Seeing is Deceiving: Systematic Vulnerability Analysis of LiDAR-Based Autonomous Driving to Mirror-Induced Perception Failures

Autonomous vehicles rely heavily on LiDAR for accurate 3D perception; however, this sensing modality remains vulnerable to several physical-world failure modes. This paper presents a systematic vulnerability analysis of LiDAR-based autonomous driving stacks against mirror-induced perception failures. We identify, formalize and evaluate two distinct failure modes arising from planar mirrors in the environment: Object Addition, where multi-path reflections create phantom obstacles in free space, and Object Removal, where beam deflection renders real obstacles invisible to the sensor. Our analysis makes three contributions. First, we conduct an end-to-end vulnerability assessment on a full-stack autonomous vehicle running Autoware, tracing how mirror-induced artifacts propagate from raw perception through planning and control, ultimately triggering emergency braking and cascade collisions. Second, we derive physics-informed empirical models that characterize artifact location, point count, and appearance probability as functions of distance, angle, and mirror area. Third, we integrate these models into a CARLA-based simulation framework for real-time artifact injection, addressing a key limitation of current simulators that do not model specular reflections. Our findings reveal that even small mirrors (0.18m², costing under $60) can cause high-confidence false detections (74% "Vehicle" classification) and trigger unsafe driving behaviors. We discuss potential countermeasures and highlight that robust defenses against physically legitimate but geometrically deceptive sensor inputs remain an open challenge.