The reviewed record of science sign in
Pith

arxiv: 2607.06094 · v1 · pith:7RPVCTNP · submitted 2026-07-07 · cs.LG

Modeling Normal Is All You Need: Joint Latent Clustering for Anomaly Detection in Multimodal Cyber-Physical Systems

Alexander Apartsin , Yehudit Aperstein This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 16:39 UTCglm-5.2pith:7RPVCTNPrecord.jsonopen to challenge →

classification cs.LG
keywords anomaly detectioncyber-physical systemsvariational deep embeddingGaussian mixture modelmultimodal normal behaviorpoint-adjusted evaluationreconstruction residuallatent space scoring
0
0 comments X

The pith

Drop reconstruction, cluster the latent: a CPS detector wins by modeling normal

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

The paper argues that anomaly detection in cyber-physical systems fails not because the algorithms are too simple but because they model normal behavior as a single blob and score anomalies through reconstruction residuals. The authors formalize the structure of CPS normal data as MIIM (Massive, Implicit, Imbalanced Multimodality): a union of many bounded, curved, imbalanced operating regimes with thin fringes, not a single Gaussian cloud. They then build a detector around a Variational Deep Embedding (VaDE), which jointly learns a low-dimensional latent representation and a Gaussian-mixture clustering of operating modes, and score anomalies entirely within that latent space. The key mechanistic move is dropping the reconstruction term: a flexible decoder faithfully rebuilds the hardest faults (correlation breaks, between-mode pockets), so the reconstruction residual is at or below chance on exactly the faults that matter. Instead, the detector uses a high-component-count mixture density head for non-Gaussian intra-mode structure and a nearest-component likelihood that does not penalize rare-but-valid modes. Two optional heads (a whitened residual and a basin-agreement rescue) are auto-gated using only train-normal statistics, so a single configuration adapts itself across datasets with different fault signatures. Under a deliberately fair evaluation protocol—raw point-wise metrics with no point adjustment, a trivial-detector difficulty split, prevalence-matched F1, and train-normal-only calibration—the detector wins both the combined and the difficult-subset AUROC columns on all three real CPS benchmarks (HAI, WADI, SKAB), with the margin tracking dataset multimodality as the MIIM thesis predicts.

Core claim

The central discovery is that scoring anomalies in a jointly learned latent space—while explicitly dropping the reconstruction residual—converts the representation from a liability into the source of detection power on the hardest CPS faults. The reconstruction residual is at or below chance on correlation-break faults because a flexible decoder rebuilds them faithfully; removing it and scoring via a high-K mixture density plus a rare-mode-safe nearest-component likelihood yields difficult-subset AUROC of 0.831 on HAI, 0.726 on WADI, and 0.610 on SKAB, beating three re-scored deep SOTA detectors (USAD, TranAD, GDN) which collapse to 0.49–0.59 on the difficult subset. The advantage is largest

What carries the argument

The detector (VaDE-hard+resid(auto)) is built on Variational Deep Embedding (VaDE): a VAE whose latent prior is a Gaussian mixture, so representation and mode clustering are learned jointly. Scoring is entirely in the latent: a high-K (K=80) diagonal Gaussian mixture density head captures non-Gaussian intra-mode fringes, and a nearest-component negative log-likelihood avoids penalizing rare-but-valid modes. The reconstruction term is dropped. Two optional heads are auto-gated on train-normal signals: a responsibility-weighted whitened residual (gated by held-out-normal generalization ratio q95(B)/q95(A)) and a basin-agreement rescue (gated by the fraction of ambiguous train-normal windows).

If this is right

  • If the MIIM thesis is correct, any CPS anomaly detector that scores through reconstruction or treats normal as unimodal will systematically miss correlation-break and between-mode faults, regardless of how sophisticated its architecture is.
  • The difficulty-stratified evaluation protocol (trivial-detector split, raw metrics, no point adjustment) could be adopted as a community standard, exposing whether published gains are real or artifacts of inflated metrics and easy anomalies.
  • The auto-gating mechanism suggests a path toward dataset-agnostic anomaly detectors: a single architecture that configures its scoring heads from train-normal statistics alone, without per-dataset tuning.
  • The finding that explicit temporal/trajectory features added nothing on these benchmarks implies that current public CPS datasets lack history-dependent faults, motivating the collection of new data that exercises trajectory-level normality.

