CAT: Customized Adversarial Training for Improved Robustness

TL;DR

This paper introduces Customized Adversarial Training (CAT), an adaptive method that improves neural network robustness by customizing perturbation levels and labels for each sample, leading to better accuracy.

Contribution

The paper presents a novel adaptive adversarial training algorithm that enhances robustness and generalization over existing methods.

Findings

Achieves higher clean and robust accuracy than previous methods

Effectively customizes perturbation levels and labels per sample

Demonstrates superior performance through extensive experiments

Abstract

Adversarial training has become one of the most effective methods for improving robustness of neural networks. However, it often suffers from poor generalization on both clean and perturbed data. In this paper, we propose a new algorithm, named Customized Adversarial Training (CAT), which adaptively customizes the perturbation level and the corresponding label for each training sample in adversarial training. We show that the proposed algorithm achieves better clean and robust accuracy than previous adversarial training methods through extensive experiments.