Pith. sign in

REVIEW 3 major objections 4 minor 46 references

SDN-Based False Data Detection With Its Mitigation and Machine Learning Robustness for In-Vehicle Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A combined SDN and LSTM system can detect and block false data injected into brake-related CAN messages within the 10 ms vehicle message deadline.

desk verdict The architecture is sensible, but the reported metrics are internally impossible; the robustness and real-time claims are unverified. read the letter →

arxiv 2506.06556 v1 pith:MS6QGEPE submitted 2025-06-06 cs.LG cs.CR

classification cs.LGcs.CR
keywords ControllerAreaNetworkFalsedatainjectionattackLongShort-TermMemorySoftware-definednetworkingAdversarialrobustnessDeepFoolIntrusiondetectionIn-vehiclesecurity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes a False Data Detection and Mitigation System (FDDMS) for in-vehicle CAN networks and argues that it can catch false data injection attacks on brake-related ECUs in real time. The system places an SDN controller and an OpenFlow switch on the CAN bus, uses an LSTM network to flag injected messages, and updates forwarding rules to redirect attack traffic away from receiving ECUs. The authors also claim that adversarially retraining the detector on only hard examples, selected by a confidence threshold, keeps accuracy near 99.47 percent against four adversarial attack methods, including a new lower-distortion DeepFool variant. They report an average end-to-end latency of about 7.76 ms, below the 10 ms message interval that defines real time for this CAN data. The reason to care is concrete: if false brake data is not caught inside that window, it may directly influence braking actions and cause a crash.

What carries the argument

The load-bearing mechanism is the FDDMS pipeline itself: an OpenFlow switch placed on the CAN bus gives the SDN controller a global view of broadcast traffic, the LSTM detector (a recurrent network with 128 memory cells that reads time-ordered CAN signal sequences) classifies each window as normal or attacked, and upon an attack the controller rewrites flow-table entries so the malicious messages are sent to backend storage instead of broadcast to the ABS, ESC, and EPB ECUs. A second mechanism is the adversarial-training loop: a score function $g(S_k, y_k)$ measures the model's confidence on each candidate adversarial example, a threshold of 0.5 admits only low-confidence challenging examples into the retraining set, and the proposed DeepFool variant uses gradient clipping bounded at $\pm 0.95$ and a scaling factor $\kappa = 0.5$ to shrink the perturbation at each iteration. This threshold-selection loop is what keeps normal and robust accuracy both near 99 percent instead of trading one against the other.

What would settle it

Run FDDMS on a hardware-in-the-loop vehicle testbed under worst-case CAN bus load and measure the attack-to-mitigation latency; if any message takes longer than the 10 ms interval, or if corrupting the SDN controller or switch can prevent the flow-rule update, the real-time mitigation claim is false. A direct check for the robustness claim is to attack the retrained detector with a black-box surrogate or an adaptive attacker who knows the 0.5 selection threshold; a large drop in detection accuracy would show the defense is specific to the four white-box attacks tested.

Watch

Extended reading notes

Core claim

The central discovery the paper is trying to establish is that false data injection on a CAN bus can be stopped inside the vehicle's own message timing: FDDMS detects an attack, updates OpenFlow flow rules, and redirects the malicious messages to backend storage, all within an average of 7.7629 ms against a 10 ms broadcast interval. The detector is a 128-neuron Long Short-Term Memory network trained on 20 decoded signals from five brake-related ECUs, with raw CAN frames reorganized into fixed-length sequences and a binary normal/attack output. To make the detector harder to evade, the authors introduce a DeepFool variant that clips gradients and scales each perturbation step, reaching a 99 percent attack success rate with lower $L_0$, $L_2$, and $L_\infty$ distortion than FGSM, BIM, or the original DeepFool. They extend an iterative adversarial-retraining procedure with a confidence-score threshold of 0.5 so that only challenging adversarial examples enter the training pool; the best retrained model keeps 99.47 percent normal accuracy and robust accuracy of 99.475, 99.47, 98.95, and 99.47 percent against FGSM, BIM, original DeepFool, and the DeepFool variant, respectively. The paper takes these numbers as evidence that the combined detection and mitigation loop is accurate and fast enough to be a practical in-vehicle defense.

