Pith. sign in

REVIEW 3 major objections 5 minor 56 references

HLSF fuses CP-APR structural anomaly scores with normalizing-flow density scores on latent factors and shows that the weighted combination ranks compromised-credential authentications better than either method alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 15:16 UTC pith:GDFYILEA

load-bearing objection New fusion of CP-APR and normalizing flows, clearly written, but the evaluation tunes on attack-period labels, so the unsupervised claim needs revision. the 3 major comments →

arxiv 2607.18479 v1 pith:GDFYILEA submitted 2026-07-20 cs.LG

Hybrid Latent-Structural Fusion (HLSF) for Cyber Anomaly Detection

Dorianis M. Perez , Maksim E. Eren , Bryan E. Kaiser This is my paper
classification cs.LG
keywords cyber anomaly detectionnon-negative tensor factorizationCP-APRnormalizing flowsRealNVPunsupervised learningcompromised credentials
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that two unsupervised views of authentication events carry complementary anomaly signals: CP-APR tensor factorization captures relational structure, while a normalizing flow trained on the factorization's latent factor rows captures how typical a latent combination is. HLSF normalizes the two scores and adds them with equal weight, producing a single ranking of suspicious events. On six tensor versions of a real enterprise red-team dataset with compromised credentials, the fused ranking is the most consistent, with the largest gains on the higher-order tensors that include source, destination, status, hour, and day. A defender would care because the method produces a better early alert list in an extremely imbalanced setting, and the base models train without labels. The caveat is that the reported tuning protocol uses a labeled temporal slice of the attack period, so the claim as tested is semi-supervised rather than fully unsupervised.

Core claim

The paper's central claim is that structural and latent-density anomaly signals are complementary, so fusing them ranks compromised-credential authentication events better than either signal alone. HLSF maps each observed tensor entry to a dense vector by concatenating log-transformed CP-APR factor rows for the entry's entities, trains a RealNVP normalizing flow on those vectors using benign-period data, and forms the final score as the weighted sum of the CP-APR Poisson-survival score and the flow's negative log-likelihood. On the four higher-order tensor configurations—user-source-destination with status, hour, and day variants—the fused score has the highest ROC-AUC and PR-AUC; on the min

What carries the argument

The load-bearing object is the latent event representation built by concatenating the log-transformed rows of the fitted CP-APR factor matrices for the event's entities, producing a dense vector of dimensionality D times the tensor rank. CP-APR supplies the structural score as the negative log of the Poisson survival function; RealNVP supplies the density score as the negative log-likelihood of that latent vector. After z-score normalization, HLSF scores each event as an equally weighted sum of the two scores. The mechanism works because the flow sees structured, continuous latent factors rather than raw sparse entries, which avoids the known tendency of likelihood-based models to over-credi

Load-bearing premise

The reported gains depend on tuning the flow and the fusion weight using a labeled temporal slice of the attack period; if those labels are unrepresentative or unavailable in deployment, the hybrid's edge over its constituents is not established.

What would settle it

Run the same six-tensor comparison while choosing the RealNVP hyperparameters and fusion weight using only benign training data, with no attack labels anywhere in model selection. If HLSF no longer tops both CP-APR and RealNVP on the four high-order tensors, the reported advantage depends on the labeled tuning slice rather than on the fusion mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • On the four higher-order tensors, HLSF achieves the best ROC-AUC and PR-AUC, with the strongest overall result on the richest tensor that includes hour and day context.
  • The hybrid surface more anomalous events in the first small fraction of records reviewed, which is the regime analysts actually work in.
  • The two signals are complementary: events that look structurally plausible under the tensor model can still land in low-density regions of latent space, and the fusion catches both failure modes.
  • The fusion does not help on the two-mode user-destination tensor, so the method's advantage is conditional on the tensor having enough relational and temporal context.

