Scalable Differential Privacy with Certified Robustness in Adversarial Learning