Load-bearing premise

The real-time mitigation claim rests on the assumption that the 7.76 ms latency measured on an SDN testbed, with a trusted OpenFlow switch and controller placed on the CAN bus, carries over to a production vehicle, and that the control plane itself cannot be compromised.

Editorial extensions

If this is right

  • If the central claim holds, an in-vehicle defense can run inside the CAN timing budget: the full detect-and-mitigate loop averages 7.7629 ms, under the 10 ms interval between messages in this dataset.
  • The adversarially retrained detector keeps accuracy near 99.47 percent against all four white-box attacks, so the claimed defense does not sacrifice normal performance for robustness.
  • The DeepFool variant gives attackers a sharper tool for testing in-vehicle detectors: 99 percent attack success with smaller $L_0$, $L_2$, and $L_\infty$ perturbations than existing methods.
  • Because mitigation happens by updating OpenFlow flow rules, attack traffic can be redirected to backend storage rather than simply dropped, preserving forensic data while protecting braking ECUs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not carry out is to move the same pipeline from the SDN testbed onto a hardware-in-the-loop or production CAN bus and measure worst-case flow-rule update latency under realistic bus load; that is the experiment most likely to confirm or break the real-time claim.
  • The threat model is white-box: every attack gradient is computed against the known LSTM. A black-box or adaptive attacker who queries the detector or trains a surrogate could plausibly bypass the threshold-selection defense, so the robustness claim is best read as specific to the four tested attacks.
  • Placing the OpenFlow switch and SDN controller in the message path introduces a new single point of failure: a compromised controller could itself drop or fabricate traffic, so the design implicitly assumes the control plane is trusted and secure.
  • The confidence-threshold sampling rule could generalize to other sequential intrusion detectors; a testable question is whether the 0.5 threshold holds across different vehicles, signal sets, and attack models or must be retuned for each deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes an SDN-based False Data Detection and Mitigation System (FDDMS) for CAN-based in-vehicle networks. It decodes a real KIA Soul CAN dataset, constructs an FDIA attack model over 20 brake-related signals, trains an LSTM detector, proposes a modified DeepFool attack with gradient clipping and a scaling factor, and augments adversarial training with a confidence-threshold sample selection. The paper reports a normal detection accuracy of 99.47%, robust accuracy around 99% under FGSM, BIM, DeepFool, and the proposed DeepFool variant, and an overall mitigation latency of 7.76 ms on a GENI testbed, from which it concludes that the system can detect and mitigate attacks in real time within the 10 ms CAN message interval. The main claimed contributions are the low-distortion DeepFool variant, the re-training technique with sample selection, and the SDN-based mitigation scheme.

Significance. The problem is relevant and timely, and the paper addresses a real CAN dataset and provides algorithm pseudocode and hyperparameter details, which are helpful for reproducibility. If the reported robustness and latency numbers were reliable, the work would be a useful data point for machine-learning-based in-vehicle intrusion detection and for SDN-assisted mitigation. However, the central quantitative claims rest on measurement tables that are internally inconsistent with the stated feature dimension and test-set size, and no code, data splits, or error bars are provided to resolve the discrepancies. Until those numbers are corrected and independently reproduced, the significance of the empirical contribution cannot be assessed.

