Pith. sign in

REVIEW 3 major objections 5 minor 35 references

A Joint Reconstruction-Triplet Loss Autoencoder Approach Towards Unseen Attack Detection in IoV Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An autoencoder trained only on benign network flows detects previously unseen attacks in IoV networks, with roughly 99% benign accuracy and 97–100% anomaly accuracy.

desk verdict The benign-only autoencoder plus triplet loss is a known combination applied to new IoT datasets; the headline zero-day numbers come from test-set hyperparameter selection, so the core claim is not actually tested. read the letter →

arxiv 2505.21703 v1 pith:YSKEPM3Y submitted 2025-05-27 cs.CR cs.AIcs.NI

Julia Boone , Tolunay Seyfi , Fatemeh Afghah This is my paper
classification cs.CRcs.AIcs.NI
keywords InternetofVehiclesanomalydetectionautoencodertripletmarginlossunseenattackzero-daynetworkintrusiontransferlearning
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 tries to establish that a purely unsupervised autoencoder, trained only on benign network flows, can detect attacks it has never seen, including types absent from training, in Internet-of-Vehicle-style networks. The authors claim that adding a triplet margin loss to the standard reconstruction error sharpens the latent-space boundary around benign behavior, and they report roughly 99% accuracy on benign traffic and 97–100% accuracy on attack traffic across two recent IoT intrusion datasets. If the claim holds, zero-day and newly deployed IoV environments could be monitored without any pre-collected attack samples, using only normal traffic baselines. The paper also argues the model transfers across application domains through fine-tuning, so a model trained in one network environment can be adapted to another without attack data.

What carries the argument

The central object is an LSTM-based autoencoder whose encoder maps each 25-step network-flow sequence to a latent vector and whose decoder reconstructs it, with the mean squared reconstruction error used as the anomaly score. The paper's key addition is a weighted joint loss $L = \lambda_{\mathrm{REC}} L_{\mathrm{REC}} + \lambda_{\mathrm{TML}} L_{\mathrm{TML}}$, where $L_{\mathrm{TML}} = \max(d(a_i, p_i) - d(a_i, n_i) + m, 0)$ is applied to latent representations, with $p_i$ a noise-augmented copy of the anchor and $n_i$ a different benign sequence. This forces benign latent representations into tighter, better-separated clusters so that attack sequences deviate more clearly in reconstruction error. The detection threshold is set at the 99th percentile of benign training reconstruction errors.

What would settle it

Train the same architecture on a genuine vehicular network dataset containing only benign vehicle-to-everything traffic, then test on attack families tied to mobility (e.g., position falsification, Sybil, or message injection): if anomaly accuracy falls substantially below the reported 97–100% range while benign accuracy stays high, the paper's IoV transfer claim is contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, an autoencoder built from LSTM layers and trained exclusively on benign network-flow sequences, with a loss $L = \lambda_{\mathrm{REC}} L_{\mathrm{REC}} + \lambda_{\mathrm{TML}} L_{\mathrm{TML}}$, detects unseen attacks by reconstruction-error thresholding. The triplet margin loss is applied to latent representations of an anchor, a noise-augmented positive, and a different benign sequence as the negative, so the model learns a compact and cohesive benign latent region instead of relying on reconstruction fidelity alone. The authors report 99.06% benign accuracy and 97.28% anomaly accuracy on the ACI-IoT-2023 dataset at the ideal loss weights, and 99.09% benign accuracy and 100% anomaly accuracy on WUSTL-IIoT-2021, with transfer-learning variants preserving high accuracy after fine-tuning on a new domain.

Load-bearing premise

The results on home and industrial IoT testbed traffic transfer to real Internet-of-Vehicles networks, even though no vehicular or V2X data is used in the experiments.

Editorial extensions

If this is right

  • A system deployed in a new IoV environment could begin monitoring immediately using only collected benign traffic, with every attack treated as a zero-day attack.
  • Because detection does not depend on attack signatures, the method is not specialized to any one attack type and can flag DoS, brute-force, reconnaissance, and similar network-level attacks alike.
  • Transfer-learning experiments suggest a model pretrained on one network's benign traffic can be fine-tuned to a new domain with a small amount of that domain's benign data, preserving high accuracy without attack labels.
  • Oversampling the benign training set with SMOTE raises average accuracy and reduces sensitivity to the loss-weight choices, which matters when only limited benign data is available.
  • Longer sequence lengths generally improve average anomaly accuracy, so the operator can trade detection periodicity against accuracy.

