How Does Frequency Bias Affect the Robustness of Neural Image Classifiers against Common Corruption and Adversarial Perturbations?

TL;DR

This paper investigates how frequency bias in neural image classifiers impacts their robustness against various corruptions and adversarial attacks, proposing Jacobian frequency regularization to improve robustness by emphasizing low-frequency features.

Contribution

It introduces Jacobian frequency regularization to control frequency bias in models, linking frequency domain features to robustness improvements against specific corruptions.

Findings

Biasing towards low-frequency features improves robustness against high-frequency corruption.

Biasing towards high-frequency features enhances robustness against low-frequency corruption.

Tradeoffs exist: improving robustness against one type of corruption may reduce performance against another.

Abstract

Model robustness is vital for the reliable deployment of machine learning models in real-world applications. Recent studies have shown that data augmentation can result in model over-relying on features in the low-frequency domain, sacrificing performance against low-frequency corruptions, highlighting a connection between frequency and robustness. Here, we take one step further to more directly study the frequency bias of a model through the lens of its Jacobians and its implication to model robustness. To achieve this, we propose Jacobian frequency regularization for models' Jacobians to have a larger ratio of low-frequency components. Through experiments on four image datasets, we show that biasing classifiers towards low (high)-frequency components can bring performance gain against high (low)-frequency corruption and adversarial perturbation, albeit with a tradeoff in performance for low (high)-frequency corruption. Our approach elucidates a more direct connection between the frequency bias and robustness of deep learning models.