Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Communication-Aware Quantum-Inspired Reinforcement Learning for Cyber-Resilient V2X Intrusion Detection and Mitigation

T0 review · 4 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A lightweight quantum-inspired deep Q-network can detect IoV intrusions at competitive accuracy while staying under 50 microseconds per sample—fast enough for vehicle safety windows.

desk verdict Solid latency-first small-DQN systems result on NIDS data; the quantum-inspired modules are not shown to be load-bearing, and the abstract/body framing do not match. read the letter →

arxiv 2606.07804 v2 pith:MF5ZN4GP submitted 2026-06-05 cs.CR

Sajid Anwer , Rohan Farooq , Anwar Shah , Tallha Akram This is my paper · ORCID
classification cs.CR
keywords InternetofVehiclesintrusiondetectionreinforcementlearningquantum-inspiredcomputingDeepQ-Networkedgelatencysmartcitiescost-sensitive
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

Connected vehicles and smart-city traffic systems need intrusion detection that is both accurate and fast enough to act inside sub-millisecond safety budgets. This paper argues that existing ensemble detectors hit high accuracy but take milliseconds per flow, while smaller models struggle to adapt to multi-stage, imbalanced attacks. It introduces QIRL, a compact deep Q-network that treats detection as a cost-sensitive sequential decision problem and adds three quantum-inspired pieces: polar amplitude-phase encoding of features, rotation-gate action exploration, and a cosine-similarity interference term that stabilizes learning. On CICIDS2017 and UNSW-NB15, under train-only SMOTE and validation-chosen thresholds, the agent reports about 97.9% and 91% accuracy with True Skill Statistics well above chance, and inference of 32.5 and 45.7 microseconds per sample—tens of times faster than reimplemented ensemble baselines. A sympathetic reader cares because delayed detection in autonomous braking and steering is a physical-safety problem, not only a cybersecurity metric.

What carries the argument

QIRL: a lightweight Deep Q-Network whose Bellman target is augmented by a Quantum Interference Module (cosine similarity of consecutive latent states, scaled by Λ) and whose exploration uses per-action rotation angles, trained under a cost-sensitive MDP reward that heavily penalizes missed attacks.

What would settle it

Re-run the same train-only SMOTE, cost weights, and tiny three-layer DQN with those three quantum modules fully disabled; if accuracy, TSS, and sub-50 μs latency remain statistically unchanged on the same CICIDS2017 and UNSW-NB15 splits, the claim that the quantum-inspired machinery is necessary for the dual objective fails.

Watch

Extended reading notes

Core claim

The paper claims that embedding amplitude-phase state encoding, rotation-based exploration, and interference-augmented rewards inside a small DQN, together with asymmetric false-negative/false-positive costs and training-only SMOTE, lets one agent jointly meet competitive detection robustness and sub-millisecond inference on standard network intrusion benchmarks relevant to IoV, closing a dual accuracy-latency gap that prior static and ensemble detectors do not simultaneously satisfy.

Load-bearing premise

The dual accuracy-plus-speed result is assumed to depend on the quantum-inspired encoding, rotation exploration, and interference reward, rather than mainly on the tiny network, cost-sensitive rewards, and careful train-only balancing.

Editorial extensions

If this is right

  • Edge nodes on vehicles or RSUs could run intrusion decisions inside the same sub-millisecond budget as braking and steering loops without waiting on cloud ensembles.
  • Cost-sensitive sequential policies become a practical alternative to static packet classifiers when multi-stage attacks and severe class imbalance dominate IoV traffic.
  • Training-only SMOTE plus validation-chosen thresholds becomes a required reporting baseline so that reported IoV IDS accuracy is not inflated by leakage.
  • Latency, not only accuracy, becomes a first-class published metric for any IDS claimed to protect Level 4/5 autonomous driving.

Reading between the lines

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

  • If the quantum modules add little beyond the tiny net and cost shaping, the same dual objective might be reachable with ordinary DQN plus asymmetric rewards—suggesting a simpler production path for vehicle ECUs.
  • Extending the same cost-sensitive MDP to true multi-agent V2V/V2I settings could expose whether interference rewards still stabilize learning when neighboring vehicles are also attackers or sensors.
  • Adversarial packet crafting aimed at the learned Q-boundary would be a natural next stress test, because the paper defers model and physical-layer attacks.
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

4 major / 6 minor

Summary. The manuscript proposes QIRL, a lightweight Deep Q-Network with three quantum-inspired components (amplitude-phase feature encoding, rotation-gate exploration, and cosine-similarity interference on consecutive latent states) plus cost-sensitive rewards and training-only SMOTE, framed as a cost-sensitive MDP for IoV intrusion detection. The central claim is that this architecture simultaneously satisfies competitive detection robustness and sub-millisecond inference (Eq. 1): 97.89% accuracy / TSS 0.9443 / AUC 0.9945 at 32.5 μs on CICIDS2017 and 91.04% / TSS 0.8244 / AUC 0.9713 at 45.7 μs on UNSW-NB15, with large speedups over ensemble baselines. Evaluation uses a careful train/val/test protocol, single-sample CPU latency measurement, ablations, comparative tables, and SHAP/PCA/t-SNE explainability.

Significance. If the dual accuracy–latency result is solid under realistic IoV constraints, the work is practically significant: sub-50 μs CPU inference with competitive TSS/AUC would be deployable on edge nodes where ensemble IDS methods exceed the ~1 ms safety window. Strengths include an explicit dual objective (Eq. 1), training-only SMOTE to avoid leakage, single-sample latency protocol, asymptotic complexity analysis showing QIM adds no inference cost, and a structured XAI section. The quantum-inspired modules are presented as the distinctive technical contribution; their necessity for the dual claim is therefore load-bearing for the paper’s novelty relative to a small cost-sensitive DQN.

major comments (4)
  1. §7.4.4, Tables 4–5: The ablations do not establish that the quantum-inspired modules are necessary for the dual objective in Eq. 1. On CICIDS2017, Cost-Sensitive Only already reaches Acc 0.9793 / TSS 0.9558 while Full QIRL is 0.9762 / 0.9586; on UNSW-NB15, Baseline DQN (no quantum modules, symmetric reward) records Acc 0.9434 / TSS 0.8597, strictly better than Full QIRL’s 0.9388 / 0.8094. Latencies remain ~30–50 μs across all variants, so the reported 51–68× speedups are explained by the tiny 64-wide MLP, not by encoding/rotation/interference. The manuscript should either (i) show statistically that Full QIRL improves the constrained objective (TSS subject to τ < 1 ms) over the best simpler variant under matched seeds/thresholds, or (ii) reframe the contribution as a lightweight cost-sensitive DQN and demote the quantum machinery to optional regularisation.
  2. Table 3 vs Tables 4–5: Main-text CICIDS2017 results (Acc 97.89%, TSS 0.9443) do not match the Full QIRL ablation row (Acc 0.9762, TSS 0.9586); UNSW main-text Acc 91.04% / TSS 0.8244 likewise diverges from ablation Full (0.9388 / 0.8094) and from Baseline (0.9434 / 0.8597). Different episode counts (20 vs 10), thresholds, or SMOTE settings appear to be mixed without reconciliation. Please report a single protocol (episodes, τ*, seeds, SMOTE) for main and ablation tables, with means±std over multiple runs, so the dual-claim numbers are reproducible and comparable.
  3. §1–§5 and §7.1: The paper is framed as V2X/IoV cyber defense with multi-stage temporal attacks and edge safety constraints, yet evaluation uses only CICIDS2017 (DDoS-focused flows) and UNSW-NB15 under binary (or effectively binary) classification of general network traffic. The MDP state is an encoded packet-flow vector and the action space is binary detect/benign (§5.2); there is no mobility, RSU load, channel busy ratio, or mitigation action beyond classification. Either add a vehicular/CAN or mobility-aware evaluation that exercises the claimed IoV threat model (§4), or narrow the claims to low-latency NIDS with IoV as motivation rather than demonstrated deployment setting.
  4. §5.3, Eqs. (11)–(12) and §7.3.3: The Quantum Interference Module adds Λ·cos_sim(ψ_t, ψ_{t+1}) to the Bellman target with fixed Λ=0.1 and no extra parameters. Given that Interference Only does not dominate Baseline on either dataset (Tables 4–5), the paper needs a clearer causal account—e.g., learning curves, TD-error variance, or sensitivity over Λ—showing when interference stabilises training versus when it is neutral. Without that, the quantum analogy remains branding rather than a demonstrated mechanism for the dual constraint.
minor comments (6)
  1. Figure 6 caption notes AUC 0.9712 vs main-text 0.9713 and mentions an “initial model” ROC; align figure, caption, and Table 3 to the final thresholded model.
  2. §5.2 reward weights: the text states both w_fn=5.0, w_fp=1.0 and w_fn=1.5, w_fp=1.0 in one sentence; clarify dataset-specific assignment as later done in §7.3.3.
  3. Table 6: several baselines lack latency; the independent re-measurements (§, ‡) are useful—state hardware and code parity more explicitly so readers can reproduce the 2248 μs ensemble figure.
  4. Notation: QIRL vs “Deep QIRL” in figures; ψ for both latent embeddings and interference similarity; standardise.
  5. Minor language/typos throughout (e.g., “not but fixed”, “TrueSkillStatisticsof”); a copy-edit pass would help.
  6. §6 complexity: d∈{40} after feature selection is stated, but §7.3.2 does not detail which 40 features or the selection method; add a short note for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation circularity: QIRL is an empirical systems paper whose accuracy/latency claims are measured on external benchmarks, not forced by construction from fitted inputs or self-citation.

full rationale

The paper’s load-bearing claims are empirical measurements (accuracy, F1, TSS, AUC-ROC, single-sample CPU latency, speedup vs re-implemented ensembles) on public CICIDS2017 and UNSW-NB15 under a stated train/val/test protocol with training-only SMOTE and validation-chosen thresholds τ*. That is standard ML evaluation, not a first-principles derivation that reduces to its inputs. Amplitude-phase encoding, rotation-angle exploration, and the Λ·cos_sim interference term are architectural choices that modify the DQN training objective; they do not algebraically define the reported test metrics. Cost weights and τ* are tuned on validation, which does not make test-set numbers tautological. Self-citations (e.g., Ahmad/Shah works on GraphGuard and surveys) are peripheral and not used as uniqueness theorems that force the dual accuracy–latency result. Ablation inconsistencies (quantum modules not clearly necessary) are a necessity/correctness issue, not circularity. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain was found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central dual accuracy-latency claim rests on standard RL/MDP formalism, classical neural nets, and several hand-chosen hyperparameters and modeling choices branded as quantum-inspired. No new physical entity is required; the invented pieces are algorithmic modules without independent quantum evidence. Free parameters (cost weights, Λ, rotation step, architecture, thresholds) are tuned to validation performance and materially affect reported operating points.

free parameters (5)
  • False-negative / false-positive cost weights (w_fn, w_fp) = CICIDS: 5.0/1.0; UNSW: 1.5/1.0
    Set to (5.0, 1.0) on CICIDS2017 and (1.5, 1.0) on UNSW-NB15 after SMOTE; directly shapes the policy and chosen threshold.
  • Interference coefficient Λ (λ) = 0.1
    Scales cosine similarity added to the Bellman target; fixed at 0.1 without principled derivation.
  • Rotation step Δθ / δ = 0.05
    Controls how fast action angles move with TD-error sign; set to 0.05.
  • Decision thresholds τ* = 0.5 and 0.7
    Chosen on validation to maximize TSS (0.5 / 0.7 reported), so headline metrics depend on post-training threshold search.
  • Network width/depth and training schedule = L=3, H=64, 20 episodes, batch 512
    Two hidden layers of 64, batch 512, 20 episodes, target sync every 200 steps, Adam 1e-3—capacity and compute budget that largely explain sub-50 μs latency.
assumptions (5)
  • domain assumption Intrusion detection can be cast as a finite-horizon/binary-action MDP with discounted return G_t = Σ γ^k r_{t+k} and Bellman optimality for Q.
    Problem formulation §3 and MDP §5.2; standard RL assumption that sequential rewards capture multi-stage attacks better than i.i.d. classification.
  • domain assumption Train-only SMOTE plus asymmetric misclassification costs adequately correct severe class imbalance without distorting test risk.
    Stated in contributions and §5.1/§7.3; methodological choice that drives UNSW operating point.
  • ad hoc to paper Polar re-encoding of feature pairs and cosine similarity of consecutive latents behave like useful quantum amplitude/interference priors for learning.
    §5.1–5.3 introduce these as quantum-inspired without a theorem linking them to sample complexity or optimality beyond analogy to Ry gates and amplitude interference.
  • domain assumption CICIDS2017 and UNSW-NB15 flows are sufficiently representative of IoV/V2X safety-critical traffic for the dual latency-accuracy claim.
    Evaluation §7.1 frames these general NIDS sets as the main evidence for vehicular deployment readiness.
  • standard math Standard DQN with replay, target network, and gradient clipping converges adequately under the asymmetric reward.
    Training pipeline Algorithm 1 and complexity §6 rely on ordinary DQN practice.
invented entities (3)
  • Quantum Interference Module (QIM)
    purpose: Augment Bellman target with Λ·cos_sim(ψ(s_t), ψ(s_{t+1})) to stabilize training without extra parameters at inference.
    Defined in §5.3; classical cosine similarity renamed via quantum analogy; no independent quantum experiment or hardware evidence.
  • Quantum Rotation Gate exploration
    purpose: Replace ε-greedy with per-action angles θ_a and P(a)∝cos²(θ_a) updated by TD-error sign.
    §5.2; adaptive categorical policy with trigonometric parameterization, not an actual quantum gate implementation.
  • Amplitude-Phase Quantum Encoding
    purpose: Map consecutive feature pairs to (r cos θ, r sin θ) as a geometric prior.
    §5.1; lossless polar reparameterization of classical features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Communication-Aware Quantum-Inspired Reinforcement Learning for Cyber-Resilient V2X Intrusion Detection and Mitigation." pith.science (2026). https://pith.science/paper/MF5ZN4GP

