Multiclass Detection of LiDAR-Based Perception Attacks in Autonomous Vehicles Using CARLA Simulation

Abstract

The growing deployment of Autonomous Vehicles (AVs) has intensified reliance on Internet of Things (IoT) sensors for managing system complexity. Among these, LiDAR plays a central role in accurate environmental perception, but its critical function also exposes it to adversarial interference. Existing approaches have largely focused on attack detection without providing a fine-grained classification of specific threat types.This thesis investigates LiDAR-based perception attacks, targeting the integrity of the LiDAR, using the high-fidelity CARLA simulation platform. A multiclass classification framework is proposed to detect and distinguish between multiple integrity-focused attack types, including False Injection, Occlusion, Obstacle Suppression, Moving Obstacles, and Clustering Modification. Each attack was systematically modeled to generate diverse and reproducible synthetic LiDAR datasets within CARLA. These scenarios were designed to specifically target the integrity of LiDAR data. Three neural network architectures—DNN, CNN1D, and RNN—were trained and tested on the generated datasets. The classification accuracies reached 95% for DNN, 86% for CNN1D, and 98% for RNN, demonstrating the framework’s ability to accurately identify distinct attack scenarios. This capability supports more robust and targeted defensive strategies. The main contribution of this work lies in combining CARLA-based simulation for integrity attack scenario generation with a machine-learning framework capable of reliable multiclass classification.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering