A Robust Defense Mechanism Against Adversarial Attacks in Maritime Autonomous Ship Using GMVAE+RL

Abstract

In this paper, we propose a robust defense frame-work combining Gaussian Mixture Variational Autoencoders (GMVAE) with Reinforcement Learning (RL) to counter adversarial attacks in Maritime Autonomous Systems, specifically targeting the Singapore Maritime Database. By modeling complex maritime data distributions through GMVAE and dynamically adapting decision boundaries via RL, our approach establishes a resilient latent representation space that effectively identifies and mitigates adversarial perturbations. Experimental evaluations using adversarial methods such as FGSM, IFGSM, DeepFool, and Carlini-Wagner attacks demonstrate that the proposed GMVAE+RL model outperforms traditional defenses in both accuracy and robustness. Specifically, it achieves a peak accuracy of 87% and robustness of 20.5%, compared to 85.8% and 19.2%for FGSM and significantly lower values for other methods. These results underscore the superiority of our method in ensuring data integrity and operational reliability within complex maritime environments facing evolving cyber threats.