CANMentor: A Large-Small Model Collaboration Framework for CAN Bus Intrusion Detection

summary

The gist

The paper proposes CANMentor, a large-small model collaboration framework that uses a cloud-side large language model teacher to supervise lightweight vehicle-side intrusion detection models for CAN…

In short

CANMentor is a framework that combines lightweight vehicle-side intrusion detection models with a powerful cloud-side large language model teacher for CAN bus security. It uses privacy-preserving proxies to let the cloud teacher supervise the local models, leading to significant performance gains of up to 25% on two datasets by improving detection of rare attacks.

Key concepts

Large-Small Model Collaboration
This framework pairs a large, powerful language model in the cloud (the teacher) with smaller, lightweight intrusion detection models on the vehicle. The small models perform real-time checks locally, while the large model guides and improves their performance through supervision.
Privacy-Aware Proxy Generation
To protect sensitive CAN data privacy when sending information to the cloud teacher, a generator creates proxies from original messages. These proxies are used for cloud inference instead of raw data, allowing the teacher to learn without exposing private vehicle traffic.
Knowledge Distillation
This technique transfers knowledge from the large teacher model to the smaller student models. The vehicle converts the teacher's predictions into 'soft targets' and uses them during training. This helps the lightweight IDS models learn more nuanced detection patterns, especially for difficult attacks.

Terminology used across episodes

This episode discusses

The paper

CANMentor: A Large-Small Model Collaboration Framework for CAN Bus Intrusion Detection · Read on arXiv

College of Cyber Security, Jinan University · School of Cyber Engineering, Xidian University

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "CANMentor: A Large-Small Model Collaboration Framework for CAN Bus Intrusion Detection".

Elias: Here's the gist of the paper. The paper proposes CANMentor, a large-small model collaboration framework that uses a cloud-side large language model teacher to supervise lightweight vehicle-side intrusion detection models for CAN bus intrusion detection The maximum relative Macro F1 improvements reach 25.01% and 20.79% on the two datasets, respectively Pruning redundant reasoning tokens CANMentor is a large-small model collaboration framework in which a cloud-side large language model (LLM) teacher supervises vehicle-side lightweight IDS models for CAN bus intrusion detection Each vehicle first performs local inference, and a privacy-aware generator then produces proxies from the original CAN messages for cloud-side teacher inference to protect data privacy The vehicle uses the teacher prediction only when the response arrives within a prescribed latency bound and otherwise retains the local prediction All returned teacher logits are converted into soft targets for periodic knowledge distillation.

Nadia: How the training signal is built The data preprocessing involves encoding each CAN message using its identifier, payload, and inter-arrival time to form a windowed sample X(u)i The vehicle-side detection and proxy sample generation utilizes a noise-conditioned U-Net proxy generator Gu that maps the windowed sample X(u)i to Xe(u)i using a noise vector z(u)i and a prescribed perturbation distance d The vehicle trains Gu to match the distribution of real windowed samples by minimizing an overall loss LG which includes adversarial loss, semantic and feature consistency losses, and a distance loss.

Elias: Where the gains come from The primary mechanism for improvement is knowledge distillation where vehicle u converts teacher logits into soft targets using temperature-scaled softmax to minimize the Kullback–Leibler divergence DKL(p Tτ p Sτ) from teacher soft targets to student probabilities The vehicle also computes anchor logits as lA,(u)i = C0u(X(u)i) and minimizes an anchor loss L(u) anchor along with the distillation objective L(u) KD to improve the lightweight IDS model.

Nadia: Teacher performance and privacy The cloud-side LLM teacher is adapted using proxy samples generated from public CAN data once per deployed generator version and reused until that generator is updated Teacher predictions on proxy samples corresponding to low-confidence inputs outperform the local predictions by 38.71–42.00 percentage points The maximum reconstruction advantage is at most 4.22% for ROAD (d = 0.40) and 5.47% for HCRL Car-Hacking (d = 0.35).

Title and authors: Nadia: The paper is titled "CANMentor: A Large-Small Model Collaboration Framework for CAN Bus Intrusion Detection," and the authors are Zhiquan Liu, Haolin Gao, Bo Wu, Jingjing Guo, Heng Sun, and Jianfeng Ma.

