A Dual-Branch Approach to Misbehavior Detection in VANETs Using Kinematic and Relational Evidence
MSc Thesis Defense by: Amitoj Birah
Date: Wednesday, September 9th , 2026
Time: 11am
Location: Dillon Hall (DH253)
Abstract:
Basic Safety Messages allow cars to share their position and movements, but their contents are self-reported. Cryptographic authentication confirms the sender’s credentials, but does not certify the message’s accuracy. An authenticated insider can send well-formed communications with manipulated position and motion. This thesis proposes a misbehaviour detection mechanism for content falsification. The mechanism is label-free, learning from benign behaviour rather than a fixed attack catalogue; dual-perspective, drawing on two independent lines of evidence to expose a falsification that is invisible to one but visible to the other. It is also receiver-centric, using information available at the receiving vehicle. Proposed mechanism is designed using dual-branch denoising autoencoder. A relational branch assesses its contextual plausibility against physical and neighbourhood evidence whereas a kinematic branch evaluates a sender’s reported motion for self consistency. Each branch is trained on benign traffic, scored by reconstruction error, calibrated against held-out benign data, and aggregated at the decision level using a shared allowance to limit overall benign false-alarm rates. Evaluated on VeReMi Extension data across urban and highway geometries at two traffic densities, and compared against three label-free baselines under a common harness, the mechanism detects content-falsification attacks in its scope at a high operating point across all four scenarios. The advantage is in detecting falsifications that keep the reported trajectory close to benign motion, which are difficult to detect using a single-perspective detector. Branch-isolation results indicate that the two branches are almost fully complimentary, covering attacks that the other cannot.
Keywords: misbehavior detection, VANETs, BSM, unsupervised machine learning, V2V
Thesis Committee:
Internal Reader: Dr. Alioune Ngom
Internal Reader: Dr. Arunita Jaekel
Advisor: Dr. Ikjot Saini
Chair: Asish Mukhopadhyay