major comments (3)
  1. [Section IV-B, Table V] The distortion metrics in Table V violate the definitions given in the same section. With 20 input features, the L0 norm (number of perturbed features) cannot exceed 20 and cannot be fractional, yet Table V reports L0 = 17,198.69 for FGSM-L∞ and 6,543.51 for the proposed DeepFool. Similarly, for FGSM-L∞ with L∞ = 13.00, the largest possible L2 norm for a 20-dimensional perturbation is sqrt(20)*13 ≈ 58.14, but Table V reports L2 = 2,199,559.0. These inconsistencies are not cosmetic: the paper's primary attack contribution is the claim that the DeepFool variant achieves lower L0/L2/L∞ distortion, and that claim is supported only by this table.
  2. [Section IV-A2, Table III] Table III reports a robust accuracy of 99.475% for the model trained on Challenging FGSM examples. Section III-D1 states that the test set has 189 samples, so accuracy can only take values k/189; 99.475% is not an attainable value (188/189 ≈ 99.47%). The same table's first row reports a normal accuracy of 100% for a model trained on FDIA, which contradicts Table II, where the best FDIA-trained optimizers achieve 99.47% and 98.95%. Because these numbers support the paper's robustness claim, the evaluation as reported is internally inconsistent.
  3. [Section IV-C, Table VI] The real-time mitigation claim depends on the GENI testbed reproducing CAN bus timing. The paper asserts that CAN messages are broadcast every 10 ms in this dataset and reports an overall latency of 7.7629 ms, but it provides no information about the GENI network's timing characteristics, link rates, or how the OpenFlow switch and controller placement maps to an actual CAN bus. In particular, a CAN bus is a shared broadcast medium without source or destination addresses, whereas the SDN mitigation redirects 'attack traffic' by flow rules; the paper does not explain how an OVS on the bus can identify and redirect individual CAN frames without breaking normal broadcast semantics. This external-validity gap leaves the real-time mitigation claim unsupported as presented.
minor comments (4)
  1. [Section III-D1, Eq. (5)] Equation (5) defines clipping with 'min value' and 'max value' and then states that minval = maxval = α = 0.95; as written, this is a degenerate clipping interval. Algorithm 1 suggests the intended range is [-α, +α], so the equation and the accompanying text should be corrected.
  2. [Section III-D2, Eqs. (7)-(8)] The score function g(S_k, y_k) is described only as being based on the LSTM model's prediction with respect to the ground-truth label; the concrete definition should be stated so that the sample-selection procedure is reproducible.
  3. [Section IV-C, Table VI] Table VI is difficult to read: the detection and mitigation time columns are not cleanly aligned in the rows, and the average detection time of 1.6217 ms and average mitigation time of 0.7991 ms cited in the introduction do not appear in the table's average row.
  4. [Throughout] There are several typographical and stylistic issues, including 'an for get gate' in Section III-D1, 'We splitted' instead of 'We split', and inconsistent capitalization of 'OpenDBC' versus 'DBC'. These should be cleaned up in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical measurements, and its self-citations are transparent building blocks rather than load-bearing reductions.

full rationale

The derivation chain in this paper is empirical rather than deductive. The LSTM detector is trained on a public in-vehicle CAN dataset with a synthetic attack model, and the robustness claims are evaluated by running external attacks (FGSM, BIM, DeepFool) and the authors' DeepFool variant on held-out test samples. The reported accuracies and distortion metrics are measurements, not quantities forced by construction from fitted parameters. Hyperparameters such as the selection threshold (0.5), the DeepFool scale factor (kappa = 0.5), and the clipping factor (alpha = 0.95) are fixed hand-set constants; they are not fitted to the test set in a way that would make the reported test accuracies tautological. The self-citations to the authors' prior work [21] for the adversarial retraining procedure and to [1] for the first author's dissertation are used transparently as prior building blocks; they are not invoked as an external 'uniqueness theorem' or as the sole justification of the central claim. The key contributions—detection accuracy on CAN data, comparison of adversarial perturbation magnitudes, and SDN mitigation latency—are supported by new measurements on an external dataset and a GENI testbed. Even if some reported numbers (e.g., the L0 and L2 values in Table V) are internally inconsistent or otherwise suspect, those are correctness or reproducibility concerns, not circularity: the numbers are not shown to be equivalent to the method's inputs by any equation in the paper. No circular step can be exhibited with a specific quote and reduction, so the appropriate score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper relies on a standard public dataset, a synthetic attack model, and a simulated SDN testbed. The free parameters are all manually chosen hyperparameters. No new physical or conceptual entities are introduced.

