A Geometrical Approach to Evaluate the Adversarial Robustness of Deep Neural Networks