@misc{pith2026260607804,
  author       = {Pith},
  title        = {Pith review of: Communication-Aware Quantum-Inspired Reinforcement Learning for Cyber-Resilient V2X Intrusion Detection and Mitigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MF5ZN4GP}},
  note         = {Machine review of arXiv:2606.07804}
}
read the original abstract

Smart cities rely on Internet of Vehicles (IoV) networks for critical services. However, this vast connectivity enlarges the attack surface, exposing vehicular systems to evolving cyber threats. Conventional static defenses struggle to autonomously adapt to these dynamic, multi-stage intrusions. To address this, we propose the Communication Aware Quantum Inspired Reinforcement Learning (CA-QIRL) framework, built on a lightweight deep Q-Network architecture for autonomous cyber defense. V2X defense is formulated as a communication-aware Markov Decision Process (MDP). The agent observes intrusion, mobility, Road Side Unit (RSU), and communication metrics to select optimal mitigation actions. CA-QIRL integrates quantum-inspired encoding, rotation exploration, and an interference reward, combined with a cost function penalizing false negatives, false positives, delay, packet loss, and RSU overload. Experimental evaluations on vehicular intrusion datasets and a mobility-aware V2X simulation demonstrate robust performance. CA-QIRL achieves competitive detection accuracies of 97.89% on CICIDS2017 and 80.31% on CAN-MIRGU, outperforming state-of-the-art ensemble methods in inference latency. Furthermore, end-to-end delay and Channel Busy Ratio (CBR) drop by up to 95.7% and 90%. Statistical significance is confirmed on ROAD and VeReMi. These findings establish CA-QIRL as a highly practical and resilient defense mechanism for next-generation V2X and IoV networks.

Figures

Figures reproduced from arXiv: 2606.07804 by the authors.