free parameters (7)
  • threshold (sample selection) = 0.5
    Used in Equation 7 to select challenging adversarial examples; chosen by the authors without sensitivity analysis.
  • kappa (DeepFool scaling factor) = 0.5
    Used in Algorithm 1 to scale accumulated perturbation; set by hand.
  • alpha (gradient clipping bound) = 0.95
    Used in Algorithm 1 to clip gradients; set by hand.
  • epsilon (DeepFool overshoot) = 0.02
    Overshooting factor in Algorithm 1; standard DeepFool setting.
  • FGSM/BIM epsilon = 13.0
    Perturbation bound for FGSM and BIM attacks; chosen for high success rate.
  • correlation threshold = 0.75
    Threshold for determining feature correlations in Section III-B; chosen qualitatively.
  • LSTM hyperparameters = 128 units, 50 epochs, lr=0.001, batch=200, retrain 30 epochs
    Model architecture and training settings from Sections IV-A and IV-B; chosen without ablations.
assumptions (4)
  • domain assumption The OTIDS dataset is representative of real CAN bus traffic and contains accurate signal values after DBC decoding.
    The paper uses the public KIA Soul dataset [22] and a generic DBC decoder; no validation of decoding accuracy is provided.
  • domain assumption The uniform-distribution attack model in Equation 2 captures realistic false data injection behavior.
    The authors generate attacks by drawing uniform random values within the normal signal range, but provide no evidence that real attackers use this distribution.
  • domain assumption The white-box threat model, where the adversary has full gradient access to the LSTM model, is relevant for in-vehicle adversarial settings.
    White-box attacks are common in the adversarial ML literature, but their practicality for automotive ECUs is not discussed.
  • domain assumption Latency measured on the GENI testbed is representative of a real in-vehicle SDN deployment over CAN.
    Section IV-C reports timing from a GENI experiment, but does not account for CAN bus contention, OS scheduling, or real ECU processing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SDN-Based False Data Detection With Its Mitigation and Machine Learning Robustness for In-Vehicle Networks." pith.science (2026). https://pith.science/paper/MS6QGEPE

@misc{pith2026250606556,
  author       = {Pith},
  title        = {Pith review of: SDN-Based False Data Detection With Its Mitigation and Machine Learning Robustness for In-Vehicle Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MS6QGEPE}},
  note         = {Machine review of arXiv:2506.06556}
}
read the original abstract

As the development of autonomous and connected vehicles advances, the complexity of modern vehicles increases, with numerous Electronic Control Units (ECUs) integrated into the system. In an in-vehicle network, these ECUs communicate with one another using an standard protocol called Controller Area Network (CAN). Securing communication among ECUs plays a vital role in maintaining the safety and security of the vehicle. This paper proposes a robust SDN-based False Data Detection and Mitigation System (FDDMS) for in-vehicle networks. Leveraging the unique capabilities of Software-Defined Networking (SDN), FDDMS is designed to monitor and detect false data injection attacks in real-time. Specifically, we focus on brake-related ECUs within an SDN-enabled in-vehicle network. First, we decode raw CAN data to create an attack model that illustrates how false data can be injected into the system. Then, FDDMS, incorporating a Long Short Term Memory (LSTM)-based detection model, is used to identify false data injection attacks. We further propose an effective variant of DeepFool attack to evaluate the model's robustness. To countermeasure the impacts of four adversarial attacks including Fast gradient descent method, Basic iterative method, DeepFool, and the DeepFool variant, we further enhance a re-training technique method with a threshold based selection strategy. Finally, a mitigation scheme is implemented to redirect attack traffic by dynamically updating flow rules through SDN. Our experimental results show that the proposed FDDMS is robust against adversarial attacks and effectively detects and mitigates false data injection attacks in real-time.

Figures

Figures reproduced from arXiv: 2506.06556 by the authors.

Figure 1
Figure 1. Format of a CAN data frame A. Decoding and Data Processing In this paper, we used the dataset from real a KIA SOUL in￾vechicle CAN bus [22] to thoroughly investigate the efficacy of proposed FDDMS. The dataset comprises of an attack free state data and three other data injected with denial of service attacks, fuzzy attacks, and impersonation attacks. The attack free file has 3,713,146 rows and four columns including… view at source ↗
Figure 3
Figure 3. Attack on TQI MAX signal analysis across all 20 features. We used Pearson’s Corre￾lation Coefficient to assess the linear correlation between pairs of features [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. SDN-based In-vehicle network generate false data. Here, i denotes the signal index, Vattack is the false data injected by the attacker, and V i max is the maximum value of the signal. For example, V 6 attack refers to the false data injected into signal 6, which corresponds to TQI ACOR. Vnormal represents the normal data without any attack [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 36 canonical work pages

