Shadow-Analyzer: An Efficient Neural Networks-Based Ghost Objects Detection for Autonomous Vehicles

Several studies have been conducted to investigate the security and reliability of object detection systems in Autonomous Vehicles (AVs), which rely on sensors such as cameras and Light Detection and Ranging (LiDAR). These studies demonstrate the low hurdle for adversaries to execute spoof attacks on LiDAR signals, deceiving non-existing objects (ghosts) as real objects in the surroundings. However, existing approaches to detect such attacks primarily depend on 3D point clouds to analyze LiDAR signals and decide whether an object is real or spoofed, thereby requiring additional processed data. Since decisions in AVs must be made in real-time to ensure safe driving and minimize the risk of exposing road users to danger, reducing the processed data without negatively impacting the reliability and precision of object detection is a significant factor in meeting real-time requirements. Furthermore, reducing the required data to achieve reliable and accurate detection enables efficient edge data processing, thus optimally utilizing available computing power in proximity. To this end, this paper introduces Shadow-Analyzer, a technique that leverages a reduced 2D dataset derived from the 3D point cloud to detect ghost object attacks. The results obtained indicate that the Shadow-Analyzer effectively identifies ghost objects with an accuracy of up to 99.17%, achieving this within an average processing time of 5.9 milliseconds. These findings suggest that our solution fulfills the criteria for real-time detection and significantly improves the standard performance of AVs.