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

TL;DR

ROSAR tackles the challenge of reliable object detection in side-scan sonar imagery under underwater noise by fusing knowledge distillation with adversarial retraining. It formalizes two safety properties, generates adversarial datasets via PGD and patches, and retrains a KD-augmented detector to enhance robustness. The work introduces three public SSS datasets and demonstrates improved robustness and detection performance, particularly under PGD-based retraining, with Patch-SWDD offering stability advantages. This framework advances underwater robotic perception by providing verifiable, robust object detection under challenging sonar conditions.

Abstract

This paper introduces ROSAR, a novel framework enhancing the robustness of deep learning object detection models tailored for side-scan sonar (SSS) images, generated by autonomous underwater vehicles using sonar sensors. By extending our prior work on knowledge distillation (KD), this framework integrates KD with adversarial retraining to address the dual challenges of model efficiency and robustness against SSS noises. We introduce three novel, publicly available SSS datasets, capturing different sonar setups and noise conditions. We propose and formalize two SSS safety properties and utilize them to generate 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 available at https://github.com/remaro-network/ROSAR-framework.

Figures (4)

• Figure 1: The structure of the ROSAR framework. Yellow boxes show the knowledge distillation process from our previous work aubard2024knowledge, while the blue boxes display the adversarial retraining process.
• Figure 2: Samples of SWDD-Clean, SWDD-Surface and SWDD-Noisy.
• Figure 3: Sample images of the three adversarial datasets.
• Figure 4: Robustness validation for $\mathcal{P}_1$ and $\mathcal{P}_2$.