Seeing Is Deceiving:Mirror-Based Lidar Spoofing for Autonomous Vehicle Deception (arxiv.org)

Researchers demonstrated a novel, low-cost attack that uses ordinary planar mirrors to spoof LiDAR-based perception on autonomous vehicles. By redirecting LiDAR beams via geometric optics, the passive mirrors can either create phantom obstacles (Object Addition Attacks) or hide real ones (Object Removal Attacks) without electronics or custom emitters. The team built analytic geometric models, validated them in outdoor experiments with a commercial LiDAR and an Autoware-equipped vehicle, and scaled testing in CARLA. Results show mirror placements can corrupt occupancy grids, produce false detections, and provoke unsafe planning and control behaviors in real driving stacks. This work is significant because it expands the threat model for AV sensing: attacks no longer require active lasers or complex hardware — simple reflective surfaces in the environment suffice. Key technical implications include the need to account for specular reflections in LiDAR point-cloud processing, limitations of single-modality trust, and the brittleness of current fusion pipelines. Proposed mitigations (thermal cameras, robust multi-sensor fusion, and light-fingerprinting) have tradeoffs and gaps, so the paper calls for new sensor verification, geometry-aware filtering, and adversarially informed training to harden AV stacks against inexpensive, physically-deployable spoofing.