TensorFlowClassifier defined similarly as in line 46. Does that work?