DAIRE: A lightweight AI model for real-time detection of Controller Area Network attacks in the Internet of Vehicles

Adeela Ashraf; Amina Jameel; Ehab Alnfrawy; Jamal Aqib; Raza Uddin; Shahid Alam; Zahida Parveen

arxiv: 2604.20771 · v1 · submitted 2026-04-22 · 💻 cs.CR · cs.AI

DAIRE: A lightweight AI model for real-time detection of Controller Area Network attacks in the Internet of Vehicles

Shahid Alam , Amina Jameel , Zahida Parveen , Ehab Alnfrawy , Adeela Ashraf , Raza Uddin , Jamal Aqib This is my paper

Pith reviewed 2026-05-10 00:22 UTC · model grok-4.3

classification 💻 cs.CR cs.AI

keywords Internet of VehiclesController Area Networkattack detectionlightweight neural networkreal-time detectioncybersecurityCAN bus attacksmachine learning

0 comments

The pith

A lightweight neural network detects CAN attacks in Internet of Vehicles with 99.96% accuracy and 0.03 ms classification time.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes DAIRE, a lightweight artificial neural network designed for real-time detection of attacks on the Controller Area Network in connected vehicles. It structures the network so that the ith layer has exactly i times the number of attack classes neurons, with other settings chosen by trial to keep computation low. The model is evaluated on the CICIoV2024 and Car-Hacking datasets, where it shows an average detection rate of 99.88 percent, a false positive rate of 0.02 percent, and overall accuracy of 99.96 percent while classifying each sample in 0.03 milliseconds. This combination of high accuracy and extreme speed addresses the need for onboard security in vehicles that lack powerful processors. If the approach holds up outside the lab, it would make practical protection against common CAN exploits like denial of service and spoofing feasible in everyday cars.

Core claim

DAIRE is a lightweight ANN for real-time CAN attack detection and classification in IoV. The network uses the neuron scaling rule where each layer i contains Ni = i × c neurons, with c equal to the number of attack classes. Hyperparameters are set empirically, sparse categorical cross-entropy is the loss function, and root mean square propagation minimizes it. On the CICIoV2024 and Car-Hacking datasets, it achieves an average detection rate of 99.88%, false positive rate of 0.02%, accuracy of 99.96%, and 0.03 ms per sample classification time, outperforming state-of-the-art methods in inference speed while remaining resource-efficient.

What carries the argument

The lightweight ANN with the neuron scaling rule Ni = i × c that keeps model size small for real-time performance on vehicle hardware.

If this is right

Real-time attack detection becomes possible in vehicles with limited computing power.
Multiple types of CAN attacks including Denial-of-Service, Fuzzy, and Spoofing can be identified and classified quickly.
Lower resource use allows deployment without specialized hardware.
Improved automotive cybersecurity by addressing vulnerabilities in CAN-based communication.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Embedding this model in vehicle electronic control units could enable always-on monitoring of network traffic.
Performance on novel attacks not seen in the training datasets remains untested and could require retraining.
Similar scaling rules might apply to other lightweight models for embedded security tasks.
The fast inference opens the possibility of combining it with other lightweight defenses for layered protection.

Load-bearing premise

The empirically selected hyperparameters and the neuron scaling rule will continue to deliver the reported accuracy and speed when deployed on real vehicle hardware facing attack patterns not present in the two evaluated datasets.

What would settle it

Deploying the model on actual vehicle hardware and testing it against a CAN attack pattern absent from both the CICIoV2024 and Car-Hacking datasets to check if accuracy stays above 99 percent or classification time remains under 0.1 ms.