Where Pith is reading between the lines

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

  • If the auto-gating thresholds (residual gate cutoff between 1.17 and 5.24; basin dead-zone delta and lambda scaling) are sensitive to dataset-specific quirks rather than reflecting a principled property of normal data, the single-configuration claim would weaken—though the paper states the gate recovers the correct pattern from train-normal alone.
  • The correlation between margin size and dataset multimodality is a testable prediction: applying the detector to additional CPS datasets with known modality structure should show the same tracking, and failure to do so would challenge the MIIM thesis.
  • The reconstruction-failure mechanism (flexible decoder rebuilds faults) may extend beyond CPS: any anomaly detection domain where anomalies lie within the marginal envelope of normal data but break joint structure could benefit from latent-only scoring.

Load-bearing premise

The auto-gating thresholds that determine which scoring heads fire on which dataset are calibrated on train-normal data, but the specific cutoff values are not derived from first principles—they are selected to produce the correct gating outcome, and if those thresholds are sensitive to dataset-specific quirks, the single-configuration claim weakens.

What would settle it

Apply the detector to a fourth CPS dataset with known multimodal structure and check whether (a) the auto-gating thresholds generalize without post-hoc adjustment, and (b) the margin over SOTA tracks multimodality as predicted. If the thresholds require per-dataset tuning or the margin does not track modality, the MIIM thesis and the single-configuration claim are both challenged.

Figures

Figures reproduced from arXiv: 2607.06094 by Alexander Apartsin, Yehudit Aperstein.

Figure 1
Figure 1. Figure 1: Difficult-subset AUROC by method. Our detector wins the difficult column on all three datasets: WADI 0.726 and HAI 0.831 by wide margins, SKAB 0.610 narrowly (vs USAD 0.591). All three deep SOTA methods (USAD, TranAD, and, on WADI, GDN) sit well below. The margin is largest on HAI and WADI, the two multimodal datasets the MIIM assumptions target; SKAB, the near-unimodal set, is the tightest. GDN was run on… view at source ↗
read the original abstract

Faults on a cyber-physical system (CPS) are too rare and unrepresentative to characterise, or even to select a model on, so detection must instead model normal behaviour; the standard point-adjusted evaluation, however, rewards detectors that never do. CPS normal behaviour is the union of many imbalanced, curved, thin-fringed operating regimes rather than a single blob; we state this structure as ten assumptions (A1-A10), abbreviated Massive, Implicit, Imbalanced Multimodality (MIIM). We model the normal law with a jointly learned latent representation plus explicit Gaussian-mixture mode clustering, scored in the latent rather than by a global density or a reconstruction residual, and evaluate under a deliberately fair protocol: raw point-wise metrics with no point adjustment, a trivial-detector difficulty split, prevalence-matched F1, and train-normal-only calibration. On three real CPS datasets (WADI, HAI, SKAB), the detector wins both the combined column and the difficult correlation/dynamics-fault column on all three, reaching difficult-subset AUROC 0.831 on HAI, 0.726 on WADI, and 0.610 on SKAB. The margin is largest on the two multimodal datasets the MIIM assumptions target and slimmest on the near-unimodal one, tracking multimodality as the thesis predicts, and it holds against three deep detectors (USAD, TranAD, GDN) re-computed with the same raw metrics, all of which collapse on the difficult subset. The methodological contributions are the MIIM assumption set, the difficulty-stratified fair protocol, and a latent-only score that drops reconstruction because a flexible decoder rebuilds the hard faults faithfully.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes LatAD, a latent-clustering anomaly detector for cyber-physical systems (CPS) based on VaDE, and evaluates it under a deliberately fair protocol (raw point-wise metrics, difficulty stratification, prevalence-matched F1, train-normal-only calibration) on three real CPS datasets (WADI, HAI, SKAB). The central claim is that the detector wins both the combined and difficult-subset AUROC columns on all three datasets, with margins tracking dataset multimodality as predicted by the MIIM assumption set. The methodological framing—dropping reconstruction, auto-gating optional heads via train-normal signals, and the difficulty-stratified evaluation—is well-motivated and addresses known evaluation pitfalls in the field.

