Date of Award

Spring 5-5-2026

Document Type

Thesis

Department

Computer and Information Sciences

Language

English

First Advisor

Panneer Selvam Santhalingam

Abstract

Safe autonomous vehicle (AV) operation depends on robust perception of the surrounding environment. While multimodal sensing—integrating cameras, LiDAR, and radar—provides comprehensive environmental awareness, radar’s robustness under adverse conditions has made it a critical component of modern perception pipelines. However, the security vulnerabilities of radar within these fusion architectures remain largely unexplored. This work presents an end-to-end analysis of radar spoofing attacks on learning-based radar–camera fusion systems. Using a simulation framework grounded in reflect-array attack models, we inject physically plausible perturbations into radar measurements by altering depth by and velocity, while maintaining crossmodal consistency. Evaluation on the nuScenes dataset shows that state-of-the-art architectures — CenterFusion and CRN are highly susceptible to such attacks. Both quantitative and qualitative analyses reveal a significant impact on perception performance, characterized by shifts in 3D bounding box centroids, missed detections, and spatial inconsistencies.

Share

COinS