Counterfactual Image Generation for adversarially robust and interpretable Classifiers