Reading between the lines

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

  • The evaluation uses a smart-home IoT dataset and an industrial IoT dataset as proxies for IoV traffic, so the paper's central claim implicitly assumes that network-layer attack patterns in those domains resemble those in moving vehicle networks; a real V2X dataset would be the direct test of that assumption.
  • The joint-loss recipe is not tied to network-specific features, so the same anchor-positive-negative construction could be applied to other benign-only time-series anomaly detection problems, such as sensor monitoring or financial fraud, where only normal observations are available.
  • Because SMOTE is applied before sequence building, the paper's oversampling results suggest that synthetic augmentation of benign flows can substitute for collecting more normal traffic, which is testable in other datasets.
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 / 5 minor

Summary. The manuscript proposes an LSTM-based autoencoder trained only on benign network flows, using a weighted combination of reconstruction loss and triplet margin loss, for unseen (zero-day) attack detection in Internet of Vehicles (IoV) networks. The model is evaluated on the ACI-IoT-2023 and WUSTL-IIoT-2021 datasets, with benchmarks against DeepSVDD, GMM, a plain autoencoder, and a VAE variant, and an additional transfer-learning study from WUSTL to ACI. The paper reports roughly 99% benign accuracy and 97–100% anomaly accuracy for the proposed method, and argues that its fully unsupervised, benign-only training makes it suitable for detecting previously unseen attacks.

Significance. If the reported performance were obtained under a protocol that genuinely keeps all attack information out of model selection and threshold determination, the contribution would be practically valuable for zero-day attack detection in resource-constrained networked systems, and the combination of reconstruction and triplet losses on benign sequences is a sensible idea. The paper includes useful baselines, ablation studies (V.B), and a direct diagnostic of latent-space cohesion (Table XII). However, the central claim of 'unseen attack detection' is currently undermined by test-set-based selection of loss weights and threshold percentiles, and the evaluation datasets are not vehicular, so the IoV-specific claim is not directly supported. The manuscript's value as published depends on fixing these protocol issues.