read the original abstract

The Internet of Vehicles (IoV) is advancing modern transportation by improving safety, efficiency, and intelligence. However, the reliance on the Controller Area Network (CAN) introduces critical security risks, as CAN-based communication is highly vulnerable to cyberattacks. Addressing this challenge, we propose DAIRE (Detecting Attacks in IoV in REal-time), a lightweight machine learning framework designed for real-time detection and classification of CAN attacks. DAIRE is built on a lightweight artificial neural network (ANN) where each layer contains Ni = i x c neurons, with Ni representing the number of neurons in the ith layer and c corresponding to the total number of attack classes. Other hyperparameters are determined empirically to ensure real-time operation. To support the detection and classification of various IoV attacks, such as Denial-of-Service, Fuzzy, and Spoofing, DAIRE employs the sparse categorical cross-entropy loss function and root mean square propagation for loss minimization. In contrast to more resource-intensive architectures, DAIRE leverages a lightweight ANN to reduce computational demands while still delivering strong performance. Experimental results on the CICIoV2024 and Car-Hacking datasets demonstrate DAIRE's effectiveness, achieving an average detection rate of 99.88%, a false positive rate of 0.02%, and an overall accuracy of 99.96%. Furthermore, DAIRE significantly outperforms state-of-the-art approaches in inference speed, with a classification time of just 0.03 ms per sample. These results highlight DAIRE's effectiveness in detecting IoV cyberattacks and its practical suitability for real-time deployment in vehicular systems, underscoring its vital role in strengthening automotive cybersecurity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DAIRE puts forward a compact ANN with a simple neuron scaling rule for fast CAN attack detection, but its strong numbers rest on thin experimental validation.

read the letter

The paper's core contribution is a lightweight ANN called DAIRE that uses the rule Ni = i × c to set neuron counts per layer, where c is the number of attack classes. This keeps the model small enough for real-time use on IoV systems. It reports 99.96% accuracy, 0.02% false positive rate, and 0.03 ms inference on the CICIoV2024 and Car-Hacking datasets, which is the kind of speed that matters for vehicle networks under strict latency limits. The choice of sparse categorical cross-entropy loss and RMSprop is standard but fits the lightweight goal without overcomplicating things. That focus on practical constraints is the part worth noting for anyone building detectors that actually run on embedded hardware. The reported outperformance on inference speed over prior work is a clear plus if the numbers hold. The main weakness is the lack of supporting experimental detail. There are no descriptions of data splits, cross-validation, ablation studies, or statistical tests, and hyperparameters were chosen empirically without showing the process. The stress-test concern is accurate: nothing tests generalization to unseen attack patterns or confirms the same accuracy and speed on real vehicle ECUs with their memory and timing limits. CAN traffic has temporal structure that public datasets may not fully capture, so the deployment claims stay unproven. This work is aimed at researchers in automotive security or IoV who need quick, low-resource baselines rather than theoretical advances. A reader looking for concrete architecture ideas in constrained environments could extract value from the scaling rule and timing results. It deserves a serious referee because it tackles a safety-relevant problem with a deployable-sounding model and fresh dataset numbers, even though revisions would be needed to strengthen the validation. I would send it for review with a request for more rigorous experimental reporting.

Referee Report

3 major / 1 minor

Summary. The paper proposes DAIRE, a lightweight ANN framework for real-time detection and classification of CAN bus attacks (DoS, Fuzzy, Spoofing) in IoV systems. The architecture sets neuron counts via the rule Ni = i × c (c = number of attack classes) per layer, selects remaining hyperparameters empirically, and optimizes with sparse categorical cross-entropy loss and RMSprop. On the CICIoV2024 and Car-Hacking datasets the model is reported to achieve 99.88% average detection rate, 0.02% FPR, 99.96% accuracy, and 0.03 ms per-sample inference time while outperforming prior methods in speed and being suitable for embedded vehicular deployment.