  1. [1]

    Countermeasures against various network attacks using machine learning methods,

    Y . Li, “Countermeasures against various network attacks using machine learning methods,” Doctoral dissertation, University of South Florida, 2020

  2. [2]

    Efficient data flow algorithms for autonomous lane changing, passing and overtaking behaviors,

    U. P. Mudalige and M. Losh, “Efficient data flow algorithms for autonomous lane changing, passing and overtaking behaviors,” Aug. 4 2015, uS Patent 9,096,267

  3. [3]

    A survey of vehicle to everything (v2x) testing,

    J. Wang, Y . Shao, Y . Ge, and R. Yu, “A survey of vehicle to everything (v2x) testing,” in Sensors, vol. 19, no. 2, 2019, p. 334

  4. [4]

    In-vehicle networking: Protocols, challenges, and solutions,

    J. Huang, M. Zhao, Y . Zhou, and C.-C. Xing, “In-vehicle networking: Protocols, challenges, and solutions,” in IEEE Network, vol. 33, no. 1. IEEE, 2018, pp. 92–98

  5. [5]

    State-of-the-Art Survey on In-Vehicle Network Communication (CAN-Bus) Security and Vulnerabilities

    O. Avatefipour and H. Malik, “State-of-the-art survey on in-vehicle network communication (CAN-Bus) security and vulnerabilities,” arXiv:1802.01725, 2018

  6. [6]

    In-vehicle networks: Attacks, vulnerabilities, and proposed solutions,

    P. Carsten, T. R. Andel, M. Yampolskiy, and J. T. McDonald, “In-vehicle networks: Attacks, vulnerabilities, and proposed solutions,” in CISRC. ACM, 2015, p. 1

  7. [7]

    Software-defined networking: A comprehensive survey,

    D. Kreutz, F. M. Ramos, P. E. Verissimo, C. E. Rothenberg, S. Azodol- molky, and S. Uhlig, “Software-defined networking: A comprehensive survey,” in Proceedings of the IEEE , vol. 103, no. 1. IEEE, 2015, pp. 14–76

  8. [8]

    Cybersecurity attacks in vehicle-to-infrastructure applications and their prevention,

    M. Islam, M. Chowdhury, H. Li, and H. Hu, “Cybersecurity attacks in vehicle-to-infrastructure applications and their prevention,” in Trans- portation Research Record, vol. 2672, no. 19. SAGE, 2018, pp. 66–78

Show all 46 references
  1. [9]

    The 2015 ukraine blackout: Implications for false data injection attacks,

    G. Liang, S. R. Weller, J. Zhao, F. Luo, and Z. Y . Dong, “The 2015 ukraine blackout: Implications for false data injection attacks,” in IEEE Transactions on Power Systems, vol. 32, no. 4. IEEE, 2016, pp. 3317– 3318

  2. [10]

    ML attack models: Ad- versarial attacks and data poisoning attacks,

    J. Lin, L. Dang, M. Rahouti, and K. Xiong, “ML attack models: Ad- versarial attacks and data poisoning attacks,” arXiv:2112.02797, 2021

  3. [11]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” in ICLR, 2015

  4. [12]

    Adversarial examples in the physical world,

    A. Kurakin, I. J. Goodfellow, and S. Bengio, “Adversarial examples in the physical world,” arXiv:1607.02533, 2016

  5. [13]

    Deepfool: A simple and accurate method to fool deep neural networks,

    S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard, “Deepfool: A simple and accurate method to fool deep neural networks,” in CVPR, 2016, pp. 2574–2582

  6. [14]

    Investigating the impact of evasion attacks against automotive intrusion detection systems,

    P. Cerracchio, S. Longari, M. Carminati, S. Zanero et al., “Investigating the impact of evasion attacks against automotive intrusion detection systems,” in Symposium on Vehicle Security and Privacy , 2024

  7. [15]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” arXiv:1706.06083, 2017

  8. [16]

    Adversarial examples are not bugs, they are features,

    A. Ilyas, S. Santurkar, D. Tsipras, L. Engstrom, B. Tran, and A. Madry, “Adversarial examples are not bugs, they are features,” in NeurIPS, vol. 32, 2019

