Towards Adversarial Robustness of SAR ATR via GANs and Deep Learning

Abstract

Deep learning has significantly advanced Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR), yet these systems remain critically vulnerable to adversarial attacks. In this paper, a comprehensive review of the state-of-the-art in adversarial robustness for SAR ATR is provided, synthesizing key studies from 2016 to 2025. The analysis covers the landscape of threat models and defenses, with a focus on the role of Generative Adversarial Networks (GANs) and hybrid architectures. The paper further analyzes hybrid architectures that combine GANs with normalizing flows and show how these yield substantial improvements in adversarial and out-of-distribution (OOD) robustness. While effective, these models incur significant computational overhead, including increased GPU memory use, inference latency, and training cost, posing constraints for real-time edge deployment. This review consolidates the primary challenges, identifies key research gaps, and concludes that future progress depends on developing certified, physics-aware, and computationally efficient defense mechanisms to enable secure real-world deployment.