Figure 1
Figure 1. The critical trade-off between detection accuracy and inference latency in current IoV security frameworks. Existing ensemble methods occupy the high-accuracy but unsafe latency zone (>1 ms), while the proposed QIRL framework simultaneously satisfies both constraints. and adversarial model perturbations are outside the current scope and are deferred to future work. Attacks within scope are modeled as non-stationary … view at source ↗
Figure 2
Figure 2. QIRL problem formulation. IoV traffic exhibits non-stationary temporal behavior. Existing IDS methods fail due to latency–accuracy trade-offs, class imbalance, and ignored temporal correlation. QIRL addresses these using a cost-sensitive MDP with quantum interference for efficient and robust detection. Each state vector is then transformed via Amplitude-Phase Quantum Encoding: for each consecutive feature pair (𝑥2𝑖 … view at source ↗
Figure 3
Figure 3. IoV threat model. The adversary A = (K, O, G, P, T) operates over a multi-layer attack surface (V2V, V2I, edge, cloud). Detection failure occurs as false negatives, while attacks evolve as a non-stationary temporal process, motivating adaptive QIRL-based defense. rotations suppress low-reward ones. To handle class imbalance, we calculate the immediate reward using a cost￾sensitive Weighted Reward Function with 𝑤fn =… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Proposed QIRL framework utilizing quantum interference to enhance adaptive cyber defense in IoV environments. Gradient descent is used to minimize the squared error between the current Q-value estimate 𝑄(𝑠𝑡 , 𝑎𝑡 ; 𝜃) and the target 𝑌 when updating the primary network p…
Figure 5
Figure 5. Figure 5: Training convergence of the QIRL agent over 20 episodes, showing rapid stabilization on CICIDS2017 (blue) versus high-variance learning on the complex UNSW-NB15 dataset (orange; test AUC-ROC: 0.9713). 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 …
Figure 6
Figure 6. Figure 6: Receiver Operating Characteristic (ROC) Analysis. (a) The agent achieves near-perfect separation on volumetric DDoS attacks (AUC=0.9945). (b) On the complex UNSW-NB15 dataset, the agent maintains robust discriminatory power. Note: the ROC curve shown reflects the initi…
Figure 7
Figure 7. Figure 7: Confusion Matrix Analysis. (a) On CICIDS2017, the model achieves 97.89% accuracy with only 4,434 missed attack packets (FNR=0.9%). (b) On UNSW-NB15 at threshold 𝜏 ∗ = 0.7, the model achieves 91.04% accuracy with 4,800 missed attacks (FNR=10.6%) and FPR of 6.97%, a bala…
Figure 8
Figure 8. Figure 8: SHAP feature attribution comparison. CICIDS2017 emphasizes traffic volume and rate features with a wide SHAP range [−50,+35], indicating a sharp DDoS-oriented decision boundary. UNSW-NB15 is driven by TCP state and timing features with a narrower range [−3,+3.5], refle…
Figure 9
Figure 9. Figure 9: PCA and t-SNE Geometric Analysis CICIDS2017. PCA (PC1=46.2%, PC2=19.4%) shows data concentrated near the origin. The t-SNE projection reveals clearly separated local clusters with boundary mixing confined to cluster edges, corresponding to the model’s false positive an…
Figure 10
Figure 10. Figure 10 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 1 linked inside Pith

  1. [1]

    Graphguard:Anadaptiveapproachforrestoringaccuracyinbackdoor-compromisedgnns

    Ahmad,A.,Shah,A.,Alnumay,W.,Ali,B.,2025a. Graphguard:Anadaptiveapproachforrestoringaccuracyinbackdoor-compromisedgnns. Neural Networks 165, 107990

  2. [2]

    Acomprehensivesurveyontheconvergenceofblockchain, digital twins, and metaverse: Shaping the future of cybersecurity frameworks

    Ahmad,A.,Shah,A.,Alnumay,W.S.,Adnan,M.,Anwer,S.,Zaman,Q.U.,2025b. Acomprehensivesurveyontheconvergenceofblockchain, digital twins, and metaverse: Shaping the future of cybersecurity frameworks. Computers and Electrical Engineering 126, 110486

  3. [3]

    Artificial intelligence for the internet of vehicles: Opportunities and challenges

    Al-Fuqaha, A., Guizani, M., Mohammadi, M., 2020. Artificial intelligence for the internet of vehicles: Opportunities and challenges. IEEE Communications Surveys & Tutorials 22, 2344–2374

    2020

  4. [4]

    Intelligent intrusion detection in external communication systems for autonomous vehicles

    Alheeti, K.M.A., McDonald-Maier, K., 2018. Intelligent intrusion detection in external communication systems for autonomous vehicles. Systems Science & Control Engineering 6, 48–56

    2018

  5. [5]

    Classification approach for intrusion detection in vehicle systems

    Alshammari, A., Zohdy, M.A., Debnath, D., Corser, G., 2018. Classification approach for intrusion detection in vehicle systems. Wireless Engineering and Technology 9, 79–94

    2018

  6. [6]

    Machine-learning-basedefficientandsecurersuplacementmechanismforsoftware-defined-iov

