Exponential Separation between Two Learning Models and Adversarial Robustness