major comments (3)
  1. [Section V, 'Implementation Details'] The paper states that the loss weights λ_TML and λ_REC are swept in [0,1] in increments of 0.1 to 'identify the ideal combinations of these weights and present the results for these ideal values' (V.2). Because the 'ideal' values are identified using the combined benign-plus-attack test set, attack labels are used during model selection. This contradicts the central claim that the method requires no attack data. The magnitude of the resulting optimism is visible in Table VII: averaged over all λ pairs without SMOTE, ACI benign accuracy is 94.11% and anomaly accuracy is 95.73%, versus 99.06% and 97.28% in Table IV for the selected ideal pair. Please re-run the evaluation using a validation split (benign-only) for weight selection, or report the full distribution of results across the grid rather than only the selected optimum.
  2. [Section V.B.3, 'Robustness Analysis Across Percentiles' and Table XI] The reconstruction-error threshold is said to be set as a percentile of benign reconstruction errors (Section IV.E), but Section V.B.3 then compares 90%, 95%, and 99% percentiles on the test set and fixes 99% 'for all other trials' based on that comparison. This is a second test-set-based selection of an operating point. Since Table XI shows substantial variation in benign accuracy across percentiles (78.30% at 90% vs. 99.06% at 99%), the headline numbers are selected, not the performance of a fixed a priori operating point. Please choose the percentile on a benign-only validation set or report the entire operating-characteristic across percentiles without presenting a single selected point.
  3. [Section III, 'IoV Network Traffic Datasets'] The paper acknowledges that ACI-IoT-2023 (a smart-home testbed) and WUSTL-IIoT-2021 (an industrial IoT plant) are not vehicular datasets, and that IoV introduces 'unique temporal patterns, mobility constraints, and attack surfaces.' Yet the title and abstract frame the contribution as IoV-specific and the conclusion states the method was evaluated on 'two state-of-the-art datasets that are well representative of modern networking patterns in distributed networked systems.' No CAN-bus, V2X, or vehicle-telemetry data is used anywhere. This gap directly limits the external validity of the IoV claim. Please either add an evaluation on a vehicular dataset (e.g., VeReMi or a real-world V2X trace) or substantially soften the IoV framing and discuss the transferability assumptions explicitly.
minor comments (5)
  1. [Section IV.A, 'Problem Definition'] The notation is inconsistent: X is defined as an element of R^{m×n}, but the following sentence says each xt is in R^m. Please clarify the intended dimensions of the time series and the feature vectors.
  2. [Section IV.B, Equations (1)-(2)] Equation (1) uses N and p but Equation (2) introduces r and t without defining r before use; the sentence 'where r = 1/T is the guessing rate' appears after Equation (2), which makes the derivation difficult to follow. Please reorder the definitions.
  3. [Section IV.D, 'Autoencoder Architecture'] The architecture description (Figure 3) omits key reproducibility details: number of LSTM layers, hidden dimensions, optimizer, learning rate, batch size, and number of training epochs. These should be reported for reproducibility.
  4. [Throughout] There are numerous typographical errors ('explicltly', 'behaivors', 'anomolous', 'guaranteened', 'V AE' spacing issues). A careful proofreading pass is recommended.
  5. [Section V.B.1, 'Transfer Learning'] The transfer-learning ablation does not isolate the effect of the triplet loss. The pre-trained model is trained with λ_TML=1.0 on WUSTL, then fine-tuned on ACI with λ_TML=0, while the 'no pre-training' baseline is trained from scratch on ACI with λ_TML=0. The observed improvement could be due to any WUSTL pre-training, not specifically to the triplet-loss-structured latent space. A pre-trained baseline with λ_TML=0 would be needed to support the stated conclusion.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported 'unseen attack' accuracies are the best hyperparameter/threshold configuration selected on the attack-labeled test set, so the zero-day claim is not independently demonstrated.

  1. fitted input called prediction [Section V, Implementation Details; used for Tables IV–VI and the Abstract's 97–100% anomaly claim]
    "We sweep across the range of λ_TML=[0,1] and λ_REC=[0,1] in increments of 0.1 to identify the ideal combinations of these weights and present the results for these ideal values unless otherwise noted."

    The headline accuracy is the grid-optimal (λ_TML, λ_REC) pair evaluated on the combined benign-plus-attack test set; no validation split is described. The attack labels are therefore an input to model selection, while the paper claims 'every attack is effectively treated as a zero-day attack ... given that no pre-existing attack data is used for training.' The reported 99% benign / 97–100% anomaly numbers are the selected pair's in-sample test performance, not the performance of a fixed unsupervised procedure on unseen attacks. Table VII shows the average over all weight pairs is lower (95.73% anomaly, 94.11% benign without SMOTE), confirming the headline depends on the test-label-guided selection.

  2. fitted input called prediction [Section IV-E and Section V-B3 (Robustness Analysis Across Percentiles), Table XI]
    "We provide analysis results for percentile values between 90% and 100% to illustrate how percentile values may impact performance and set the percentile value to 99% in all other trials."

    The operating threshold is a percentile of benign reconstruction errors, but the choice of which percentile to use is made after comparing benign and anomaly accuracy across 90–100% on the labeled test set (Table XI, Figures 5–6). Thus the 99th-percentile operating point is selected in view of attack labels; the associated benign/anomaly accuracies are not the output of a threshold rule fixed before seeing attacks.

full rationale

This paper is an empirical ML study, not a derivation, and it contains no self-citation chain or imported uniqueness theorem. The analytical framework in Sec IV-B is not used to derive the loss, so no mathematical equivalence is smuggled in there. However, the central claim—detection of attacks never seen during training—is compromised by two selection steps that use the attack labels. The loss weights are chosen by a grid sweep 'to identify the ideal combinations' and the reported tables present those ideal values; because no validation split is described, 'ideal' means best on the attack-containing test set. Similarly, the percentile threshold is set to 99% after reporting how accuracy varies across percentile values on the labeled test set. The headline 99% benign / 97–100% anomaly accuracy is therefore the test-set-optimal operating point, not the performance of a fixed unsupervised rule on unseen attacks. This is fitted-input-as-prediction rather than a definitional equivalence, so it is partial circularity (score 6) rather than total collapse (8–10). The method's architecture and benign-only training objective retain independent content; the circularity is in the evaluation protocol that selects hyperparameters and threshold on the labels it claims not to use.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The ledger shows the paper's contribution is empirical: the only quantities tuned are loss weights, threshold percentile, sequence length, and augmentation noise, all selected with test data visibility. No new theoretical entities are introduced. The strongest unstated dependency is that two non-vehicular IoT datasets stand in for IoV traffic.

free parameters (8)
  • lambda_REC (reconstruction loss weight) = 0.8 for ACI (Table VI), 0.6 for WUSTL pretraining (Tables IX-X)
    Swept over [0,1] in 0.1 increments and ideal value selected using test-set performance (Sec V.2).
  • lambda_TML (triplet margin loss weight) = 0.9 for ACI (Table VI), 1.0 for WUSTL pretraining (Tables IX-X)
    Swept together with lambda_REC; chosen on test performance.
  • Reconstruction error threshold percentile = 99th percentile
    Values from 90% to 100% were examined (Table XI); 99% is used in all other trials (Sec V.E).
  • Sequence length = 25 fixed, with trials at 10, 50, and 100
    Chosen by hand; results vary noticeably with length (Table VIII).
  • Noise augmentation for positive triplets = 0.01
    Fixed noise level for positive sample generation (Sec V.1).
  • Triplet margin m = not reported
    The margin in Eq. 11 is never given a value in the paper.
  • SMOTE usage = used for ACI in some trials
    Oversampling included or excluded based on observed performance improvements (Tables V-VII).
  • Unspecified architecture and training hyperparameters
    LSTM layer counts, hidden sizes, epochs, batch size, optimizer, and learning rate are not stated.
assumptions (4)
  • domain assumption Benign network flows are available for training and are the only class needed
    Assumption 2 in Sec IV-C; the entire method trains only on benign data.
  • domain assumption Attacks manifest as statistically distinguishable deviations in reconstruction error of fixed-length sequences
    Assumption 1 in Sec IV-C; no attack data is used in training.
  • domain assumption ACI-IoT-2023 and WUSTL-IIoT-2021 are representative of modern IoV traffic
    Sec III argues these proxy IoT datasets are better suited than NSL-KDD or VeReMi, but no actual IoV data is tested.
  • standard math Autoencoder reconstruction error is a valid anomaly score
    Standard assumption motivating reconstruction-based anomaly detection (Sec II-A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Joint Reconstruction-Triplet Loss Autoencoder Approach Towards Unseen Attack Detection in IoV Networks." pith.science (2026). https://pith.science/paper/YSKEPM3Y

@misc{pith2026250521703,
  author       = {Pith},
  title        = {Pith review of: A Joint Reconstruction-Triplet Loss Autoencoder Approach Towards Unseen Attack Detection in IoV Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YSKEPM3Y}},
  note         = {Machine review of arXiv:2505.21703}
}
read the original abstract

Internet of Vehicles (IoV) systems, while offering significant advancements in transportation efficiency and safety, introduce substantial security vulnerabilities due to their highly interconnected nature. These dynamic systems produce massive amounts of data between vehicles, infrastructure, and cloud services and present a highly distributed framework with a wide attack surface. In considering network-centered attacks on IoV systems, attacks such as Denial-of-Service (DoS) can prohibit the communication of essential physical traffic safety information between system elements, illustrating that the security concerns for these systems go beyond the traditional confidentiality, integrity, and availability concerns of enterprise systems. Given the complexity and volume of data generated by IoV systems, traditional security mechanisms are often inadequate for accurately detecting sophisticated and evolving cyberattacks. Here, we present an unsupervised autoencoder method trained entirely on benign network data for the purpose of unseen attack detection in IoV networks. We leverage a weighted combination of reconstruction and triplet margin loss to guide the autoencoder training and develop a diverse representation of the benign training set. We conduct extensive experiments on recent network intrusion datasets from two different application domains, industrial IoT and home IoT, that represent the modern IoV task. We show that our method performs robustly for all unseen attack types, with roughly 99% accuracy on benign data and between 97% and 100% performance on anomaly data. We extend these results to show that our model is adaptable through the use of transfer learning, achieving similarly high results while leveraging domain features from one domain to another.

Figures

Figures reproduced from arXiv: 2505.21703 by the authors.

Figure 1
Figure 1. t-SNE visualization of the ACI-IoT-2023 dataset [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of the WUSTL-2021 dataset [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Proposed joint triplet-reconstruction loss autoencoder architecture utilizing LSTM layers. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sequencing for IoT network flows 2) Implementation Details: For both datasets, we utilize 80% of the benign set for training purposes and 20% of the benign set for testing benign data. We sweep across the range of λTML = [0, 1] and λREC = [0, 1] in increments of 0.1 to…
Figure 5
Figure 5. Figure 5: ACI precision-recall curve across percentile values for joint [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Benign representations with and without contrastive loss. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [1]

    Indus- trial internet of things: Challenges, opportunities, and directions,

    E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, “Indus- trial internet of things: Challenges, opportunities, and directions,” IEEE Transactions on Industrial Informatics , vol. 14, no. 11, pp. 4724–4734, 2018

    work page 2018

  2. [2]

    Iot practices in military applications,

    V . Gotarane and S. Raskar, “Iot practices in military applications,” in 2019 3rd International Conference on Trends in Electronics and Informatics (ICOEI), 2019, pp. 891–894

    work page 2019

  3. [3]

    Security issues in internet of vehicles (iov): A comprehensive survey,

    H. Taslimasa, S. Dadkhah, E. C. P. Neto, P. Xiong, S. Ray, and A. A. Ghorbani, “Security issues in internet of vehicles (iov): A comprehensive survey,”Internet of Things, vol. 22, p. 100809, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2542660523001324

    work page 2023

  4. [4]

    An in- depth analysis of the mirai botnet,

    J. Margolis, T. T. Oh, S. Jadhav, Y . H. Kim, and J. N. Kim, “An in- depth analysis of the mirai botnet,” in 2017 International Conference on Software Security and Assurance (ICSSA) , 2017, pp. 6–12

    work page 2017

  5. [5]

    The mirai botnet and the iot zombie armies,

    G. Kambourakis, C. Kolias, and A. Stavrou, “The mirai botnet and the iot zombie armies,” in MILCOM 2017 - 2017 IEEE Military Communications Conference (MILCOM) , 2017, pp. 267–272

    work page 2017

  6. [6]

    Mirai ddos attack against kreb- sonsecurity cost device owners $300,000,

    C. Osborne, “Mirai ddos attack against kreb- sonsecurity cost device owners $300,000,” May

    work page

  7. [7]

    Predicting machine failures from multivariate time series: An industrial case study,

    N. O. Pinciroli Vago, F. Forbicini, and P. Fraternali, “Predicting machine failures from multivariate time series: An industrial case study,” Machines, vol. 12, no. 6, 2024. [Online]. Available: https: //www.mdpi.com/2075-1702/12/6/357

    work page 2024

  8. [8]

    Deep learning for time series classification: a review,

    H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: a review,” Data Mining and Knowledge Discovery , vol. 33, no. 4, pp. 917– 963, 2019. [Online]. Available: https://link.springer.com/article/10.1007/ s10618-019-00619-1

    work page 2019

Show all 35 references
  1. [9]

    Machine learning-based network vulnerability analysis of industrial internet of things,

    M. Zolanvari, M. A. Teixeira, L. Gupta, K. M. Khan, and R. Jain, “Machine learning-based network vulnerability analysis of industrial internet of things,” IEEE Internet of Things Journal , vol. 6, no. 4, pp. 6822–6834, 2019

    2019

  2. [10]

    Online and scalable unsupervised network anomaly detection method,

    J. Dromard, G. Roudiere, and P. Owezarski, “Online and scalable unsupervised network anomaly detection method,” IEEE Transactions on Network and Service Management , vol. 14, no. 1, pp. 34–47, 2016

