Research Projects

Design a new adversarial attack framework specifically for LiDAR-based 3D object detection models, using Grad-CAM-generated attention maps to identify key regions to perturb. This targeted approach will maximize model misclassification while preserving the integrity of the input data. Investigate how current architectures like PointNet and 3DETR react to adversarial perturbations. The study will assess their weaknesses and potential failure modes in safety-critical environments, such as autonomous driving, to understand the scope and severity of adversarial threats. Integrate adversarial examples into the training process to enhance model resilience. This will help mitigate the effect of adversarial noise, improving the overall robustness of 3D object detection models in challenging scenarios. Apply Bayesian Neural Networks (BNNs) to model uncertainty in predictions, accounting for both adversarial perturbations and natural environmental variations. This will contribute to more reliable decision-making in uncertain conditions, a critical aspect of autonomous systems.