Analysis of the Effect of Feature Denoising from the Perspective of Corruption Robustness

Adversarial attack is a method that aims to cause incorrect predictions in a deep learning model by making slight perturbations to the input. As a result of this vulnerability, various studies have been conducted to improve adversarial robustness. However, deep learning models are also vulnerable to distribution mismatch between training data and test data. This mismatch can occur due to natural corruption in test data. Research on corruption robustness has been less explored compared to adversarial robustness. This paper analyzes the effect of feature denoising network, which is for improving adversarial robustness, in terms of corruption robustness. Experimental results show that feature denoising network can also lead to improved robustness against common corruptions.