Where Pith is reading between the lines

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

  • Because hyperparameters and the fusion weight are chosen on a labeled temporal slice of the test period, the practical claim actually supported is semi-supervised; a fully label-free deployment would need a rule for setting the fusion weight without attack labels, and that version remains untested.
  • The fusion recipe is not tied to cyber authentication: any sparse event tensor with fitted latent factors could be paired with a density estimator the same way, so it should transfer to fraud, system-log monitoring, or other rare-event settings if the tensor is high-order enough.
  • A sharper test of complementarity would compare HLSF against a normalizing flow trained directly on raw tensor entries or on a different latent embedding; the current design cannot separate the benefit of fusion from the benefit of feeding the flow the specific log-concatenated CP-APR factors.
  • The fixed equal weight chosen across all datasets hints that the two signals are roughly equally informative for these tensors; an adaptive weight that shifts more weight to the structural score as tensor order grows is a natural untested extension.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Hybrid Latent-Structural Fusion (HLSF), which combines CP-APR tensor-factorization anomaly scores (Poisson survival p-values) with RealNVP normalizing-flow negative-log-likelihood scores computed on CP-APR latent factor representations. The method is evaluated on six tensor formulations of the LANL authentication dataset (US, UD, USD, USDs, USDHs, USDHDs) and compared against standalone CP-APR and standalone RealNVP using ROC-AUC, PR-AUC, recall-versus-review, and lift curves. The central claim is that HLSF consistently improves anomaly ranking over either constituent method, particularly for higher-order tensor representations. The main experiments report mean±1σ over 10 ensemble runs. The paper acknowledges a 'semi-supervised model-selection procedure' in which RealNVP hyperparameters and the fusion weight α are tuned on a labeled temporal subset of the test period, with final evaluation on the remaining held-out test subset.

Significance. If the empirical claim is supported, HLSF is a simple and plausible contribution: it combines two complementary anomaly signals (low-rank structural deviations and latent-space density deviations) and demonstrates gains on a public real-world cybersecurity dataset. The paper builds on a strong baseline (CP-APR) and uses a realistic, severely imbalanced evaluation setting. The inclusion of ensemble variability and operational metrics (recall/lift) is a strength. However, the current evaluation protocol prevents the central claim from being accepted: the use of test-period labels for model selection means the 'held-out' test subset is not fully held out, and the reported PR-AUC gains are not supported by significance testing. The manuscript also lacks released code for the HLSF fusion itself, although the pyCP_APR library is public.

major comments (3)
  1. [Section 4.2] The evaluation protocol is not fully unsupervised. RealNVP hyperparameters are selected by maximizing AP on a labeled temporal subset of the test period, and Section 3.4 states that the fusion weight α was selected using validation data. Because the validation and held-out anomalies come from the same red-team exercise, tuning on the validation subset can adapt to the specific attack pattern, inflating HLSF's reported gains relative to CP-APR, whose rank selection uses only the benign validation period (days 49–56). This is a load-bearing issue for the central claim that HLSF outperforms standalone methods on held-out data. Please re-run the comparison under a protocol that does not use test-period labels for any model selection (e.g., select flow hyperparameters and α using a benign validation period or a label-free criterion), or explicitly reframe the method and comparisons as semi-su
  2. [Table 4] The PR-AUC improvements are statistically fragile. For example, on USD, HLSF PR-AUC is 0.0573±0.0273 versus RealNVP 0.0357±0.0220 and CP-APR 0.0092±0.0011; the HLSF-vs-RealNVP difference is less than 1σ, and the HLSF-vs-CP-APR difference is about 1.8σ. Similar overlap appears for USDs (0.0532±0.0296 vs 0.0399±0.0281) and USDHs (0.0458±0.0312 vs 0.0235±0.0159). No significance tests or paired comparisons are reported. Given that the conclusion emphasizes 'more robust ranking,' please provide per-run paired significance tests (e.g., Wilcoxon signed-rank, bootstrap confidence intervals on the difference) or report the full per-run metric distributions.
  3. [Sections 3.4 and 4.2] The fusion weight α=0.5 is reported as the value that 'consistently achieved strong performance across datasets' on validation data, but no sensitivity analysis or α-selection outcome is shown. If α is tuned on the labeled test-period validation subset, the HLSF results reflect this tuning, and the advantage over standalone methods may not generalize to a deployment where attack-period labels are unavailable. Please provide α sensitivity curves or a table showing performance across candidate α values, and clarify whether the chosen α would remain fixed under the proposed unsupervised selection protocol.
minor comments (5)
  1. [Abstract / Introduction] The abstract and introduction describe the methods as 'unsupervised,' but Section 4.2 explicitly uses a 'semi-supervised model-selection procedure' with test-period labels. Please qualify these statements to avoid inconsistency.
  2. [Eq. (28)] The log-offset ε is introduced but its numerical value is never specified. State the value used in the experiments.
  3. [Section 4.2] The temporal split within the test period is described only verbally; the exact day ranges for the model-selection validation subset and the fully held-out test subset are not given. Please provide these split boundaries so the reader can assess the overlap with the red-team activity.
  4. [Figures 2–7] The figures show ensemble results but no error bands or shading. Adding confidence intervals or at least per-run curves would make the variability visible, especially where Table 4 reports large standard deviations.
  5. [Results / Table 4] The text states HLSF obtains the best ROC-AUC on 'all tensors except UD,' but on US the PR-AUC of RealNVP (0.0424±0.0361) is higher than HLSF (0.0409±0.0350). The 'strongest overall performance' claim should be qualified to ROC-AUC and the four higher-order tensors.