    Anbalagan,S.,etal.,2021. Machine-learning-basedefficientandsecurersuplacementmechanismforsoftware-defined-iov. IEEEInternetof Things Journal 8, 13950–13957

    2021

  7. [7]

    IEEETransactionsonIntelligentTransportationSystems22,4507– 4518

    Ashraf,J.,Bakhshi,A.D.,Moustafa,N.,Khurshid,H.,Javed,A.,Beheshti,A.,2021.Noveldeeplearning-enabledlstmautoencoderarchitecture fordiscoveringanomalouseventsfromintelligenttransportationsystems. IEEETransactionsonIntelligentTransportationSystems22,4507– 4518

    2021

  8. [8]

    Smote: Synthetic minority over-sampling technique

    Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P., 2002. Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research 16, 321–357

    2002

Show all 29 references
  1. [9]

    An intelligent intrusion detection system in IoV using machine learning and deep learning models

    Dasari, S., Menasinkai, V.K., 2025. An intelligent intrusion detection system in IoV using machine learning and deep learning models. International Journal of Communication Systems Early Access/In Press

    2025

  2. [10]

    Quantumreinforcementlearning

    Dong,D.,Petersen,I.R.,2008. Quantumreinforcementlearning. IEEETransactionsonSystems,Man,andCybernetics,PartB(Cybernetics) 38, 1207–1220. First Author et al.:Preprint submitted to ElsevierPage 20 of 21 QIRL

    2008

  3. [11]

    Reliable data transmission for a vanet-ioit architecture: A dnn approach

    Ghosh, J., Kumar, N., Al-Utaibi, K.A., Sait, S.M., Vo, V.N., So-In, C., 2024. Reliable data transmission for a vanet-ioit architecture: A dnn approach. Internet of Things 25, 101129. URL:https://www.sciencedirect.com/science/article/pii/S2542660524000714, doi:https://doi.org/1...

    2024 doi

  4. [12]

    Ensemble learning approaches for multi-class intrusion detection systems for the Internet of Vehicles: A comprehensive survey

    Hosen, A.S.M.S., Akhter, S., Ra, I.H., 2025. Ensemble learning approaches for multi-class intrusion detection systems for the Internet of Vehicles: A comprehensive survey. Future Internet 17, 317

    2025

  5. [13]

    Multi-stageoptimizedmachinelearningframeworkfornetworkintrusiondetection

    Injadat,M.,Moubayed,A.,Nassif,A.B.,Shami,A.,2021. Multi-stageoptimizedmachinelearningframeworkfornetworkintrusiondetection. IEEE Transactions on Network and Service Management 18, 1803–1816

    2021

  6. [14]

    Milccde: A metaheuristic improved decision-based ensemble framework for intrusion detection in autonomous vehicles, in: Proceedings of SCIS 2024, Springer, Singapore

    Jha, J., Yadav, J., Naqvi, H., 2025. Milccde: A metaheuristic improved decision-based ensemble framework for intrusion detection in autonomous vehicles, in: Proceedings of SCIS 2024, Springer, Singapore

    2025

  7. [15]

    Intrusiondetectionsystemusingdeepneuralnetworkforin-vehiclenetworksecurity

    Kang,M.J.,Kang,J.W.,2016. Intrusiondetectionsystemusingdeepneuralnetworkforin-vehiclenetworksecurity. PLoSONE11,e0155781

    2016

  8. [16]

    Asecureandefficientdeeplearning-basedintrusiondetectionframeworkfortheInternetofVehicles

    Khonde,S.R.,Ulagamuthalvi,V.,2025. Asecureandefficientdeeplearning-basedintrusiondetectionframeworkfortheInternetofVehicles. Scientific Reports 15

    2025

  9. [17]

    Stackedsparseautoencodersbasedoutlierdiscoveryforin-vehiclecontroller area network (CAN)

    Lokman,S.F.,Othman,A.T.,Bakar,M.H.A.,Razuwan,R.,2018. Stackedsparseautoencodersbasedoutlierdiscoveryforin-vehiclecontroller area network (CAN). International Journal of Engineering & Technology 7, 375–380

    2018

  10. [18]

    Ciciov2024: Advancing realistic ids approaches against dos and spoofing attack in iov can bus

    Neto, E.C.P., Taslimasa, H., Dadkhah, S., Iqbal, S., Xiong, P., Rahman, T., Ghorbani, A.A., 2024. Ciciov2024: Advancing realistic ids approaches against dos and spoofing attack in iov can bus. Internet of Things 26, 101209. URL:https://www.sciencedirect.com/ science/article/pi...

