Answers are grounded in this paper and cite their sources.
LiDAR is an important remote sensing technology for acquiring three-dimensional spatial information, providing data for object identification, localization, and scene understanding in applications such as autonomous driving. As these systems rely on deep learning models, adversarial attacks have become an important tool for evaluating their robustness. Among attack paradigms, transfer-based attacks are particularly practical, crafting adversarial point clouds on accessible surrogate models and transferring them to unknown target models. However, most existing methods either perturb all points under a uniform constraint or concentrate distortion on critical regions, ignoring spatial variations in geometry and sensitivity and often damaging the structures that recognition relies on. Moreover, they target object-level classification and transfer poorly to scene-level detection. Therefore, we propose MIST (Multi-source Imperceptible Saliency-guided Transferable), an adversarial attack framework for LiDAR point cloud perception in remote sensing. MIST fuses source-model gradients, local geometric variation, and predictive uncertainty into a multi-source saliency score that guides perturbation allocation. Guided by this score, Reverse control tightens displacement limits at high-saliency sites, while Intensity scaling modulates update strength to maintain attack effectiveness. Their coordination allocates the budget adaptively by saliency, improving cross-model transferability while preserving geometric imperceptibility. We evaluate MIST through classification on ModelNet40 and ScanObjectNN, and extend to full-scene 3D object detection on KITTI and Waymo under the same transfer-based setting. Results show MIST improves the transferability-fidelity balance in classification and achieves effective cross-model attacks in detection, offering insights for developing more robust perception models in LiDAR-based remote sensing.
No takes yet. Share an insight, caveat, or question.
Ping et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: