Reproduce “Robust Local Features for Improving the Generalization of Adversarial Training” on OpenResearch

About this paper

Chuanbiao Song, Kun He, Jiadong Lin, Liwei Wang, John E. Hopcroft

This paper introduces Robust Local Features for Adversarial Training (RLFAT), a novel framework that improves both the adversarial robustness and standard generalization of deep learning models. The method trains models to learn robust local features by using randomly shuffled adversarial examples and then transferring these features, achieving consistent accuracy gains across various datasets and attack types.

Key insights

• Adversarial training biases models towards global structure features, hindering robust generalization.
• Learning robust local features can bridge the robust generalization gap in adversarially trained models, improving both standard and adversarial accuracy.
• Explicitly training models with random block shuffled adversarial examples and transferring these features encourages reliance on robust local cues.

Summary by alphaXiv.