  9. [17]

    Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples,

    A. Athalye, N. Carlini, and D. Wagner, “Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples,” in ICML, 2018, pp. 274–283

  10. [18]

    Robust machine learning against adversarial samples at test time,

    J. Lin, L. L. Njilla, and K. Xiong, “Robust machine learning against adversarial samples at test time,” in IEEE International Conference on Communications, 2020, pp. 1–6

  11. [19]

    Smooth adversarial training,

    C. Xie, M. Tan, B. Gong, A. Yuille, and Q. V . Le, “Smooth adversarial training,” arXiv:2006.14536, 2020

  12. [20]

    Theoretically principled trade-off between robustness and accuracy,

    H. Zhang, Y . Yu, J. Jiao, E. Xing, L. El Ghaoui, and M. Jordan, “Theoretically principled trade-off between robustness and accuracy,” in ICML, 2019, pp. 7472–7482

  13. [21]

    An adversarial attack defending system for securing in-vehicle networks,

    Y . Li, J. Lin, and K. Xiong, “An adversarial attack defending system for securing in-vehicle networks,” in IEEE Consumer Communications & Networking Conference , 2021, pp. 1–6

  14. [22]

    Otids: A novel intrusion detection system for in-vehicle network by using remote frame,

    H. Lee, S. H. Jeong, and H. K. Kim, “Otids: A novel intrusion detection system for in-vehicle network by using remote frame,” in PST. IEEE, 2017, pp. 57–5709

  15. [23]

    In-vehicle network attacks and countermeasures: Challenges and future directions,

    J. Liu, S. Zhang, W. Sun, and Y . Shi, “In-vehicle network attacks and countermeasures: Challenges and future directions,” in IEEE Network , vol. 31, no. 5. IEEE, 2017, pp. 50–58

  16. [24]

    A survey on security attacks and defense techniques for connected and autonomous vehicles,

    M. Pham and K. Xiong, “A survey on security attacks and defense techniques for connected and autonomous vehicles,” Computers & Security, vol. 109, p. 102269, 2021

  17. [25]

    A structured approach to anomaly detection for in-vehicle networks,

    M. M ¨uter, A. Groll, and F. C. Freiling, “A structured approach to anomaly detection for in-vehicle networks,” in IAS. IEEE, 2010, pp. 92–98

  18. [26]

    Entropy-based anomaly detection for in-vehicle networks,

    M. M ¨uter and N. Asaj, “Entropy-based anomaly detection for in-vehicle networks,” in IEEE IV. IEEE, 2011, pp. 1110–1115

  19. [27]

    Potential sources of sensor data anomalies for autonomous vehicles: An overview from road vehicle safety perspective,

    X. Zhao, Y . Fang, H. Min, X. Wu, W. Wang, and R. Teixeira, “Potential sources of sensor data anomalies for autonomous vehicles: An overview from road vehicle safety perspective,” Expert Systems with Applications, vol. 236, p. 121358, 2024

  20. [28]

    Anomaly detection in connected and autonomous vehicles: A survey, analysis, and research challenges,

    S. Baccari, M. Hadded, H. Ghazzai, H. Touati, and M. Elhadef, “Anomaly detection in connected and autonomous vehicles: A survey, analysis, and research challenges,” IEEE Access, 2024

  21. [29]

    Anomaly diagnosis of connected autonomous vehicles: A survey,

    Y . Fang, H. Min, X. Wu, W. Wang, X. Zhao, B. Martinez-Pastor, and R. Teixeira, “Anomaly diagnosis of connected autonomous vehicles: A survey,” Information Fusion, vol. 105, p. 102223, 2024

  22. [30]

    Fingerprinting electronic control units for vehicle intrusion detection,

    K.-T. Cho and K. G. Shin, “Fingerprinting electronic control units for vehicle intrusion detection,” in U SEN IXSecurity Symposium, 2016, pp. 911–927

  23. [31]

    Intrusion detection system using deep neural network for in-vehicle network security,

    M.-J. Kang and J.-W. Kang, “Intrusion detection system using deep neural network for in-vehicle network security,” in PloS one , vol. 11, no. 6, 2016, p. e0155781

