Attackers can now fool a self-driving car into seeing a "phantom" object by making its camera and LiDAR sensors agree on a lie.
April 24, 2026
Original Paper
Cross-Modal Phantom: Coordinated Camera-LiDAR Spoofing Against Multi-Sensor Fusion in Autonomous Vehicles
arXiv · 2604.21841
The Takeaway
Self-driving systems rely on sensor fusion to ensure that if one sensor is tricked, the other will provide the truth. This new attack creates cross-modal consistency, where the visual and laser data are synchronized to show the same fake obstacle. This bypasses the safety redundancy that was previously thought to be foolproof. By spoofing both inputs at once, an attacker can force a car to slam on its brakes or swerve into another lane. This proves that hardware redundancy is not a complete defense against a sophisticated, coordinated attack. The industry must develop new ways to verify sensor data that go beyond simple consensus.
From the abstract
Autonomous Vehicles (AVs) increasingly depend on Multi-Sensor Fusion (MSF) to combine complementary modalities such as cameras and LiDAR for robust perception. While this redundancy is intended to safeguard against single-sensor failures, the fusion process itself introduces a subtle and underexplored vulnerability. In this work, we investigate whether an attacker can bypass MSF's redundancy by fabricating cross-sensor consistency, making multiple sensors agree on the same false object. We desig