Source-linked AI summary
CANet: An Unsupervised Intrusion Detection System for High Dimensional CAN Bus Data
Markus Hanselmann, Thilo Strauss, Katharina Dormann, Holger Ulmer
TL;DR
Connected vehicles expose CAN-connected ECUs to remote attacks, while heterogeneous multi-ID traffic complicates intrusion detection and limits conventional approaches. The paper introduces CANet, an unsupervised neural architecture that processes individual messages across different IDs and detects anomalies through reconstruction. CANet reports strong detection of unknown attacks with high true negative rates, although its anomaly score may require refinement for deployment.
Problem
CAN lacks security mechanisms, and existing approaches are limited by costly rule configuration, difficulty detecting unknown attacks, and the heterogeneous structure of messages from different IDs.
Method
CANet uses separate LSTM subnetworks for each CAN ID, combines them into a joint latent vector, and reconstructs payloads for unsupervised anomaly detection.
Results
CANet detects a large amount of unknown attacks on real and synthetic data, maintains typically over 0.99 true negative rates, and reliably finds replay anomalies.
Takeaways & Limitations
CANet extends CAN intrusion detection to messages with different IDs simultaneously and provides a publicly available synthetic benchmark for reproducibility.
Takeaways & Limitations
The anomaly score may need refinement for in-vehicle use, where false positives are costly and a 100% true negative rate is required.
Abstract
from arXiv · showhide
We propose a novel neural network architecture for detecting intrusions on the CAN bus. The Controller Area Network (CAN) is the standard communication method between the Electronic Control Units (ECUs) of automobiles. However, CAN lacks security mechanisms and it has recently been shown that it can be attacked remotely. Hence, it is desirable to monitor CAN traffic to detect intrusions. In order to detect both, known and unknown intrusion scenarios, we consider a novel unsupervised learning approach which we call CANet. To our knowledge, this is the first deep learning based intrusion detection system (IDS) that takes individual CAN messages with different IDs and evaluates them in the moment they are received. This is a significant advancement because messages with different IDs are typically sent at different times and with different frequencies. Our method is evaluated on real and synthetic CAN data. For reproducibility of the method, our synthetic data is publicly available. A comparison with previous machine learning based methods shows that CANet outperforms them by a significant margin.
1 INTRODUCTION
CANet addresses the difficulty of detecting unknown attacks in heterogeneous CAN traffic, where rule-based systems require costly expert configuration and standard neural networks do not directly fit the message structure. The paper introduces an unsupervised architecture for multi-ID signal-space anomaly detection and reports performance superior to baseline methods.
- Motivation: Remote attacks on connected vehicles can compromise ECUs and inject messages that may cause dangerous vehicle behavior.The paper motivates CAN intrusion detection as a traffic-safety concern, including in autonomous vehicles.
- Research gap: Rule-based and statistical CAN IDS methods can detect known attacks efficiently but require time-consuming, expert-driven configuration and may miss unknown scenarios.They also struggle to encode complex physical dependencies between signals.
- Research gap: Consecutive CAN messages use different IDs, frequencies, and signal types, making the data structure difficult for standard neural networks to process directly.Prior neural approaches generally model a single signal or ID rather than the full CAN signal space.
- Contribution: CANet assigns each message ID an LSTM subnetwork and combines their representations to model the joint state of CAN traffic in an unsupervised autoencoder framework.The architecture reconstructs input payloads and supports detection of unknown attack types.
- Contribution: The authors report that CANet outperforms baseline methods by a significant margin and is especially strong against signal manipulations difficult for classical approaches.The paper also identifies anomaly detection in CAN signal space as potentially applicable to early technical-failure detection.
2 RELEVANT BACKGROUND
The paper frames CAN as the vehicle bus connecting ECUs and focuses on detecting manipulated signals that violate normal behavior or physical relationships. Existing methods cover rules, statistics, and single-series learning, while CANet extends learning to signals from multiple message IDs simultaneously.
- CAN terminology: CAN is a vehicle bus standard that enables automotive ECUs to communicate with one another.CAN messages contain an ID identifying the message type and a payload carrying current vehicle-signal values.
- Threat model: The threat model assumes an attacker has accessed the CAN bus through a compromised ECU and manipulates messages to influence vehicle behavior.The detection objective is intrusion into CAN communication under this access assumption.
