Computational limitations in robust classification and win-win results