Core Concepts
Resolving the trade-off between robustness and accuracy in adversarial training through Logit-Oriented Adversarial Training (LOAT).
Abstract
The content discusses the challenges in adversarial training, introduces the Fisher-Rao norm to analyze model complexity, and proposes LOAT as a regularization framework to enhance adversarial training algorithms. It includes empirical evidence, theoretical concepts, and experimental results.
Introduction
Adversarial training aims to improve deep neural network robustness.
Existing methods compromise standard generalization performance.
LOAT proposed to mitigate trade-off between robustness and accuracy.
Related Works
Various factors influence the effectiveness of adversarial training.
Trade-off between robustness and accuracy is a key challenge.
Model complexity plays a crucial role in generalization performance.
Preliminaries
Definition of Rademacher complexity and its relation to model complexity.
Basic notions of image classification tasks and adversarial training objectives.
Proposed Methods
Rademacher complexity analysis via CE loss and Fisher-Rao norm.
Sensitivity of complexity-related factors on generalization gap.
Introduction of Logit-Oriented Adversarial Training (LOAT) framework.
Experiments
Evaluation of LOAT on different models and attacks.
Boost in standard accuracy observed with LOAT.
Comparison of different regularization strategies on model performance.
Stats
"Our code will be available at https://github.com/TrustAI/LOAT."
Quotes
"Can we interpret the degradation of standard accuracy in a unified and principled way?"
"Model complexity offers a potential approach to analyze the trade-off between robustness and accuracy."
"Our extensive experiments demonstrate that the proposed regularization strategy can boost the performance of the prevalent adversarial training algorithms."