LiDAR jamming has been a stubborn problem for autonomous driving: attackers flood a sensor with high-frequency laser pulses until authentic returns are drowned out and the point cloud dissolves into noise. New research from Keio University suggests the raw data behind that noise still holds the truth.
The team found that while jamming randomizes the point clouds LiDAR produces, the underlying full-waveform data keeps distinguishable signatures separating attack signals from legitimate returns. They built PULSAR-Net, a U-Net-style network that exploits those signatures, along with simultaneous laser sensing in modern LiDAR systems, to reconstruct authentic point clouds even while an attack is underway.
Because no real-world datasets of jammed full-waveform LiDAR existed, the researchers generated training data with a synthetic-data generator built from physics models. PULSAR-Net trained exclusively on those simulated examples still delivered substantial reconstruction in real-world static and driving scenarios, which the authors highlight as evidence the approach generalizes.
The demo, shown at the VehicleSec 2026 demo session in Baltimore, frames the work as a defense layer rather than a detection-only fix. Reconstruction under jamming keeps downstream perception, planning, and control running when a spoofing attack would otherwise blind the vehicle.
With jamming gear now cheap and documented at security conferences, defensive work that operates on sensor internals rather than just flagging anomalies is a step toward keeping autonomous systems safe in hostile environments.
