Testing Procedure for DeepFool?

[bendoesai]

Hello, I've been using ART for research for a while, but I keep hitting a roadblock when I try to use deepfool (usually via autoattack) on a binary classifier. What needs to be done in order for deepfool to be "tested on a binary classifier with a single output neuron"? Do you just need someone to implement the binary case? I'd love to be able to contribute to a framework I've been using for so long.

[beat-buesser]

Hi @bendoesai Thank you very much for using ART for such a long time! We'd love to have you contribute to ART. Can you wrap your single-output binary classification model with a thin wrapper model class that exposes two outputs?

[bendoesai]

Hi @beat-buesser, thanks for the response. My project involves extending a deepfake detection framework (DeepfakeBench) such that its infrastructure can be used to generate adversarial samples and run adversarial training.

I could absolutely emulate a second output logit as a workaround, but would it not make more sense to simply add the support for the single output neuron case?