Understanding and combating robust overfitting via input loss landscape analysis and regularization

Adversarial training is widely used to improve the robustness of deep neural networks adversarial attack. However, prone overfitting, and cause far from clear. This work sheds light on mechanisms underlying overfitting through analyzing loss landscape w.r.t. input. We find that robust results standard training, specifically minimization clean loss, can be mitigated by regularization gradients. Moreover, we turns severer during partially because gradient effect becomes weaker due increase in landscapes curvature. To generalization, propose a new regularizer smooth penalizing weighted logits variation along direction. Our method significantly mitigates achieves highest efficiency compared similar previous methods. Code available at https://github.com/TreeLLi/Combating-RO-AdvLC.