    2016

  3. [11]

    Detection of eavesdrop- ping attack in uav-aided wireless systems: Unsupervised learning with one-class svm and k-means clustering,

    T. M. Hoang, N. M. Nguyen, and T. Q. Duong, “Detection of eavesdrop- ping attack in uav-aided wireless systems: Unsupervised learning with one-class svm and k-means clustering,” IEEE Wireless Communications Letters, vol. 9, no. 2, pp. 139–142, 2019

    2019

  4. [12]

    Deep learning for anomaly detection: Challenges, methods, and opportunities,

    G. Pang, L. Cao, and C. Aggarwal, “Deep learning for anomaly detection: Challenges, methods, and opportunities,” in Proceedings of the 14th ACM international conference on web search and data mining , 2021, pp. 1127–1130

    2021

  5. [13]

    Anomaly detection for iot time- series data: A survey,

    A. A. Cook, G. Mısırlı, and Z. Fan, “Anomaly detection for iot time- series data: A survey,” IEEE Internet of Things Journal , vol. 7, no. 7, pp. 6481–6494, 2019

    2019

  6. [14]

    Lstm learning with bayesian and gaussian processing for anomaly detection in industrial iot,

    D. Wu, Z. Jiang, X. Xie, X. Wei, W. Yu, and R. Li, “Lstm learning with bayesian and gaussian processing for anomaly detection in industrial iot,” IEEE Transactions on Industrial Informatics , vol. 16, no. 8, pp. 5244–5253, 2019

    2019

  7. [15]

    Online anomaly detection with concept drift adaptation using recurrent neural networks,

    S. Saurav, P. Malhotra, V . TV , N. Gugulothu, L. Vig, P. Agarwal, and G. Shroff, “Online anomaly detection with concept drift adaptation using recurrent neural networks,” in Proceedings of the acm india joint international conference on data science and management of data, 20...

    2018

  8. [16]

    Unsupervised anomaly detection in time series using lstm-based autoencoders,

    O. I. Provotar, Y . M. Linder, and M. M. Veres, “Unsupervised anomaly detection in time series using lstm-based autoencoders,” in 2019 IEEE International Conference on Advanced Trends in Information Theory (ATIT), 2019, pp. 513–517

    2019

  9. [17]

    Anomaly detection methods based on gan: a survey,

    H. Li and Y . Li, “Anomaly detection methods based on gan: a survey,” Applied Intelligence, vol. 53, no. 7, pp. 8209–8231, 2023

    2023

  10. [18]

    Dynamic thresholding for video anomaly detection,

    D. Jia, X. Zhang, J. T. Zhou, P. Lai, and Y . Wei, “Dynamic thresholding for video anomaly detection,” IET Image Processing , vol. 16, no. 11, pp. 2973–2982, 2022

    2022

  11. [19]

    An adversarial contrastive autoencoder for robust multivariate time series anomaly detection,

    J. Yu, X. Gao, F. Zhai, B. Li, B. Xue, S. Fu, L. Chen, and Z. Meng, “An adversarial contrastive autoencoder for robust multivariate time series anomaly detection,” Expert Systems with Applications , vol. 245, 2024. [Online]. Available: https://www.sciencedirect.com/science/ ar...

    2024

  12. [20]

    Contrastive autoencoder for anomaly detection in multivariate time series,

    H. Zhou, K. Yu, X. Zhang, G. Wu, and A. Yazidi, “Contrastive autoencoder for anomaly detection in multivariate time series,” Information Sciences, vol. 610, pp. 266–280, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0020025522008775

    2022

  13. [21]

    Deep convolutional autoencoder for assessment of anomalies in multi-stream sensor data,

    X. Wang, Y . Liu, and H. Zhang, “Deep convolutional autoencoder for assessment of anomalies in multi-stream sensor data,” arXiv preprint, vol. arXiv:2202.07592, 2022. [Online]. Available: https: //arxiv.org/abs/2202.07592

    2022 arXiv

  14. [22]

    Structural attention-based recurrent variational autoencoder for highway vehicle anomaly detection,

