Research map / Point clouds / Point cloud classification

Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing

NeurIPS 2026 · Point clouds › Point cloud classification · approach: Adversarial point cloud attacks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PULSAR-Net, a 3D U-Net with axial attention that segments jamming-attack pulses in LiDAR full-waveform data using simultaneous laser sensing, trained only on physics-based synthetic jammed KITTI/nuScenes waveforms, and reports point-level recovery and detector mAP gains including real-world tests.

Problem

LiDAR jamming attacks (high-frequency-repetition laser pulses) randomize point clouds and blind autonomous vehicles, and point-cloud-level defenses cannot recover the original geometry.

Approach

Exploit intermediate full-waveform data and simultaneous multi-laser sensing, where attack pulses appear identically across simultaneously received waveforms, and segment attack pulses with a depthwise-separable 3D U-Net with axial attention, trained on a physics-aware synthetic jamming dataset.

Key results

• On nuScenes-J with 5 simultaneous lasers, PULSAR-Net reaches 86.13% point-level accuracy versus 38.26% for naive average subtraction.
• Real-world static scenes: PULSAR-Net restores mAP to 49.8/69.4/49.8 (PointPillars/CenterPoint/PV-RCNN) from near 0 under attack, with 91.7% vehicle reconstruction accuracy (CenterPoint).
• Real-world driving scenes: PULSAR-Net achieves 73.2% vehicle reconstruction accuracy and 50.7/63.1/53.8 mAP, versus 0.70/0.20/0.00 without defense; runs at 31.74 FPS on RTX 5090.

Surprising findings

• Naive average subtraction produces many false positives and fails to remove attack pulses reliably, while Neural DSP collapses under dense jamming peaks.
• Model trained on nuScenes-J transfers to KITTI-J with 81.73% point accuracy, whereas KITTI-J-trained model drops to 58.81% on nuScenes-J.

Limitations

• No false-positive analysis in driving scenarios; the model may introduce ghost objects.
• Relies on full-waveform access and simultaneous laser sensing, which not all LiDARs provide.
• Assumes attackers cannot selectively inject different patterns into simultaneously sensed receivers.

Experimental setup

Datasets: KITTI nuScenes

Benchmarks: KITTI-J nuScenes-J real-world jamming case study (self-collected LiDAR, 700 frames, 3,106 annotated vehicles)

Metrics: point-level recovery accuracy object detection mAP FPS segmentation IoU

Baselines: Naive Average Subtraction Neural DSP No Defense

Builds on: Neural DSP DW-3D HFR attack PointPillars CenterPoint

Closest papers by meaning

• A New Adversarial Perspective for LiDAR-based 3D Object Detection (AAAI 2025 · similarity 0.83)
• Enduring, Efficient and Robust Trajectory Prediction Attack in Autonomous Driving via Optimization-Driven Multi-Frame Perturbation Framework (CVPR 2025 · similarity 0.83)
• TULIP: Transformer for Upsampling of LiDAR Point Clouds (CVPR 2024 · similarity 0.83)
• GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.82)
• SwiftPillars: High-Efficiency Pillar Encoder for Lidar-Based 3D Detection (AAAI 2024 · similarity 0.82)
• Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · similarity 0.82)