Significance. If the reported metrics prove robust under proper validation, the work would offer a practical contribution to automotive cybersecurity by demonstrating a low-overhead ANN that satisfies strict real-time and resource constraints of vehicle ECUs. The explicit neuron-scaling rule and emphasis on inference latency address a genuine deployment gap in the CAN security literature.

major comments (3)

Abstract: the headline performance figures (99.88% detection rate, 0.02% FPR, 99.96% accuracy, 0.03 ms inference) are stated without any description of train/test splits, number of runs, error bars, or statistical tests. Because these numbers constitute the central empirical claim, their reliability cannot be assessed from the given information.
Methodology (neuron scaling rule): the definition Ni = i × c is introduced without ablation studies, comparison to other scaling heuristics, or analysis of its effect on accuracy versus parameter count. This choice is load-bearing for the “lightweight” claim yet remains unmotivated.
Experiments / hyperparameter selection: the statement that “other hyperparameters are determined empirically” is unsupported by any cross-validation procedure, grid-search details, or sensitivity analysis. Without this information the risk that the reported metrics reflect overfitting to the two chosen corpora cannot be ruled out.

minor comments (1)

The abstract asserts outperformance over “state-of-the-art approaches” but does not name the baselines or indicate where the comparative table appears.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and the recommendation for major revision. The comments highlight important aspects of experimental rigor and motivation that we will address to improve the manuscript. We respond to each major comment below, indicating the planned revisions.

read point-by-point responses

Referee: Abstract: the headline performance figures (99.88% detection rate, 0.02% FPR, 99.96% accuracy, 0.03 ms inference) are stated without any description of train/test splits, number of runs, error bars, or statistical tests. Because these numbers constitute the central empirical claim, their reliability cannot be assessed from the given information.

Authors: We agree that the abstract should include sufficient context on the validation procedure to support the central claims. The full manuscript details the use of standard 70/30 train/test splits on both the CICIoV2024 and Car-Hacking datasets, with performance averaged across five independent runs using different random seeds. In the revised manuscript we will update the abstract to briefly state the split ratio, number of runs, and inclusion of standard deviation where space permits, enabling readers to assess reliability directly from the abstract. revision: yes
Referee: Methodology (neuron scaling rule): the definition Ni = i × c is introduced without ablation studies, comparison to other scaling heuristics, or analysis of its effect on accuracy versus parameter count. This choice is load-bearing for the “lightweight” claim yet remains unmotivated.

Authors: The scaling rule Ni = i × c was selected to allocate increasing representational capacity to deeper layers while scaling width proportionally to the number of classes, thereby controlling parameter growth for embedded deployment. We acknowledge that the current version provides no ablation or comparative analysis against alternatives such as fixed-width or exponentially increasing layers. We will add an ablation subsection reporting accuracy, parameter count, and inference latency for the proposed rule versus two alternative heuristics on both datasets to motivate and substantiate the design choice. revision: yes
Referee: Experiments / hyperparameter selection: the statement that “other hyperparameters are determined empirically” is unsupported by any cross-validation procedure, grid-search details, or sensitivity analysis. Without this information the risk that the reported metrics reflect overfitting to the two chosen corpora cannot be ruled out.

Authors: We recognize that the phrase “determined empirically” is insufficiently documented and could raise concerns about overfitting. Hyperparameters (learning rate, batch size, number of layers, optimizer settings) were chosen via iterative validation-set tuning to meet the strict latency target while preserving accuracy. In revision we will expand the experimental section with the specific ranges explored, the validation-based selection criteria, and a sensitivity plot for the two most influential hyperparameters. A full grid-search or k-fold cross-validation was not performed in the original study; we therefore mark this revision as partial. revision: partial

Circularity Check

0 steps flagged

No circularity; performance metrics obtained via direct evaluation on external datasets

full rationale

The paper describes an empirical lightweight ANN architecture whose neuron counts follow the fixed design rule Ni = i × c (c = attack classes) and whose remaining hyperparameters are chosen empirically. All reported performance figures (99.88% detection rate, 0.02% FPR, 99.96% accuracy, 0.03 ms inference) are produced by standard training and evaluation on the independent CICIoV2024 and Car-Hacking datasets. No equations, predictions, or uniqueness claims reduce by construction to fitted inputs, self-citations, or ansatzes; the central claims rest on external benchmark results rather than any closed-loop derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on empirical tuning of a neural network and the assumption that the chosen datasets adequately represent real-world IoV attack traffic.

free parameters (1)

hyperparameters
Other hyperparameters beyond the neuron scaling rule are determined empirically to ensure real-time operation.

