Exploiting epistemic uncertainty of the deep learning models to generate adversarial samples