A Blockchain-Assisted Hybrid AI and LLM Framework for Misbehavior Detection in Connected Autonomous Vehicles

Abstract

Connected Autonomous Vehicles (CAVs) are a key component of Intelligent Transportation Systems (ITS), enabling cooperative perception and automated decision-making through Vehicle-to-Everything (V2X) communication. Despite their benefits, CAVs introduce critical security vulnerabilities at the perception layer. In particular, Non-Data-Dependent (NDD) misbehaviour attacks manipulate physical-world objects, such as traffic signs, without altering transmitted messages, exploiting weaknesses in vision-based perception models and evading conventional detection mechanisms based on message plausibility or consistency. Existing approaches primarily rely on isolated machine learning or deep learning models under centralised trust assumptions, resulting in limited robustness against adversarial visual perturbations, insufficient semantic awareness, and poor explainability. This paper proposes a blockchain-enabled hybrid framework for detecting misbehaviour in image-based NDD attacks in CAV networks. The framework integrates classical machine learning, deep learning, and Large Language Model (LLM)- based reasoning to jointly capture perceptual and semantic anomalies. Convolutional neural networks perform visual feature extraction, while LLMs conduct high-level contextual reasoning over structured metadata. An entropy-based adaptive fusion mechanism produces robust and explainable detection decisions. Detected events are recorded as structured reports and anchored onto a permissioned blockchain, ensuring immutability, traceability, and secure information sharing. Experimental results on adversarial traffic sign datasets demonstrate that the proposed framework outperforms state-of-the-art methods, validating its effectiveness for safe and trustworthy intelligent transportation systems.