LOT: Layer-wiseOrthogonalTrainingonImproving ℓ2CertifiedRobustness

Neural Information Processing Systems 

Under the supervised scenario, we improve the state-of-the-art certified robustness for allarchitectures(e.g.