An Adversarial Machine Learning Approach for Generating Stealthy BSM Data in V2V Communication
MSc Thesis Proposal by:
Saumya Buch
Date: 21-05-2026
Time: 12:00 pm
Location: 122 Essex Hall
Abstract:
Misbehavior detection systems (MDSs) are vital in determining whether a Basic Safety Message (BSM) message has anomalies or is physically implausible. In this thesis, we propose an adversarial data generation framework using neural networks to simulate evasive attacks against an MDS in Vehicle-to-Vehicle (V2V) communication. The objective is to generate realistic yet adversarially modified BSM data that can bypass detection systems, thereby enabling the study of adversarial attack scenarios and improving MDS in being able to detect adversarial attacks. The results demonstrate that while the pretrained MDS achieves perfect classification performance on the test set, it remains susceptible to adversarial attacks under constrained feature modifications and a fixed detection model. In addition, the framework is able to create perturbations that do not largely change the datasets from their original state but are able to induce an MDS to detect them as benign. These findings highlight methods of exploration that can improve the effectiveness and reliability of vehicular network security solutions.
Thesis Committee:
Internal Reader: Dr. Jessica Chen
Internal Reader: Dr. Boubakeur Boufama
Advisor: Dr. Ikjot Saini
