Towards Adversarial And Backdoor Robustness Of Deep Learning