  24. [32]

    Supervised and unsupervised intrusion detection based on can message frequencies for in-vehicle network,

    T. Kuwahara, Y . Baba, H. Kashima, T. Kishikawa, J. Tsurumi, T. Haga, Y . Ujiie, T. Sasaki, and H. Matsushima, “Supervised and unsupervised intrusion detection based on can message frequencies for in-vehicle network,” in Journal of Information Processing, vol. 26, 2018, pp. 306– 313

  25. [33]

    False data injection attacks against state estimation in electric power grids,

    Y . Liu, P. Ning, and M. K. Reiter, “False data injection attacks against state estimation in electric power grids,” in ACM TISSEC, vol. 14, no. 1. ACM, 2011, p. 13

  26. [34]

    A comprehensive survey of false data injection in smart grid,

    Z. Guan, N. Sun, Y . Xu, and T. Yang, “A comprehensive survey of false data injection in smart grid,” in International Journal of Wireless and Mobile Computing , vol. 8, no. 1. Inderscience Publishers, 2015, pp. 27–33

  27. [35]

    False data injection attack and its countermeasures in wireless sensor networks,

    S. A. Jeba and B. Paramasivan, “False data injection attack and its countermeasures in wireless sensor networks,” in European Journal of Scientific Research, vol. 82, no. 2, 2012, pp. 248–257

  28. [36]

    Proof-of-relevance: Filtering false data via authentic consensus in vehicle ad-hoc networks,

    Z. Cao, J. Kong, U. Lee, M. Gerla, and Z. Chen, “Proof-of-relevance: Filtering false data via authentic consensus in vehicle ad-hoc networks,” in IEEE INFOCOM Workshops. IEEE, 2008, pp. 1–6

  29. [37]

    Modeling inter-signal arrival times for accurate detection of can bus signal injection attacks: a data-driven approach to in-vehicle intrusion detection,

    M. R. Moore, R. A. Bridges, F. L. Combs, M. S. Starr, and S. J. Prowell, “Modeling inter-signal arrival times for accurate detection of can bus signal injection attacks: a data-driven approach to in-vehicle intrusion detection,” in The Annual Conference on Cyber and Informatio...

  30. [38]

    SDVN: enabling rapid network innovation for heterogeneous vehicular communication,

    Z. He, J. Cao, and X. Liu, “SDVN: enabling rapid network innovation for heterogeneous vehicular communication,” in IEEE network, vol. 30, no. 4. IEEE, 2016, pp. 10–15

  31. [39]

    Ml-based approach to detect ddos attack in v2i communication under sdn architecture,

    P. K. Singh, S. K. Jha, S. K. Nandi, and S. Nandi, “Ml-based approach to detect ddos attack in v2i communication under sdn architecture,” in TENCON. IEEE, 2018, pp. 0144–0149

  32. [40]

    In-vehicle false information attack detection and mitigation framework using machine learning and software defined networking,

    Z. Khan, M. Chowdhury, M. Islam, C.-Y . Huang, and M. Rah- man, “In-vehicle false information attack detection and mitigation framework using machine learning and software defined networking,” arXiv:1906.10203, 2019

  33. [41]

    OpenDBC,

    CommaAI, “OpenDBC,” [Online]. 2019, Available: https://github.com/ commaai/opendbc

  34. [42]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville, Deep Learning. MIT Press, 2016, http://www.deeplearningbook.org

  35. [43]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014

  36. [44]

    GENI: A federated testbed for innovative network experiments,

    B. Mark et al. , “GENI: A federated testbed for innovative network experiments,” in Computer Networks, 2014

  37. [45]

    Project Floodlight,

    Floodlight, “Project Floodlight,” [Online]. 2012, Available: http://www. projectfloodlight.org/

  38. [46]

    Adversarial robustness toolbox v1.2.0,

    M.-I. Nicolae, M. Sinn, M. N. Tran, B. Buesser, A. Rawat, M. Wistuba, V . Zantedeschi, N. Baracaldo, B. Chen, H. Ludwig, I. Molloy, and B. Edwards, “Adversarial robustness toolbox v1.2.0,” arXiv:1807.01069, 2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.