DART3D: Depth-Aware Robust Adversarial Training for Monocular 3D Object Detection

Abstract

Monocular 3D object detection plays a pivotal role in the field of autonomous driving and numerous deep learning-based methods have made significant breakthroughs in this area. Despite the advancements in detection accuracy and efficiency, these models tend to fail when faced with adversarial attacks, rendering them ineffective. Therefore, bolstering the adversarial robustness of 3D detection models has become a critical issue. To mitigate this issue, we propose a depth-aware robust adversarial training method for monocular 3D object detection, dubbed DART3D. Specifically, we first design an adversarial attack that iteratively degrades the 2D and 3D perception capabilities of 3D object detection models (iterative deterioration of perception), serving as the foundation for our subsequent defense mechanism. In response to this attack, we propose an uncertainty-based residual learning method for adversarial training. Our adversarial training leverages inherent uncertainty to boost robustness against attacks while incorporating depth-aware information enhances resistance to perturbations in both 2D and 3D domains. We conducted extensive experiments on the KITTI 3D dataset, showing that DART3D outperforms direct adversarial training in 3D object detection (Formula presented.) for the car category, with improvements of 4.415%, 4.112% and 3.195% in easy, moderate and hard settings, respectively.