Elias: They're proposing this collaboration framework because standard vehicle-side intrusion detection systems have trouble generalizing when they encounter different types of data patterns on the CAN bus.

Nadia: So instead of just running one local model by itself, this paper sets up a system where you have a cloud-side large language model teacher guiding the vehicle's lightweight models.

Priya: What I see immediately is that they are trying to overcome the privacy issue too, which is huge since uploading raw CAN messages to the cloud would expose sensitive vehicle information.

Elias: Exactly, and they propose using a privacy-aware generator that makes proxies from the original CAN messages so the cloud teacher can learn without seeing everything directly.

Nadia: The implication here is that we can leverage the reasoning power of a massive model without compromising the security of the vehicle's data stream itself.

Priya: But how does this actually work in practice for someone sitting in a car? That part sounds really critical to me.

The paper's summary: Elias: The core idea is that each vehicle runs its own local inference first, and then it uses a proxy generator to create samples from the original CAN messages for the cloud-side teacher.

Nadia: Then, the vehicle only actually uses that teacher prediction if it comes back within a certain latency bound they set, otherwise, it sticks with what its own model predicted locally.

Priya: That sounds like a smart way to balance getting expert guidance from the cloud with maintaining real-time performance on the vehicle itself.

Elias: Right, and all those teacher predictions that come back are then used in a process called periodic knowledge distillation to periodically improve the lightweight IDS model on the vehicle.

Nadia: So essentially, it’s a loop: local detection, proxy generation for cloud guidance, conditional teacher use for immediate action, and then using those results to refine the local model over time.

Priya: It addresses that non-IID data problem by letting the massive LLM teacher see these carefully constructed proxies instead of just raw vehicle traffic.

The paper's improvements: Nadia: One of the main things they point out is that this knowledge distillation process, where you convert those teacher logits into soft targets, improves the vehicle-side IDS models significantly on non-IID data.

Elias: They claim a maximum relative Macro F1 improvement of twenty-five point zero one percent on the ROAD dataset and twenty point seven nine percent on the HCRL Car-Hacking dataset, respectively.

Priya: Those are pretty solid numbers, but we need to be careful about what those numbers actually mean when you look at different types of attacks.

Elias: The paper also shows that using these proxies gives a reconstruction advantage over just not using any proxy samples on the ROAD dataset, which is four point two two percent when the perturbation distance is set to zero point four zero.

Nadia: And for HCRL Car-Hacking, that advantage goes up to five point four seven percent when the perturbation distance is zero point three five; it shows that this method helps reconstruct the original data better than a baseline, which is interesting because it ties into the privacy aspect of making those proxies.

Priya: So so the method isn't just about getting higher detection scores; it’s also providing a measurable benefit in how well you can reconstruct what was originally sent, which is important for understanding the data distribution.

Conclusion: Elias: To wrap up, CANMentor combines these vehicle-side lightweight models with the cloud-side LLM teacher using fixed-distance proxy samples to supervise them.

Nadia: The main implication is that this framework helps vehicles detect intrusion blind spots for attack classes that are underrepresented in their training data, especially when the data isn't identically distributed.

Priya: It also shows a way to use the cloud teacher for improvement without directly exposing sensitive vehicle messages by relying on these privacy-preserving proxies.

Elias: The system is designed for real-time operation, using teacher predictions only when they arrive within a set latency bound while still using all those returned logits to periodically distill knowledge back into the lightweight model.

Nadia: So, in summary of CANMentor: A Large-Small Model Collaboration Framework for CAN Bus Intrusion Detection, we have a method that uses cloud intelligence to make local detection better without needing access to the raw data stream for every single decision.

Elias: It's a solid contribution because it shows how you can use large models as supervisors for resource-constrained systems in a way that respects strict real-time requirements.

Priya: It’s important to remember that the authors state their method involves adapting the teacher using proxy samples generated from public CAN data once per deployed generator version and reusing them until that generator is updated.

Nadia: We'll keep an eye on how this approach scales as more complex vehicle networks come online, because that's where the real security testing begins.

More episodes

← Home