Label Noise in Adversarial Training: A Novel Perspective to Study Robust Overfitting

TL;DR

This paper reveals the presence of label noise in adversarial training caused by mismatched label distributions, providing new insights into robust overfitting and proposing a calibration method to improve model performance.

Contribution

It introduces the concept of label noise in adversarial training, explains its impact on robust overfitting, and proposes a calibration method that enhances performance without extra tuning.

Findings

Label noise exists in adversarial training due to distribution mismatch.

The proposed calibration method improves robustness across models and datasets.

Insights into the dependence of robust overfitting on perturbation radius and data quality.

Abstract

We show that label noise exists in adversarial training. Such label noise is due to the mismatch between the true label distribution of adversarial examples and the label inherited from clean examples - the true label distribution is distorted by the adversarial perturbation, but is neglected by the common practice that inherits labels from clean examples. Recognizing label noise sheds insights on the prevalence of robust overfitting in adversarial training, and explains its intriguing dependence on perturbation radius and data quality. Also, our label noise perspective aligns well with our observations of the epoch-wise double descent in adversarial training. Guided by our analyses, we proposed a method to automatically calibrate the label to address the label noise and robust overfitting. Our method achieves consistent performance improvements across various models and datasets without introducing new hyper-parameters or additional tuning.