    2024 doi

  11. [19]

    Deepreinforcementlearningforcybersecurity

    Nguyen,T.T.,Reddi,V.J.,2018. Deepreinforcementlearningforcybersecurity. IEEETransactionsonNeuralNetworksandLearningSystems 29, 2392–2405

    2018

  12. [20]

    Data driven intrusion detection for intelligent internet of vehicles: A deep convolutional neural network-based method

    Nie, L., Ning, Z., Wang, X., Hu, X., Cheng, J., Li, Y., 2020. Data driven intrusion detection for intelligent internet of vehicles: A deep convolutional neural network-based method. IEEE Transactions on Network Science and Engineering 7, 2219–2230

    2020

  13. [21]

    An integrated ids for the internet of vehicles using a large language model framework

    R., A., V., V., Srinivas, N., A., A.M., 2025. An integrated ids for the internet of vehicles using a large language model framework. Internet of Things 33, 101666. URL:https://www.sciencedirect.com/science/article/pii/S2542660525001805, doi:https: //doi.org/10.1016/j.iot.2025.101666

    2025 doi

  14. [22]

    Ascalablehierarchicalintrusiondetectionsystemforinternetofvehicles

    Uddin,M.A.,Chu,N.H.,Rafeh,R.,Barika,M.,2025. Ascalablehierarchicalintrusiondetectionsystemforinternetofvehicles. arXivpreprint arXiv:2505.16215

    2025 arXiv

  15. [23]

    An ensemble-based hybrid model for the detection of attacks in the internet of vehicular things

    Ullah, I., Khalil, I., Bai, X., Garg, S., Kaddoum, G., Shamim, M., 2025. An ensemble-based hybrid model for the detection of attacks in the internet of vehicular things. IEEE Transactions on Intelligent Transportation Systems 26, 17914

    2025

  16. [24]

    An explainable three-way neural network approach for intrusion detection in iot ecosystem

    Wahab, F., Ma, S., Zhao, Y., Shah, A., 2025. An explainable three-way neural network approach for intrusion detection in iot ecosystem. Internet of Things 33, 101722

    2025

  17. [25]

    Ansdn-basedhybrid-dl-drivencognitiveintrusiondetectionsystemforiotecosystem

    Wahab,F.,Shah,A.,Khan,I.,Ali,B.,Adnan,M.,2024. Ansdn-basedhybrid-dl-drivencognitiveintrusiondetectionsystemforiotecosystem. Computers and Electrical Engineering 119, 109545

    2024

  18. [26]

    Performance comparison of explainable dqn and ddpg models for cooperative lane change decision-making in multi- intelligent industrial iot vehicles

    bai ZHAN, H., 2025. Performance comparison of explainable dqn and ddpg models for cooperative lane change decision-making in multi- intelligent industrial iot vehicles. Internet of Things 31, 101552. URL:https://www.sciencedirect.com/science/article/pii/ S2542660525000654, doi...

    2025 doi

  19. [27]

    Fair federated learning for multi-task 6g nwdaf network anomaly detection

    Zhang, C., Shan, G., Roh, B.H., 2024. Fair federated learning for multi-task 6g nwdaf network anomaly detection. IEEE Transactions on Intelligent Transportation Systems doi:10.1109/TITS.2024.3461679. early Access

    2024 doi

  20. [28]

    Intrusion detection for iot based on improved genetic algorithm and deep belief network

    Zhang, Y., Li, P., Wang, X., 2019. Intrusion detection for iot based on improved genetic algorithm and deep belief network. IEEE Access 7, 31711–31722

    2019

  21. [29]

    Taskoffloadingschemeofvehicularcloudedgecomputingbasedondigitaltwinandimproveda3c

    Zhu,L.,Tan,L.,2024. Taskoffloadingschemeofvehicularcloudedgecomputingbasedondigitaltwinandimproveda3c. InternetofThings 26, 101192. URL:https://www.sciencedirect.com/science/article/pii/S2542660524001331, doi:https://doi.org/10. 1016/j.iot.2024.101192. First Author et al.:Prep...

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.