Demotivate adversarial defense in remote sensing

TL;DR

This paper investigates whether adversarial defenses in remote sensing neural networks also improve their robustness to over-fitting and data variability, finding no clear correlation between adversarial robustness and other robustness types.

Contribution

The study provides empirical evidence that adversarial robustness does not necessarily enhance geographic or over-fitting robustness in remote sensing models.

Findings

Adversarial robustness is uncorrelated with over-fitting robustness.

Adversarial training does not improve resilience to domain change.

Adversarial regularization does not enhance geographic robustness.

Abstract

Convolutional neural networks are currently the state-of-the-art algorithms for many remote sensing applications such as semantic segmentation or object detection. However, these algorithms are extremely sensitive to over-fitting, domain change and adversarial examples specifically designed to fool them. While adversarial attacks are not a threat in most remote sensing applications, one could wonder if strengthening networks to adversarial attacks could also increase their resilience to over-fitting and their ability to deal with the inherent variety of worldwide data. In this work, we study both adversarial retraining and adversarial regularization as adversarial defenses to this purpose. However, we show through several experiments on public remote sensing datasets that adversarial robustness seems uncorrelated to geographic and over-fitting robustness.