Can standard training with clean images outperform adversarial one in robust accuracy?

no code implementations • 29 Sep 2021 • Jing Wang, Jiahao Hu, Guanrong Li

Thus the perturbations carefully generated by the attacker can be diminished.