Any proposed methods of attack mitigation comes to mind after this research? I would suspect, and take this with a grain of salt, I haven't been working with machine learning for quite some time, that smaller subsets of the same data each representing a different kind of attack mapped to another dimension in the network output (that would be then recurrently fed to the same network) could prevent the issue. Kind of teaching the network to defend itself, recognising patterns of attacks and discarding or compensating them as it would noise or other kind of interventions.

Are the attacks statistically identifiable? Can they be translated to a training subset? What would you propose?