    J. Kim, S. Park, and M. Lee, “Structural attention-based recurrent variational autoencoder for highway vehicle anomaly detection,” arXiv preprint , vol. arXiv:2301.03634, 2023. [Online]. Available: https://arxiv.org/abs/2301.03634

    2023 arXiv

  15. [23]

    Location anomalies detection for connected and autonomous vehicles,

    T. Nguyen, L. Chen, and K. Wang, “Location anomalies detection for connected and autonomous vehicles,” arXiv preprint , vol. arXiv:1907.00811, 2019. [Online]. Available: https://arxiv.org/abs/1907. 00811

    1907 arXiv

  16. [24]

    Xai-ads: An explainable artificial intelligence framework for enhancing anomaly detection in autonomous driving systems,

    N. Sazid, L. Lingxi, and A. Mustafa, “Xai-ads: An explainable artificial intelligence framework for enhancing anomaly detection in autonomous driving systems,” IEEE Access, vol. 12, pp. 48 583 – 48 607, 2024

    2024

  17. [25]

    Vanet network traffic anomaly detection using gru- based deep learning model,

    A. Ghayth, A. Yassine, S. Mohammad, A. Anurag Vijay, Y . Ghazaala, A. Esraa, Saleh, A.-K. Hamza, Mohammed Ridha, D. Debabrata, and M. Renato, Racelis, “Vanet network traffic anomaly detection using gru- based deep learning model,” IEEE TRANSACTIONS ON CONSUMER ELECTRONICS,, vo...

    2024

  18. [26]

    Securing vanets: Multi-objective intrusion detection with variational autoencoders,

    N. Nissar, N. Naja, and A. Jamali, “Securing vanets: Multi-objective intrusion detection with variational autoencoders,” IEEE Transactions on Consumer Electronics , vol. 70, no. 1, pp. 3867–3874, 2024

    2024

  19. [27]

    An introduction to autoencoders,

    U. Michelucci, “An introduction to autoencoders,” arXiv preprint arXiv:2201.03898, 2022

    2022 arXiv

  20. [28]

    Triplet loss with multistage outlier suppression and class-pair margins for facial expres- sion recognition,

    W. Xie, H. Wu, Y . Tian, M. Bai, and L. Shen, “Triplet loss with multistage outlier suppression and class-pair margins for facial expres- sion recognition,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 2, pp. 690–703, 2022

    2022

  21. [29]

    Two-stage method based on triplet margin loss for pig face recognition,

    Z. Wang and T. Liu, “Two-stage method based on triplet margin loss for pig face recognition,” Computers and Electronics in Agriculture, vol. 194, p. 106737, 2022. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0168169922000540

    2022

  22. [30]

    Facenet: A unified embed- ding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embed- ding for face recognition and clustering,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2015

    2015

  23. [31]

    Smote: synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: synthetic minority over-sampling technique,” J. Artif. Int. Res. , vol. 16, no. 1, p. 321–357, jun 2002

    2002

  24. [32]

    Pyod: A python toolbox for scalable outlier detection,

    Y . Zhao, Z. Nasrullah, and Z. Li, “Pyod: A python toolbox for scalable outlier detection,” Journal of Machine Learning Research , vol. 20, no. 96, pp. 1–7, 2019. [Online]. Available: http://jmlr.org/papers/v20/ 19-011.html

    2019

  25. [33]

    Deep one-class classification,

    L. Ruff, R. A. Vandermeulen, N. Goernitz, L. Deecke, S. A. Siddiqui, A. Binder, E. M ¨uller, and M. Kloft, “Deep one-class classification,” in Proceedings of the 35th International Conference on Machine Learning (ICML) . PMLR, 2018, pp. 4393–4402. [Online]. Available: https://...

    2018

  26. [34]

    C. C. Aggarwal, Outlier Analysis , 2nd ed. Springer, 2017. [Online]. Available: https://charuaggarwal.net/outlierbook.pdf

    2017

  27. [2018]

    Available: https://www.zdnet.com/article/ mirai-botnet-attack-against-krebsonsecurity-cost-device-owners-300000/

    [Online]. Available: https://www.zdnet.com/article/ mirai-botnet-attack-against-krebsonsecurity-cost-device-owners-300000/

Pith tools

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