Description
AdvProp is an adversarial training scheme which treats adversarial examples as additional examples, to prevent overfitting. Key to the method is the usage of a separate auxiliary batch norm for adversarial examples, as they have different underlying distributions to normal examples.
Papers Using This Method
Improving Model Generalization by On-manifold Adversarial Augmentation in the Frequency Domain2023-02-28How explainable are adversarially-robust CNNs?2022-05-25Fast AdvProp2022-04-21Pyramid Adversarial Training Improves ViT Performance2021-11-30Advanced Graph and Sequence Neural Networks for Molecular Property Prediction and Drug Discovery2020-12-02Adversarial Examples Improve Image Recognition2019-11-21