BENCHMARKING ADVERSARIAL ATTACKS AND DEFENSES IN REMOTE SENSING IMAGES

Deep neural networks are prone to being fooled by adversarial examples, which are created by adding imperceptible perturbations to deceive the classifier and induce incorrect predictions. Several defense methods have been proposed to address this issue. However, little attention has been given to the application of these methods in the field of remote sensing image (RSI) area. In this research, we have developed a comprehensive open-source benchmark that aims to evaluate the effectiveness of adversarial attacks and defenses in the context of RSI scene classification 1. Extensive experiments have been conducted on both SAR and optical image datasets to thoroughly analyze white-box attacks, black-box attacks, and defense scenarios. We anticipate that our benchmark will assist researchers in constructing deep neural networks with enhanced resistance against adversarial attacks in the RSI field.