Circularity Check

0 steps flagged

No significant circularity: the HLSF fusion claim is empirical and not forced by definition or by self-citation.

full rationale

The paper's central claim is empirical: that the convex combination in Eq. (31) of CP-APR structural scores and RealNVP negative-log-likelihood scores on CP-APR latent factors improves ROC-AUC/PR-AUC on held-out LANL tensor entries. Nothing in the definition of the HLSF score entails the measured ranking improvement; the claim is supported by Table 4 and the accompanying evaluation. The fitted quantities—RealNVP hyperparameters in Table 3 and the fusion weight alpha in Section 3.4—are selected on a labeled validation slice of the test period and then evaluated on a separate fully held-out test subset, as disclosed in Section 4.2. This is an evaluation-protocol limitation rather than a circular reduction: alpha=0.5 does not encode the held-out labels or the final test ranking. The paper's reliance on Eren et al. [18,19,20] supplies the CP-APR implementation, rank-selection procedure, and smoothing constant; these are baseline ingredients, and the pyCP_APR library is open-source, code-reproduced software. No equation reduces to another by construction, and no fitted parameter is renamed as a prediction. The acknowledged limitation that HLSF depends on the quality of the tensor factorization is a substantive scope restriction, not a circularity. I therefore find no significant circularity and assign a score of 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method introduces no new entities, forces, or conserved quantities. The free parameters are the CP-APR rank, the fusion weight α, the inherited smoothing weight a, the RealNVP hyperparameters, and the unspecified log-offset ϵ. The key domain assumptions are the relevance of CP-APR latent factors to test entities and the validity of flow likelihood as an anomaly score in this latent space.

free parameters (5)
  • CP-APR rank R = 20, 20, 4, 4, 5, 45 (per tensor, Table 2)
    Selected by maximizing validation log-likelihood over ranks 1–100 (Section 4.2). Directly shapes the latent representations the flow sees.
  • fusion weight α = 0.5
    Selected on validation data from candidate values (Section 3.4). Equal weighting was used in all experiments; this is a fitted choice, not derived.
  • smoothing weight a = 0.1
    Inherited from Eren et al. [20] for Poisson-rate smoothing in Eq. (15). Not fitted here but a free parameter of the underlying method.
  • RealNVP hyperparameters (per tensor) = Table 3: learning rate, hidden features, batch size, layers
    Selected via Hyperband using the validation subset of the labeled test period (Section 4.2). These choices affect the density estimates and anomaly scores.
  • log-offset ϵ = not specified
    Appears in Eq. (28) to avoid log(0). The value is never given; with nonnegative CP-APR factors that can be exactly zero, this choice affects the latent features materially.
axioms (5)
  • domain assumption Each tensor entry is an independent Poisson draw (Eq. 8)
    CP-APR models counts as Poisson; the paper actually binarizes the counts (Section 4.1), which the Poisson model approximates only loosely.
  • domain assumption CP-APR latent factors learned on the benign training period provide valid representations for test-period entities and events (Eq. 28)
    The paper does not state how out-of-vocabulary entities or exactly-zero factor rows are handled; the entire flow pipeline depends on a factor row existing for every test entity.
  • domain assumption A normalizing flow trained on latent representations of benign events assigns low likelihood to anomalous events (Sections 3.3–3.4)
    The core anomaly-detection premise; known to fail for raw data in some OOD settings [26, 32, 41], and here transferred to the latent space without proof.
  • domain assumption Validation labels from the test period are representative enough for hyperparameter selection (Section 4.2)
    RealNVP hyperparameters and α are chosen on a temporal subset of the labeled test period; if those anomalies differ from the held-out ones, the selected configuration may be biased.
  • standard math Standard change-of-variables formula for flows and Poisson survival for p-values (Eqs. 16, 21)
    Background results used without proof; they are correct and standard.

pith-pipeline@v1.3.0-alltime-deepseek · 15010 in / 15073 out tokens · 135775 ms · 2026-08-01T15:16:55.460426+00:00 · methodology

