Assessment of BERT and Lightweight LLMs for Intrusion Detection in Internet of Vehicles

Securing communication within vehicular systems is becoming more and more crucial as smart technologies are added to them. Despite its efficiency, the Controller Area Network (CAN) bus is vulnerable to cyberattacks due to its inherent lack of protection. Intelligent intrusion detection systems are desperately needed as autonomous and connected vehicles continue to improve. The goal of this research is to create sophisticated transformer designs that can comprehend semantic and sequence information included in CAN communication. Modern vehicular systems are becoming increasingly intelligent with the advent of the Internet of Vehicles (IoV) but they are also more vulnerable to cyberattacks, especially through the CAN bus. Conventional intrusion detection systems (IDS) frequently miss contextually relevant material and serial dependencies in CAN messages. This paper, explore the use of transformer-built large language models (LLMs) for CAN-bus intrusion detection. Using the Car Hacking dataset, we experimentally compared four transformer models: BERT, TinyBERT, AlBERT, and ELECTRA. Each model was refined using different data sets, ranging from 10,000 to 40,000 cases per class and raw CAN messages were transformed into byte sentence representations to maintain semantic structure. According to this research, TinyBERT achieved BERT-level performance while drastically lowering complexity, while ALBERT and ELECTRA models all performed accurately across a range of dataset sizes. These findings show that tiny transformers are feasible for in-vehicle, low-resource, and time-sensitive IDS. The study offers a solid standard for later transformer-based IDS in the IoV context by shedding light on model scalability, performance trade-offs and embedded applicability.