ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar Object Detection

Robust object detection in side-scan sonar (SSS) is challenging due to the impact of adversarial perturbations and environmental noise, which limit reliable deployment in autonomous underwater vehicles (AUVs). While existing approaches primarily focus on denoising techniques and data augmentation, they do not explicitly enhance the model’s robustness against adversarial perturbations. To overcome these limitations, we introduce ROSAR, a framework that extends our prior work on knowledge distillation (KD) by integrating adversarial retraining, jointly tackling the dual challenges of model robustness and effi-ciency in SSS object detection. To support adversarial retraining and robustness evaluation, we introduce three novel, publicly available SSS datasets, each capturing different sonar setups and noise conditions. Furthermore, we define two SSS safety properties that guide the generation of adversarial datasets for retraining. Through a comparative analysis of projected gradient descent (PGD) and patch-based adversarial attacks, ROSAR demonstrates significant improvements in model robustness and detection accuracy under SSS-specific conditions, enhancing the model’s robustness by up to 1.85%. ROSAR is publicly available at https://github.com/remaro-network/ROSAR-framework.