A Polarization-Aware Adversarial Attack Method for Deceiving Polarized Vision Systems

To address the vulnerability of deep learning-based polarized vision models to adversarial sample attacks, this paper proposes a polarization-aware adversarial attack (PA³) generation method based on polarization parameter space optimization. The method generates adversarial samples that are imperceptible to the human eye while specifically optimizing the multi-channel characteristics of sky polarization images. By incorporating perturbation constraints into the Carlini & Wagner (C&W) attack objective function, the generated adversarial samples remain visually indistinguishable from the original images yet significantly degrade the polarization vision model's classification accuracy. Experimental results demonstrate that this method achieves a high attack success rate on the sky polarization image dataset, inducing a heading angle error rate exceeding 70% while maintaining a significantly lower perturbation amplitude than existing methods. Furthermore, this paper presents the first systematic study of adversarial samples' impact on sky polarization image processing tasks, introduces a visually concealed adversarial attack strategy, and exposes the polarization vision system's vulnerabilities. Finally, we preliminarily explore defense mechanisms for sky polarization vision systems, highlighting their security risks and offering novel insights for adversarial sample generation and defense.