ReAL: Reflective Attack Detection for LPLiDARometry

In this project, we present ReAL, a machine learning-based detection framework for identifying reflective surface attacks against LiDAR sensing systems in real time. Developed as part of the paper "ReAL: Machine Learning Detection of Reflective Attacks against Lidarometry", published in IEEE SoutheastCon 2025, this work addresses a critical security challenge in perception systems that rely on LiDAR measurements for environmental awareness. Reflective surfaces can distort LiDAR returns and introduce misleading sensor readings, which may compromise downstream perception and decision-making in autonomous and edge-based systems.

Our approach focuses on building a lightweight and effective detection pipeline that can operate on resource-constrained edge hardware, including the Jetson Orin platform. Using data collected from an RPLiDAR A1M8-R6 sensor, we constructed multiple experimental scenarios involving different object placements, reflective conditions, distances, and angular configurations to model a wide range of reflective interference behaviors. The resulting dataset was used to train and evaluate machine learning models capable of distinguishing normal LiDAR observations from reflective attack conditions with high confidence and low inference latency.

The system was evaluated across three increasingly complex scenarios. In Scenario 1, the model achieved an inference accuracy of 92.71% with an F1-score of 92.70 and latency of 2.763 ms. In Scenario 2, performance improved to 95.53% accuracy and 95.52 F1-score with 4.727 ms latency. In Scenario 3, which simulated more realistic multi-object reflective conditions, the detector reached 99.97% accuracy, 99.97 F1-score, and only 0.083 ms latency. These results demonstrate that the proposed defense can detect reflective attacks both accurately and efficiently, making it practical for deployment in real-world edge AI perception systems.

Implementation of entire project can be found here: Code