0 comments
read the original abstract

Malicious anomalous activity detection is a fundamental challenge for cyber security systems. Both tensor decomposition under statistical framework with CANDECOMP-PARAFAC alternating Poisson regression (CP-APR) and normalizing flows have proven to be powerful unsupervised machine learning methods that model multi-dimensional data and capture complex and multi-faceted details of behavior profiles in cyber security applications. In this study, we propose Hybrid Latent-Structural Fusion (HLSF), a weighted anomaly fusion framework integrating CP-APR structural anomaly scores with latent-space density scores derived from normalizing flows. In our experiments, we show that the HLSF framework improves anomaly detection performance on a dataset of real-world compromised user credentials collected from the large enterprise network of Los Alamos National Laboratory (LANL) during a red-teaming exercise, compared with using CP-APR or normalizing flows alone.

Figures

Figures reproduced from arXiv: 2607.18479 by Bryan E. Kaiser, Dorianis M. Perez, Maksim E. Eren.

Figure 1
Figure 1. Figure 1: Binary tensor with User-Source-Destination dimensions showing background [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method comparisons using ensemble runs of the US dataset for (a) ROC, (b) [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Method comparisons using ensemble runs of the UD dataset for (a) ROC, (b) [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Method comparisons using ensemble runs of the USD dataset for (a) ROC, (b) [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Method comparisons using ensemble runs of the USDs dataset for (a) ROC, (b) [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Method comparisons using ensemble runs of the USDHs dataset for (a) ROC, (b) [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Method comparisons using ensemble runs of the USDHDs dataset for (a) ROC, [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 8 canonical work pages

  1. [1]

    A step-by- step training method for multi generator gans with application to anomaly detection and cybersecurity

    Adiban, M., Siniscalchi, S.M., Salvi, G., 2023. A step-by- step training method for multi generator gans with application to anomaly detection and cybersecurity. Neurocomputing 537, 296–

    2023

  2. [2]

    Ahmadian, A., Lindsten, F., 2021. Likelihood-free out-of-distribution detection with invertible generative models, in: Proceedings of the Thir- tieth International Joint Conference on Artificial Intelligence (IJCAI- 21), pp. 2119–2125. doi:10.24963/ijcai.2021/292

    work page doi:10.24963/ijcai.2021/292 2021

  3. [3]

    Ammar, A., Al Bassam, A., 2025. Normalizing-flow-based anomaly scor- ing for intelligent network intrusion detection, in: 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Network- ing (QPAIN), IEEE, Rangpur, Bangladesh

    2025

  4. [4]

    Computer security threat monitoring and surveil- lance

    Anderson, J.P., 1980. Computer security threat monitoring and surveil- lance. Technical Report, James P. Anderson Company

    1980

  5. [5]

    A review on application of gans in cybersecurity domain

    Arora, A., Shantanu, 2022. A review on application of gans in cybersecurity domain. IETE Technical Re- view 39, 433–441. URL:https://doi.org/10.1080/ 02564602.2020.1854058, doi:10.1080/02564602.2020.1854058, arXiv:https://doi.org/10.1080/02564602.2020.1854058

    arXiv 2022

  6. [6]

    A benchmark of medical out-of-distribution detection

    Cao, T., Huang, C.W., Hui, D.Y.T., Cohen, J.P., 2020. A benchmark of medical out-of-distribution detection. arXiv preprint arXiv:2007.04250 URL:https://arxiv.org/abs/2007.04250

    Pith/arXiv arXiv 2020

  7. [7]

    Deep learning for anomaly detection: A survey

    Chalapathy, R., Chawla, S., 2019. Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407 URL:https://arxiv.org/ abs/1901.03407

    Pith/arXiv arXiv 2019

  8. [8]

    Deep generative model with hierarchical latent factors for time series anomaly detection, in: Camps-Valls, G., Ruiz, F.J.R., Valera, I

    Challu, C.I., Jiang, P., Nian Wu, Y., Callot, L., 2022. Deep generative model with hierarchical latent factors for time series anomaly detection, in: Camps-Valls, G., Ruiz, F.J.R., Valera, I. (Eds.), Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, PMLR.pp.1643–1654. URL:https://proceedings.mlr.press/v151/ challu22a.html

    2022

  9. [9]

    Anomaly detection: A survey

    Chandola, V., Banerjee, A., Kumar, V., 2009. Anomaly detection: A survey. ACM Computing Surveys 41. doi:10.1145/1541880.1541882

    arXiv 2009

  10. [10]

