Towards an Intrinsic Definition of Robustness for a Classifier