INTACT: Inducing Noise Tolerance through Adversarial Curriculum Training for LiDAR-Based Safety-Critical Perception and Autonomy

In this work, we present INTACT, a two-phase framework to improve the robustness of deep neural networks (DNNs) against noisy LiDAR data in safety-critical perception tasks. INTACT combines meta-learning with adversarial curriculum training (ACT) to address corruption and sparsity in 3D point clouds. In the meta-learning phase, a teacher network acquires task-agnostic priors to generate saliency maps highlighting critical data regions. In the ACT phase, these maps guide a student network through progressively complex noise, enabling targeted perturbation and stronger noise resilience. Evaluations on KITTI, Argoverse, and ModelNet40 show that INTACT improves robustness by up to 20% across detection, tracking, and classification, outperforming adversarial and curriculum training baselines. On KITTI, INTACT raised Multiple Object Tracking Accuracy (MOTA) by 11% (64.1% → 75.1%) and by 21.2% under Gaussian noise (52.5% → 73.7%). Mean Average Precision (mAP) increased from 59.8% to 69.8% (50% point drop) and from 49.3% to 70.9% (Gaussian noise).