    Semi-supervised anomaly detection via reinforcement learning-enabled method with causal inference

    Chen, X., Xiao, R., Zeng, Z., Zhang, S., Du, X., 2025. Semi-supervised anomaly detection via reinforcement learning-enabled method with causal inference. Information Sciences 719, 122463. URL:https://www. 27 sciencedirect.com/science/article/pii/S002002552500595X, doi:https://doi.org/10.1016/j.ins.2025.122463

    arXiv 2025

  11. [11]

    On tensors, sparsity, and nonnega- tive factorizations

    Chi, E.C., Kolda, T.G., 2012. On tensors, sparsity, and nonnega- tive factorizations. SIAM Journal on Matrix Analysis and Applica- tions 33, 1272–1299. URL:https://doi.org/10.1137/110859063, doi:10.1137/110859063

    work page doi:10.1137/110859063 2012

  12. [12]

    Classification of red team authentication events in an enterprise network, in: Machine Learning and Knowledge Dis- covery for Cybersecurity

    Conroy, J.M., 2018. Classification of red team authentication events in an enterprise network, in: Machine Learning and Knowledge Dis- covery for Cybersecurity. World Scientific, pp. 179–194. doi:10.1142/ 9781786345646_009

    2018

  13. [13]

    An intrusion-detection model

    Denning, D.E., 1987. An intrusion-detection model. IEEE Transac- tions on Software Engineering SE-13, 222–232. doi:10.1109/TSE.1987. 232894

    work page doi:10.1109/tse.1987 1987

  14. [14]

    Nice: Non-linear independent components estimation

    Dinh, L., Krueger, D., Bengio, Y., 2014. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516 URL:https: //arxiv.org/abs/1410.8516

    Pith/arXiv arXiv 2014

  15. [15]

    Density estimation using real nvp

    Dinh, L., Sohl-Dickstein, J., Bengio, S., 2016. Density estimation using real nvp. arXiv preprint arXiv:1605.08803 URL:https://arxiv.org/ abs/1605.08803

    Pith/arXiv arXiv 2016

  16. [16]

    A compre- hensive survey of generative adversarial networks (gans) in cybersecu- rity intrusion detection

    Dunmore, A., Jang-Jaccard, J., Sabrina, F., Kwak, J., 2023. A compre- hensive survey of generative adversarial networks (gans) in cybersecu- rity intrusion detection. IEEE Access 11, 76071–76094. doi:10.1109/ ACCESS.2023.3296707

    arXiv 2023

  17. [17]

    Variationalautoencodersusingconvolutional neural network for highly advanced cyber threats, in: 2024 IEEE Inte- grated STEM Education Conference (ISEC), pp

    DuraiPandian, A.P., 2024. Variationalautoencodersusingconvolutional neural network for highly advanced cyber threats, in: 2024 IEEE Inte- grated STEM Education Conference (ISEC), pp. 01–06. doi:10.1109/ ISEC61299.2024.10664944

    arXiv 2024

  18. [18]

    Multi-dimensional anomalous entity detection via poisson tensor factorization, in: 2020 IEEE International Conference on Intelligence and Security Informatics (ISI), pp

    Eren, M.E., Moore, J.S., Alexandrov, B.S., 2020. Multi-dimensional anomalous entity detection via poisson tensor factorization, in: 2020 IEEE International Conference on Intelligence and Security Informatics (ISI), pp. 1–6. doi:10.1109/ISI49825.2020.9280524. 28

    arXiv 2020

  19. [19]

    pycp_apr.https://github.com/lanl/pyCP_ APR

    Eren, M.E., Moore, J.S., Skau, E., Bhattarai, M., Chennupati, G., Alexandrov, B.S., 2021. pycp_apr.https://github.com/lanl/pyCP_ APR. doi:10.5281/zenodo.4840598

    work page doi:10.5281/zenodo.4840598 2021

  20. [20]

    General-purpose unsupervised cy- ber anomaly detection via non-negative tensor factorization

    Eren, M.E., Moore, J.S., Skau, E., Moore, E., Bhattarai, M., Chen- nupati, G., Alexandrov, B.S., 2023. General-purpose unsupervised cy- ber anomaly detection via non-negative tensor factorization. Digital Threats: Research and Practice 4, 1–28. doi:10.1145/3519602

    work page doi:10.1145/3519602 2023

  21. [21]

