Curriculum-Guided Adversarial Learning for Enhanced Robustness in 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D object detection is crucial for autonomous systems.
• LiDAR-based detectors like PointPillars face challenges in adversarial robustness and class imbalance.
• Existing methods often lack inherent resilience to sophisticated attacks.

Purpose of the Study:

• To develop a framework that enhances adversarial robustness and detection accuracy for LiDAR-based 3D object detectors.
• To introduce a novel 3D object detector with intrinsic adversarial robustness.
• To address the class imbalance problem in 3D object detection datasets.

Main Methods:

• Proposed a curriculum-guided adversarial learning (CGAL) framework.
• Developed a novel 3D object detector, Pillar-RBFN, integrating a nonlinear enhancement block (NEB) with radial basis function networks.
• Introduced a data augmentation technique (SFGTS) and an adaptive focal loss to create an adversarial dataset (Adv-KITTI) and mitigate class imbalance.

Main Results:

• The CGAL framework improved mean average precision (mAP) by 0.8–2.5 percentage points over conventional training.
• Models trained with Adv-KITTI showed at least a 15 percentage point mAP enhancement.
• Pillar-RBFN demonstrated intrinsic adversarial robustness without adversarial training.

Conclusions:

• The proposed CGAL framework significantly enhances adversarial robustness and detection accuracy for LiDAR-based 3D object detection.
• The novel Pillar-RBFN detector offers inherent resilience to adversarial attacks.
• The developed data augmentation and loss function effectively address class imbalance issues, improving overall model performance.