Generating Realistic Datasets for VANET Misbehavior Detection
MSc Thesis Defense by: Joshua Picchioni
Date: Friday, August 21st, 2026
Time: 10:00am to 11:30am
Location: Essex Hall 122
Abstract:
Connected vehicles in Vehicular Ad-Hoc Networks (VANETs) use Basic Safety Messages (BSMs) to share information such as position, speed, and heading with nearby vehicles and infrastructure. Although these messages can be authenticated, authentication does not actually guarantee that the information inside them is correct. A compromised vehicle could send a technically valid message containing false position data. Many studies attempt to detect such false information using ML models trained with simulated traffic data. Simulation provides controlled and repeatable conditions, but it may not fully capture the measurement noise and irregular movements found in data collected from vehicles on real roads. This thesis presents a pipeline for creating labelled position-falsification real-world datasets from BSMs collected through the Wyoming Connected Vehicle Pilot. After data cleaning and geographic filtering, the Wyoming dataset is cut down to 12 million messages distributed across three 100 km regions. The pipeline converts the BSM message latitude and longitude values into local Cartesian coordinates, allowing the positional offsets to be applied and measured in metres. Thirty percent of the vehicle IDs are selected as attackers before vehicles are separated into training and testing sets. Two position attacks are then applied. Random Position Offset (RPO) generates a random offset for each malicious vehicle message, and Constant Position Offset (CPO) applies a fixed offset to every single message sent by the vehicles. Each message is also assigned a label saying if it is an attacker or not, and more trajectory features are calculated to create more features to test.
The generated datasets are evaluated using 3 Machine Learning (ML) algorithms. Random Forest, Decision Tree, and K-Nearest Neighbors. These experiments cover the three regions, fourteen displacement ranges, several raw BSM feature sets (no trajectory), four trajectory-based feature sets, and sets both including and excluding the raw vehicle ID. Mean RPO test F1 increases significantly after including Motion Consistency features, but CPO does not as shifting by a constant offset technically maintains a valid trajectory. The results show that the pipeline can produce useful, labelled datasets from real-world BSM data (which can be modifiable to different types of attacks), while also showing that different types of position attacks may require different features and ways of detecting them.
Keywords: Connected Vehicles, Basic Safety Messages, Misbehaviour Detection, Position Falsification, Random Position Offset, Constant Position Offset, Machine Learning, Trajectory Consistency
Thesis Committee:
Internal Reader: Dr. Jianguo Lu
External Reader: Dr. Bala Balasingam
Advisor: Dr. Arunita Jaekel
Chair: Dr. Alioune Ngom
