Advancing SAR Image Robustness: Integrating Diffusion Models for Adversarial Purification and Speckle Noise Suppression

Abstract

Synthetic Aperture Radar (SAR) is a powerful tool for ground target detection, but SAR images often suffer from coherent speckle noise, which complicates automatic target recognition (ATR). Recent advances in deep learning have shown promise in SAR-ATR, yet there is a notable gap in addressing adversarial attacks and despeckling challenges. This study presents SAR Adversarial Diffusion Purification (SAR-ADP), a novel model designed to enhance SAR image analysis by tackling both adversarial attacks and coherent speckle noise. Our main contributions include pioneering methods for SAR image adversarial defense, simultaneous adversarial noise removal and despeckling, guidance mechanisms within the diffusion model, and a streamlined architecture for SAR image purification. SAR-ADP integrates adversarial defense strategies, diffusion modeling, and despeckling techniques, resulting in significant improvements in image recognition accuracy and post-processing. Empirical results validate its effectiveness, achieving a robustness accuracy of approximately 97% and a standard accuracy of around 99%.