SAR-PAA: A Physically Adversarial Attack Approach Against SAR Intelligent Target Recognition

Abstract

Synthetic aperture radar (SAR) target recognition has entered a new era of intelligence due to the rapid development of deep learning. Naturally, an accompanying challenge arises in countering SAR intelligent target recognition technology and protecting the targets of interest from exposure risks. In this article, a novel adversarial attack approach against SAR intelligent target recognition is proposed, which is physically easy to implement in practice. Just with strategic deployment of simple scatterers near the target, such as corner reflectors, the SAR intelligent target recognition system will be effectively attacked. First, an adversarial attack framework against SAR intelligent target recognition is constructed based on the SAR imaging process and target recognition mechanism. Then, the corresponding adversarial attack optimization model is established considering the adversarial attack effectiveness, target and environment constraints. Finally, integrating electromagnetic computation and differential evolution algorithm, a solution method for the adversarial attack optimization model is designed to achieve the physical implementation of adversarial attacks against SAR target recognition systems. To comprehensively evaluate the performance of the proposed method, six typical SAR target recognition networks are attacked based on the MSTAR dataset, resulting in an average fooling rate of 81.83%. Extensive experiments demonstrate the effectiveness and superiority of the proposed method.