    Using collab- orative filtering to weave an information tapestry

    Goldberg, D., Nichols, D., Oki, B.M., Terry, D., 1992. Using collab- orative filtering to weave an information tapestry. Commun. ACM 35, 61–70. URL:https://doi.org/10.1145/138859.138867, doi:10. 1145/138859.138867

    arXiv 1992

  22. [22]

    Gudovskiy, D., Ishizaka, S., Kozuka, K., 2022. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional nor- malizing flows, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 98–107

    2022

  23. [23]

    A normalizing flow- based semi-supervised method for imbalanced network intrusion detec- tion

    Guo, C., Yang, S., Cheng, J., Zhang, D., 2025. A normalizing flow- based semi-supervised method for imbalanced network intrusion detec- tion. International Journal of Computers Communications & Control

    2025

  24. [24]

    Hong, Z., Yue, Y., Chen, Y., Cong, L., Lin, H., Luo, Y., et al.,

  25. [25]

    doi:10.15837/ijccc.2025.4.6890

    work page doi:10.15837/ijccc.2025.4.6890 2025

  26. [26]

    Why normal- izing flows fail to detect out-of-distribution data, in: Advances in Neural Information Processing Systems (NeurIPS), pp

    Kirichenko, P., Izmailov, P., Wilson, A.G., 2020. Why normal- izing flows fail to detect out-of-distribution data, in: Advances in Neural Information Processing Systems (NeurIPS), pp. 20578– 20589. URL:https://proceedings.neurips.cc/paper/2020/file/ ecb9fe2fbb99c31f567e9823e884dbec-Paper.pdf. 29

    2020

  27. [27]

    Normalizing flows: An introduction and review of current methods

    Kobyzev, I., Prince, S.J., Brubaker, M.A., 2020. Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 3964–3979

    2020

  28. [28]

    Msattnflow: Normalizing flow for unsuper- vised anomaly detection with multi-scale attention

    Hu, Z., Zeng, X., Li, Y., Yin, Z., Meng, E., Wei, Z., Zhu, L., Wang, Z., 2025. Msattnflow: Normalizing flow for unsuper- vised anomaly detection with multi-scale attention. Pattern Recogni- tion 161, 111220. URL:https://www.sciencedirect.com/science/ article/pii/S0031320324009713, doi:https://doi.org/10.1016/j. patcog.2024.111220

    arXiv 2025

  29. [29]

    The program with a personality: Analysis of elk cloner, the first personal computer virus

    Levy, S., Crandall, J.R., 2020. The program with a personality: Analysis of elk cloner, the first personal computer virus. CoRR abs/2007.15759. URL:https://arxiv.org/abs/2007.15759,arXiv:2007.15759

    Pith/arXiv arXiv 2020

  30. [30]

    Hyperband: A novel bandit-based approach to hyperparameter optimization

    Li, L., Jamieson, K., 2018. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of Machine Learning Research 18, 1–52. URL:http://jmlr.org/papers/v18/16-558.html

    2018

  31. [31]

    Anomaly detection in large-scale networks with latent space mod- els

    Lee, W., McCormick, T.H., Neil, J., Sodja, C., Cui, Y., 2022. Anomaly detection in large-scale networks with latent space mod- els. Technometrics 64, 241–252. URL:https://doi.org/10.1080/ 00401706.2021.1952900, doi:10.1080/00401706.2021.1952900, arXiv:https://doi.org/10.1080/00401706.2021.1952900

    arXiv 2022

  32. [32]

    Do deep generative models know what they don’t know?, in: International Conference on Learning Representations (ICLR)

    Nalisnick, E., Matsukawa, A., Teh, Y.W., Görür, D., Lakshmi- narayanan, B., 2019. Do deep generative models know what they don’t know?, in: International Conference on Learning Representations (ICLR). URL:https://openreview.net/forum?id=H1xwNhCcYm

    2019

  33. [33]

    Phishnet- vae cybersecurity approach: An integrated variational autoencoder and deep neural network approach for enhancing cybersecurity strategies by detecting phishing attacks

    Nuiaa Al Ogaili, R.R., Mahdi, M.I., Neamah, A.F., Alradha Alsaidi, S.A.A., Alsaeedi, A.H., Dashoor, Z.A., Manickam, S., 2025. Phishnet- vae cybersecurity approach: An integrated variational autoencoder and deep neural network approach for enhancing cybersecurity strategies by detecting phishing attacks. International Journal of Intelligent Engi- neering &...

