Fix SOREL model about secml_malware

[an9236868]

hi @zangobot did you successfully run this on SOREL for goodware/malware class? All of the attacking methods that I run seem to require attributes from MalConv networks such as below, so I'm not sure what is the purpose of CClassifierrSorel

---> 52                 self.embedding_size = end2end_model.get_embedding_size()
     53                 self.max_input_length = end2end_model.get_input_max_length()
     54                 self.invalid_pos = end2end_model.get_embedding_value()

AttributeError: 'CClassifierSorel' object has no attribute 'get_embedding_size'

[zangobot]

Hello!

The classifier uses a different feature set (while malconv is end to end), hence these attacks (that alter single bytes) are not suitable to attack Sorel.
Otherwise, you would need to compute a gradient from features to single bytes (and this is a very hard problem, since you are trying to inverse the mapping).
I advise using a blackbox attack against it!

[an9236868]

thanks @zangobot for your responses, I would be really appreciated it if you can provide me with some of the black-box attacks that you know that can work with SOREL.

[zangobot]

As it is now, the way for using the SOREL model is by creating a c_wrapper_phi that only consider the score you wish to decrease, and you're done!
If you code that, feel free to open a pull request.
I'll put this into my todo list in the meantime.

[zangobot]

Fixed, pull the new version to use it!