A Robust Adversarial Defense Algorithm for Enhancing Salient Object Detection in Remote Sensing Image

Deep neural networks (DNNs) have been widely applied to salient object detection in optical remote sensing images (ORSIs-SOD) with significant progress. However, the introduction of adversarial perturbations can severely degrade their detection performance. Moreover, the relatively limited labeled samples in ORSI-SOD make the model more susceptible for overfitting, increasing sensitivity to adversarial perturbations. Therefore, researching defense algorithms against adversarial samples to enhance the robustness of salient object detection (SOD) models in the field of ORSIs is of great importance. However, existing defense studies have seldom focused on the ORSI-SOD task and have not considered the unique challenges it faces to propose effective defense solutions. Thus, this article proposes a robust adversarial defense algorithm specifically for ORSI-SOD for the first time. This algorithm first proposes a fixed luminance channel multimodal data random fusion adversarial defense strategy to enhance the model’s robustness against the color and contour information of targets. Then, the algorithm proposes a sinusoidal noise-driven adversarial feature adaptive masking module, which adaptively reduces the impact of adversarial perturbations and enhances the model’s generalization ability. In addition, the algorithm proposes a global context-aware denoising perception module (GCA-DPM) into the network to strengthen the model’s understanding of the global content of images, further suppressing the impact of adversarial noise and enhancing the detection accuracy. Finally, experimental results demonstrate that the proposed defense algorithm effectively enhances the adversarial robustness of various SOD models. Against three different attack strategies, the algorithm increases the average $F_{\beta }$ of four SOD models by 23.94% and 21.98% on the optical remote sensing saliency detection (ORSSD) and extended ORSSD (EORSSD) datasets, respectively.

I. Introduction

Salient object detection (SOD) [1] is a method that simulates human visual attention mechanisms. It assists machines to automatically identify the most attractive regions or objects within an image. As a crucial preprocessing step, SOD has driven advancements in fields, such as object segmentation [2], image quality assessment [3], and object tracking [4]. In recent years, convolutional neural networks (CNNs), as a widely used deep learning architecture in the field of computer vision, have provided strong support for the further development of SOD tasks.