    2025

  34. [34]

    Momtazpour, M., Zhang, J., Rahman, S., Sharma, R., Ramakrishnan, N., 2015. Analyzinginvariantsincyber-physicalsystemsusinglatentfac- tor regression, in: Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Association for Computing Machinery, New York, NY, USA. p. 2009–2018. URL: https://doi.org/10.1145/27832...

    arXiv 2015

  35. [35]

    Normalizing flows for probabilistic modeling and inference

    Papamakarios, G., Nalisnick, E., Rezende, D.J., Mohamed, S., Laksh- minarayanan, B., 2021. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research 22, 1–64. URL: http://jmlr.org/papers/v22/19-1028.html

    2021

  36. [36]

    Graph link prediction in computer networks using poisson matrix factorisation

    Passino, F.S., Turcotte, M.J.M., Heard, N.A., 2020. Graph link prediction in computer networks using poisson matrix factorisation. arXiv preprint arXiv:2001.09456 URL:https://arxiv.org/abs/2001. 09456

    Pith/arXiv arXiv 2020

  37. [37]

    Understanding likelihood of normalizing flow and image complexity through the lens of out-of- distribution detection

    Osada, G., Takahashi, T., Nishide, T., 2024. Understanding likelihood of normalizing flow and image complexity through the lens of out-of- distribution detection. Proceedings of the AAAI Conference on Artificial 30 Intelligence 38, 21492–21500. URL:https://ojs.aaai.org/index. php/AAAI/article/view/30146, doi:10.1609/aaai.v38i19.30146

    doi:10.1609/aaai.v38i19.30146 2024

  38. [38]

    Likelihood ratios for out-of-distribution de- tection, in: Advances in Neural Information Processing Sys- tems (NeurIPS)

    Ren, J., Liu, P.J., Fertig, E., Snoek, J., Poplin, R., DePristo, M., et al., 2019. Likelihood ratios for out-of-distribution de- tection, in: Advances in Neural Information Processing Sys- tems (NeurIPS). URL:https://papers.neurips.cc/paper/ 9611-likelihood-ratios-for-out-of-distribution-detection. pdf

    2019

  39. [39]

    Same same but differnet: Semi-supervised defect detection with normalizing flows, in: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp

    Rudolph, M., Wandt, B., Rosenhahn, B., 2021. Same same but differnet: Semi-supervised defect detection with normalizing flows, in: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1907–1916

    2021

  40. [40]

    Time of day anomaly detection, in: 2018 European Intelligence and Security Infor- matics Conference (EISIC), IEEE

    Price-Williams, M., Turcotte, M.J., Heard, N., 2018. Time of day anomaly detection, in: 2018 European Intelligence and Security Infor- matics Conference (EISIC), IEEE. pp. 1–6. doi:10.1109/EISIC.2018. 00009

    work page doi:10.1109/eisic.2018 2018

  41. [41]

    Input complexity and out-of-distribution detection with likelihood-based generative models, in: International Conference on Learning Representations (ICLR)

    Serrà, J., Álvarez, D., Gómez, V., Slizovskaia, O., Núñez, J.F., Luque, J., 2020. Input complexity and out-of-distribution detection with likelihood-based generative models, in: International Conference on Learning Representations (ICLR). URL:https://openreview.net/ forum?id=SyxIWpVYvr. 31

    2020

  42. [42]

    Generative Adversarial Networks (GAN) In- sights for Cyber Security Applications

    Shaikh, M.S., 2025. Generative Adversarial Networks (GAN) In- sights for Cyber Security Applications. Springer Nature Switzer- land, Cham. URL:https://doi.org/10.1007/978-3-031-81481-5_ 11, doi:10.1007/978-3-031-81481-5_11

    work page doi:10.1007/978-3-031-81481-5_ 2025

  43. [43]

    Variational Autoencoder (VAE) for Anomaly De- tection in Network Traffic

    Sayyad, S.A., 2024. Variational Autoencoder (VAE) for Anomaly De- tection in Network Traffic. Master’s thesis. National College of Ireland. Dublin. URL:https://norma.ncirl.ie/8624/

    2024

  44. [44]

    Using variational autoen- coders with machine learning algorithms in cyber security applications

    Taylor, T., Eleyan, A., Al-Khalidi, M., 2025. Using variational autoen- coders with machine learning algorithms in cyber security applications. Artificial Intelligence and Applications 3, 428–442. URL:https:// ojs.bonviewpress.com/index.php/AIA/article/view/4151, doi:10. 47852/bonviewAIA52024151

    2025

  45. [45]