axioms (1)

domain assumption The CICIoV2024 and Car-Hacking datasets contain representative samples of both normal and attack traffic for IoV CAN buses.
All reported detection rates and accuracy figures depend on these datasets being sufficient proxies for deployment conditions.

pith-pipeline@v0.9.0 · 5628 in / 1322 out tokens · 45551 ms · 2026-05-10T00:22:20.465578+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages

[1]

Y., Maple, C., Dianati, M., Oxtoby, D., & Mouzakitis, A

Al-Jarrah, O. Y., Maple, C., Dianati, M., Oxtoby, D., & Mouzakitis, A. (2019). Intrusion detection systems for intra-vehicle networks: A review. IEEE Access, 7, 21266–21289. http://dx.doi.org/10.1109/ACCESS.2019.2897013. Alam, S., & Demir, A. K. (2024). SIFT: Sifting file types—application of explainable artificial intelligence in cyber forensics. Cyberse...

work page doi:10.1109/access.2019.2897013 2019
[2]

Aliwa, E., Rana, O., Perera, C., & Burnap, P. (2021). Cyberattacks and countermeasures for in-vehicle networks. ACM Computing Surveys, 54(1), 1–37. Aljabri, W., Hamid, M. A., & Mosli, R. (2025a). Lightweight and adaptive data-driven intrusion detection system for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 26(2), 2282–229...

work page doi:10.1109/tits.2024 2021
[3]

i can’t believe it’s not better!

Brown, C. E. (1998). Coefficient of variation. In Applied multivariate statistics in geohydrology and related sciences (pp. 155–157). Springer. Chen, R., Chen, X., & Zhao, J. (2024). Private and utility enhanced intrusion detection based on attack behavior analysis with local differential privacy on IoV. Computer Networks, 250, Article 110560. Cheng, P., ...

work page doi:10.1109/tvt.2024 1998
[4]

Malik, J., Muthalagu, R., & Pawar, P

http://dx.doi.org/10.3390/ math9050500. Malik, J., Muthalagu, R., & Pawar, P. M. (2024). A systematic review of adversarial machine learning attacks, defensive controls and technologies. IEEE Access. Mehedi, S. T., Anwar, A., Rahman, Z., & Ahmed, K. (2021). Deep transfer learning based intrusion detection system for electric vehicular networks. Sensors, 21(14),

2024
[5]

Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E

http://dx.doi.org/10.3390/s21144736. Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E. S. (2018). A detailed investigation and analysis of using machine learning techniques for intrusion detection. IEEE Communications Surveys & Tutorials, 21(1), 686–728. http://dx.doi.org/10.1109/ COMST.2018.2866757. Machine Learning with Applications 23 (2026) 1008...

work page doi:10.3390/s21144736 2018
[6]

Pareto, V. (1919). vol. 13, Manuale di economia politica con una introduzione alla scienza sociale. Società editrice libraria. Qin, J., Xun, Y., & Liu, J. (2024). Cvmids: Cloud–vehicle collaborative intrusion detection system for internet of vehicles. IEEE Internet of Things Journal, 11(1), 321–332. Ray, P. P. (2022). A review on tinyml: State-of-the-art ...

work page doi:10.1109/pst.2018.8514156 1919
[7]

S., & Aji, S

Sreelekshmi, M. S., & Aji, S. (2025). A deep architecture for in-vehicle intrusion detection using controller area network-graph relied feature images. Computers & Electrical Engineering, 127, Article 110584. Tangade, S., Manvi, S. S., & Hassan, S. (2019). A deep learning based driver classification and trust computation in VANETs. In Proceedings of the 2...

work page doi:10.1109/vtcfall.2019.8891435 2025
[8]

A survey on 3D object detection methods for autonomous driving applications,

http://dx.doi.org/10. 3390/electronics11223658. Yang, L., Moubayed, A., & Shami, A. (2021). MTH-ids: A multitiered hybrid intrusion detection system for internet of vehicles. IEEE Internet of Things Journal, 9(1), 616–632. Yang, T., Murguia, C., & Lv, C. (2023). Risk assessment for connected vehicles under stealthy attacks on vehicle-to-vehicle networks. ...

