moo_hax

The BlackboxClassifer class takes an arbitrary predict function. You should usually return a list of probabilities. But for regression models, I don't know ART has attacks for them. Perhaps integrating...

@abigailgold , updated my comment above. You effectively use Optuna (sampler) as an attack. I don't know if this is in the "literature", but I use it to optimize regular...

I use 2 classes to represent a threshold - a return value is either above or below the threshold. But the `BlackboxClassifier` doesn't check `nb_classes`, any checks on a required...

@beat-buesser is right (obviously), we shouldn't use the `BlackboxClassifier` because of the name, and making all the attacks available. I think a plain "BlackboxModel" could also be a solution. @abigailgold...

Though, I think the Optuna thing might be unpublished. I put together some research scripts to compare it against other evasion attacks. Should probably confirm our intuitions are correct 😁

We use an overridable function in the estimator (our target) https://github.com/Azure/counterfit/blob/dae55c29e9f27ac5d9a99a280bdc23f2a4b26bbd/counterfit/core/targets.py#L179 It's not perfect, and it only works for evasion.

Perhaps a `functools.partial`?