- Detection objective: The detection goal is to identify signals that deviate from normal behavior or break complex physical relationships that are difficult to derive as rules.These relationships may be complex and potentially unknown.
- Related work: Prior neural CAN IDS methods model a single signal or the signals associated with one ID, whereas CANet processes signals from multiple IDs simultaneously.This extension is intended to exploit dependencies across the broader CAN traffic state.
- Related work: Earlier non-neural CAN anomaly detectors include signature, fingerprinting, clustering, fuzzy-logic, Hidden-Markov-Model, and entropy-based methods.The paper places these approaches within a broader review of CAN IDS techniques.
3 CANET
CANet uses one recurrent subnetwork per CAN ID to maintain a joint latent representation, reconstruct payloads, and detect anomalies from reconstruction errors. Its anomaly score is thresholded using normal validation data, but the paper notes practical limitations for large signal sets and in-vehicle false-positive requirements.
- Network architecture: CANet uses a separate LSTM for each CAN ID, and concatenates their states into a joint latent vector representing the current traffic state.The design stores temporal dependencies per ID while combining information across IDs.
- Network architecture: When one ID payload arrives, its LSTM updates the corresponding latent representation, which is used to reconstruct payloads for all considered IDs.The reconstruction is defined across the ordered set of IDs and their associated signals.
- Training objective: The model compares the current ID’s true signal values with their reconstructions using a quadratic error loss.Backpropagation computes gradients for the current ID’s LSTM and its connection to the corresponding output.
- Training objective: Because each ID stores temporal dependencies separately, training is not sensitive to the exact order of consecutive message IDs.This accommodates variability in ID order within real CAN data.
- Anomaly score: A signal is marked anomalous when its reconstruction error exceeds a threshold set from normal validation errors at the 99.99% percentile.The global anomaly score is one if at least one stored signal indicator is anomalous.
- Anomaly score: The anomaly score becomes less feasible as the number of signals grows because multiple-testing effects reduce the probability of correctly identifying normal data.The authors also state that the score may need refinement for in-vehicle use, where false positives are costly and a 100% true negative rate is required.
4 EXPERIMENTS
CANet is evaluated on real and synthetic CAN traffic across normal data and five attack types, then compared with adapted baseline methods. It detects correlated signal anomalies and unknown attacks effectively, while exposing limitations for suppress attacks, large models, and deployment-oriented scoring.
- Detection performance: Normal-data accuracy is typically above 0.99, while plateau and playback attacks usually achieve true positive rates of 0.85–0.955 and continuous change exceeds 0.70.The evaluation is pointwise, so these rates reflect successfully detected attack points rather than necessarily complete attack intervals.
- Baseline comparison: CANet outperforms the predictive and autoencoder baselines by a significant margin, which typically detect only the first few attacked messages.The baselines perform particularly poorly on playback because single-signal models cannot exploit dependencies among correlated signals.
- Attack-specific scope: Flooding attacks are detected with a high true positive rate, but suppress attacks remain difficult because CANet has no timestamp input and is not designed specifically for frequency-based attacks.Rule-based methods can detect suppress and flooding attacks straightforwardly by analyzing message frequencies.
- Model size: Small models perform reasonably well, but synthetic-data performance decreases significantly at hscale = 30, which the authors attribute to overfitting.The authors suggest changing other architectural parameters, such as the autoencoder bottleneck, to mitigate overfitting.
- Detection behavior: CANet detects anomalies across functionally dependent signals, while unrelated signals remain unaffected by an attack on another signal.In the synthetic playback example, deviations appear in B and correlated A and C, but not in independent D.
5 CONCLUSION
CANet is an unsupervised neural-network IDS for CAN-bus intrusions, designed to detect attacks on connected vehicles. It achieves high normal-data discrimination, detects unknown attacks, and is evaluated using reproducible synthetic data.
- CANet is an unsupervised neural-network architecture for detecting intrusions and anomalies on the CAN bus.
- The models achieve a true negative rate typically over 0.99, a level described as necessary for real-world applications.
- For attack-interval evaluation, CANet finds most anomaly intervals when the required point-wise detection percentage is not too large, outperforming both baselines.
- CANet detects a large amount of unknown attacks on both real and synthetic data, including replay anomalies reliably.
- The synthetic CAN data is publicly available as a reproducible benchmark for forthcoming approaches.