Modern car networks still ship with no built-in authentication, which leaves the controller area network open to message-injection attacks. Detecting those intrusions on real hardware has been a balancing act: powerful transformer models understand message context but demand more compute than a typical ECU can spare.
The Ohio State team behind CANBERT+, including Jun Yeon Won, Qadeer Ahmed, and Zhiqiang Lin, designed it for the kind of hardware vehicles actually carry. Each CAN frame becomes a single token, with separate embedding layers for the identifier, data length, and payload bytes, and redundant layers stripped out. The result keeps byte-level context without long sequences, and the network runs about six times faster than earlier transformer detectors.
Training happens in two stages: self-supervised pre-training on benign traffic, then supervised fine-tuning on attack data. Evaluated across eight public datasets covering spoofing, fuzzing, denial-of-service, and replay attacks, CANBERT+ beat traditional machine learning baselines and prior BERT-based detectors. The biggest gain came on replay attacks, where the F1 score jumped from 0.60 to 0.72, a 20% improvement.
Two compact variants shrink the model by as much as 205 times while keeping detection accuracy high, which the authors say makes the approach viable for the limited memory and compute budgets found in vehicles today.
The team presented the work at VehicleSec 2026 in Baltimore, part of the USENIX Security week. With regulators pushing automakers toward in-vehicle monitoring under UNECE R155 and related frameworks, efficiency-focused detectors like CANBERT+ target the gap between research-grade models and software that can actually ship inside a car.
