RAMP: Boosting Adversarial Robustness Against Multiple l p Perturbations for Universal Robustness

A T models' multiple-norm robustness (union accuracy) is still low, which is crucial since in the real-world an adversary is not necessarily bounded by a single