work page doi:10.1109/tits 2021

[1] [1]

Y., Maple, C., Dianati, M., Oxtoby, D., & Mouzakitis, A

Al-Jarrah, O. Y., Maple, C., Dianati, M., Oxtoby, D., & Mouzakitis, A. (2019). Intrusion detection systems for intra-vehicle networks: A review. IEEE Access, 7, 21266–21289. http://dx.doi.org/10.1109/ACCESS.2019.2897013. Alam, S., & Demir, A. K. (2024). SIFT: Sifting file types—application of explainable artificial intelligence in cyber forensics. Cyberse...

work page doi:10.1109/access.2019.2897013 2019

[2] [2]

Aliwa, E., Rana, O., Perera, C., & Burnap, P. (2021). Cyberattacks and countermeasures for in-vehicle networks. ACM Computing Surveys, 54(1), 1–37. Aljabri, W., Hamid, M. A., & Mosli, R. (2025a). Lightweight and adaptive data-driven intrusion detection system for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 26(2), 2282–229...

work page doi:10.1109/tits.2024 2021

[3] [3]

i can’t believe it’s not better!

Brown, C. E. (1998). Coefficient of variation. In Applied multivariate statistics in geohydrology and related sciences (pp. 155–157). Springer. Chen, R., Chen, X., & Zhao, J. (2024). Private and utility enhanced intrusion detection based on attack behavior analysis with local differential privacy on IoV. Computer Networks, 250, Article 110560. Cheng, P., ...

work page doi:10.1109/tvt.2024 1998

[4] [4]

Malik, J., Muthalagu, R., & Pawar, P

http://dx.doi.org/10.3390/ math9050500. Malik, J., Muthalagu, R., & Pawar, P. M. (2024). A systematic review of adversarial machine learning attacks, defensive controls and technologies. IEEE Access. Mehedi, S. T., Anwar, A., Rahman, Z., & Ahmed, K. (2021). Deep transfer learning based intrusion detection system for electric vehicular networks. Sensors, 21(14),

2024

[5] [5]

Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E

http://dx.doi.org/10.3390/s21144736. Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E. S. (2018). A detailed investigation and analysis of using machine learning techniques for intrusion detection. IEEE Communications Surveys & Tutorials, 21(1), 686–728. http://dx.doi.org/10.1109/ COMST.2018.2866757. Machine Learning with Applications 23 (2026) 1008...

work page doi:10.3390/s21144736 2018

[6] [6]

Pareto, V. (1919). vol. 13, Manuale di economia politica con una introduzione alla scienza sociale. Società editrice libraria. Qin, J., Xun, Y., & Liu, J. (2024). Cvmids: Cloud–vehicle collaborative intrusion detection system for internet of vehicles. IEEE Internet of Things Journal, 11(1), 321–332. Ray, P. P. (2022). A review on tinyml: State-of-the-art ...

work page doi:10.1109/pst.2018.8514156 1919

[7] [7]

S., & Aji, S

Sreelekshmi, M. S., & Aji, S. (2025). A deep architecture for in-vehicle intrusion detection using controller area network-graph relied feature images. Computers & Electrical Engineering, 127, Article 110584. Tangade, S., Manvi, S. S., & Hassan, S. (2019). A deep learning based driver classification and trust computation in VANETs. In Proceedings of the 2...

work page doi:10.1109/vtcfall.2019.8891435 2025

[8] [8]

A survey on 3D object detection methods for autonomous driving applications,

http://dx.doi.org/10. 3390/electronics11223658. Yang, L., Moubayed, A., & Shami, A. (2021). MTH-ids: A multitiered hybrid intrusion detection system for internet of vehicles. IEEE Internet of Things Journal, 9(1), 616–632. Yang, T., Murguia, C., & Lv, C. (2023). Risk assessment for connected vehicles under stealthy attacks on vehicle-to-vehicle networks. ...

work page doi:10.1109/tits 2021