How to use PoisoningAttackBackdoor

[EvilWatermelon]

I want to use the backdoor attack against my own dataset. The notebook only shows a variant with load_mnist(). This didn't really help because everything depends on ART functions.

[beat-buesser]

Hi @EvilWatermelon Could you please describe the issue in more detail?

[EvilWatermelon]

For my masterthesis I implement a framework which measure risks of ML models. One of my goals is to attack a ML model and then measure the differences. I want to use ART because there are already implemented backdoor attacks but I have issues to use the attacks of ART.

I tried to implement the attacks with the notebooks and examples. But the examples load training data from special functions load_mnist() etc.

Can I use the attack functions with my own training data or is only possible to load training data from the ART functions?

[beat-buesser]

That sounds interesting! You can use any image dateset, you don't have to use ART's utility functions to load data. The attacks usually expect images x to be of shape NHWC or NCHW, depending on what your estimator/model is expecting as input shape, as type np.ndarray. Let us know if you have more questions!