RAFS-Net: A Robust Adversarial Fusion Framework for Enhanced Maritime Surveillance in Hostile Environments

Deep learning-based intelligent ship surveillance technology has become an indispensable component of modern maritime intelligent perception, with its adversarial defense capabilities serving as a crucial guarantee for reliable and stable monitoring. However, current research on deep learning-based ship surveillance primarily focuses on minimizing the discrepancy between predicted labels and ground truth labels, overlooking the equal importance of enhancing defense capabilities in the adversarial technology-laden maritime environment. To address this challenge and improve model robustness and stability, this study proposes a novel framework termed the Robust Adversarial Fusion Surveillance Net Framework (RAFS-Net). Utilizing ResNet as the backbone network foundation, the framework constructs a ship adversarial attack chain through an adversarial generation module. An adversarial training module enables the model to comprehensively learn adversarial perturbation features. These dual modules effectively rectify abnormal decision boundaries via a synergistic mechanism, compelling the model to learn robust feature representations resilient to malicious interference. Experimental results demonstrate that the framework maintains stable and efficient detection capabilities even in marine environments saturated with interfering information. By systematically integrating gradient-driven adversarial sample generation and an end-to-end training mechanism, it achieves a performance breakthrough of 9.1% in mean Average Precision (mAP) on the ship adversarial benchmark dataset, providing technical support for maritime surveillance models in complex adversarial environments.