Advancing Radar Cybersecurity: Defending Against Adversarial Attacks in SAR Ship Recognition Using Explainable AI and Ensemble Learning

This paper investigates the vulnerability of Synthetic Aperture Radar (SAR)-based ship recognition models to adversarial attacks. We employ the Fast Gradient Sign Method (FGSM) to generate adversarial examples, adding imperceptible perturbations to SAR ship images to mislead a pre-trained convolutional neural network (CNN). To analyze the impact of these attacks, we utilize the Local Interpretable Model-agnostic Explanations (LIME) algorithm, an Explainable Artificial Intelligence (XAI) method, to explain the contributing area in the input image to the CNN's decision-making process under adversarial conditions. Finally, we propose an ensemble learning strategy combining multiple transfer learning-based architectures to enhance the robustness of ship recognition systems against adversarial examples and mitigate their transferability. Our real data experiment is conducted on OpenSARShip dataset, which consists of different ship images extracted from 41 images captured by Sentinel-1 SAR satellite.