    Poisson fac- torization for peer-based anomaly detection, in: 2016 IEEE Confer- ence on Intelligence and Security Informatics (ISI), IEEE

    Turcotte, M.J., Moore, J., Heard, N., McPhall, A., 2016. Poisson fac- torization for peer-based anomaly detection, in: 2016 IEEE Confer- ence on Intelligence and Security Informatics (ISI), IEEE. pp. 208–210. doi:10.1109/ISI.2016.7745472

    arXiv 2016

  46. [46]

    Prediction of industrial cyber attacks using normalizing flows

    Stepashkina, V.P., Hushchyn, M.I., 2024. Prediction of industrial cyber attacks using normalizing flows. Doklady Mathematics 110, S95–S102. Translated from Doklady Rossiiskoi Akademii Nauk. Matematika, Infor- matika, Protsessy Upravleniya, Vol. 520, No. 2, pp. 107–115

    2024

  47. [47]

    Maximizing anomalydetectionperformanceusinglatentvariablemodelsinindustrial systems

    Wang, K., Guo, Z., Mo, Y., Wang, Y., Yuan, X., 2024. Maximizing anomalydetectionperformanceusinglatentvariablemodelsinindustrial systems. IEEE Transactions on Automation Science and Engineering 21, 4808–4816. doi:10.1109/TASE.2023.3302794

    arXiv 2024

  48. [48]

    Application of uncertainty to out-of-distribution detection for autonomous driving perception safety

    Wang, K., Ma, Q., Shen, C., Lu, J., 2025. Application of uncertainty to out-of-distribution detection for autonomous driving perception safety. IEEE Transactions on Intelligent Transportation Systems 26, 11276– 11293

    2025

  49. [49]

    Unified host and network data set

    Turcotte, M.J.M., Kent, A.D., Hash, C.L., 2017. Unified host and network data set. ArXiv abs/1708.07518. URL:https://api. semanticscholar.org/CorpusID:1407948

    Pith/arXiv arXiv 2017

  50. [50]

    Understanding fail- ures in out-of-distribution detection with deep generative models, in: 32 Proceedings of the 38th International Conference on Machine Learning (ICML), pp

    Zhang, L., Goldstein, M., Ranganath, R., 2021. Understanding fail- ures in out-of-distribution detection with deep generative models, in: 32 Proceedings of the 38th International Conference on Machine Learning (ICML), pp. 12427–12436. URL:https://proceedings.mlr.press/ v139/zhang21m.html

    2021

  51. [51]

    Improving out-of-distribution detection in normalizing flows with synthetic outliers

    Zhao, Y., Ding, Q., Zhang, X., 2026. Improving out-of-distribution detection in normalizing flows with synthetic outliers. Neurocomput- ing 677, 133081. URL:https://www.sciencedirect.com/science/ article/pii/S0925231226004789, doi:https://doi.org/10.1016/j. neucom.2026.133081

    arXiv 2026

  52. [52]

    Generalized out-of-distribution detection: A survey

    Yang, J., Zhou, K., Li, Y., Liu, Z., 2024. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision 132, 5635–5662. doi:10.1007/s11263-024-02117-4

    doi:10.1007/s11263-024-02117-4 2024

  53. [53]

    Msflow: Multi- scale flow-based framework for unsupervised anomaly detection

    Zhou, Y., Xu, X., Song, J., Shen, F., Shen, H.T., 2025. Msflow: Multi- scale flow-based framework for unsupervised anomaly detection. IEEE Transactions on Neural Networks and Learning Systems 36, 2437–2450. doi:10.1109/TNNLS.2023.3344118. 33

    arXiv 2025

  54. [55]

    Semi- supervised anomaly detection via neural process

    Zhou, F., Wang, G., Zhang, K., Liu, S., Zhong, T., 2023. Semi- supervised anomaly detection via neural process. IEEE Transactions on Knowledge and Data Engineering 35, 10423–10435. doi:10.1109/ TKDE.2023.3266755

    arXiv 2023

  55. [308]

    URL:https://www.sciencedirect.com/science/article/pii/ S0925231223003065, doi:https://doi.org/10.1016/j.neucom.2023. 03.056. 26

    doi:10.1016/j.neucom.2023 2023

  56. [2024]

    arXiv preprint arXiv:2404.18279 URL:https://arxiv.org/abs/2404

    Out-of-distribution detection in medical image analysis: A survey. arXiv preprint arXiv:2404.18279 URL:https://arxiv.org/abs/2404. 18279

    Pith/arXiv arXiv