Attack for multi-label classification

Is there an attack available for multi-label classification problems? I went through the repo and I could see the ```self.run``` ends with ```ep.crossentropy```. However, ```BCE with logits``` is required for the above-stated problem. Is there an attack or method which can be over-written to implement this loss?

Start by reading the attack path around self.run and the ep.crossentropy call mentioned in the issue. Determine whether the existing method supports replacing that loss with BCE with logits for multi-label classification, and document or test the supported approach; done means a clear multi-label attack path is established.

Written by the indexing model from the issue text.