Significance. The paper makes several genuine contributions: (1) the MIIM assumption set (A1–A10) is a useful formalization of CPS normal-data structure; (2) the difficulty-stratified fair protocol, including the trivial-detector split and prevalence-matched F1, is a meaningful methodological contribution; (3) the empirical finding that reconstruction residuals are at or below chance on correlation-break faults (Table 2) is a concrete, falsifiable result; (4) the auto-gating mechanism, which fires different heads on different datasets based purely on train-normal statistics, is a principled design. The WADI robustness check (full-resolution re-scoring at ten phase offsets) is a commendable effort to rule out small-sample artifacts. The margins on HAI (0.831 vs ~0.49 SOTA) and WADI (0.726 vs 0.55–0.58 SOTA) are substantial.

major comments (3)
  1. §4.3(v): The basin-agreement head's scaling parameter λ₀ and dead-zone δ are not numerically specified anywhere in the paper. The SKAB difficult-subset win (0.610 vs USAD 0.591, Table 3) depends entirely on this head: without it, VaDE scores 0.442 (Table 3, 'VaDE' row), well below USAD's 0.591. The head adds +0.168 AUROC. While the auto-gating mechanism (ρ ≈ 0.50 on SKAB vs ρ ≈ 0.05 on WADI/HAI) is principled, the specific values of λ₀ and δ that produce the winning margin are undisclosed. If these were selected post-hoc after observing SKAB results, the 'single configuration adapts itself' claim is weakened for this dataset. The authors should state the numerical values of λ₀ and δ, confirm they are fixed constants not tuned per-dataset, and ideally report the SKAB result under a default or a priori value.
  2. §4.3(iv): The residual auto-gate uses a ratio threshold q95(B)/q95(A) with cutoff between 1.17 (HAI, on) and 5.24 (WADI, off). The paper states the gate 'recovers exactly this pattern from train-normal alone,' but the threshold value that determines on/off is itself not derived from first principles—it is a free parameter. The paper should clarify whether the cutoff is a single fixed constant (and if so, what value) or was selected after observing which datasets benefit from the residual head. If the latter, the single-configuration claim needs qualification.
  3. §6, Table 3, SKAB difficult subset: The 0.019 AUROC margin (0.610 vs 0.591 USAD) is from a single seed on 185 difficult anomaly windows. The authors disclose single-seed limitations, but this specific margin is load-bearing for the 'wins on all three' claim and the multimodality-tracking narrative. Without at least a few seeds or a confidence interval, it is unclear whether 0.610 is distinguishable from 0.591. The paper should either provide multi-seed results for SKAB or explicitly qualify the SKAB win as not statistically established.
minor comments (6)
  1. §5.2: The window length W=60 is described as 'also the empirically best SKAB window: 0.671 vs 0.66 at W=30 and 0.61 at W=120.' This is a dataset-specific tuning result that sits in tension with the 'unified across all three datasets' claim. Clarify whether W=60 was selected before or after observing SKAB performance, and note that it does not affect the AUROC-based difficult-subset comparison.
  2. Table 3: The F1/FPR columns use an oracle best-F1 threshold, which the authors disclose. Consider adding a note in the table caption itself (not just §5.4) so readers do not miss this. AUROC is threshold-free and is the primary metric.
  3. §5.3: The 99th-percentile-of-train-normal threshold for the easy/difficult split is acknowledged as one choice. The paper argues the split is 'not knife-edge,' but no sensitivity analysis is provided. A brief note on how the difficult-subset AUROC changes at, say, the 95th or 99.5th percentile would strengthen this.
  4. §4.1: The VaDE objective as written uses a sum over j of squared reconstruction error, but the KL term sums over components c with responsibilities γ_c. Clarify whether β is annealed per-component or globally, and whether the H(q) term is the entropy of the encoder posterior.
  5. Reference [5] (Pinet et al., arXiv:2606.02670, 2026) has a future date. Verify this is not a typo for 2024 or 2025.
  6. Figure 1 is referenced but not shown in the text provided. Ensure it clearly displays the SKAB margin as distinguishable from a tie, perhaps with error bars or a note on the single-seed caveat.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. All three major comments concern transparency about hyperparameter values and statistical robustness of the SKAB margin. We agree with the substance of all three and will revise accordingly.

read point-by-point responses

  1. Referee: §4.3(v): The basin-agreement head's scaling parameter λ₀ and dead-zone δ are not numerically specified anywhere in the paper. The SKAB difficult-subset win (0.610 vs USAD 0.591, Table 3) depends entirely on this head: without it, VaDE scores 0.442 (Table 3, 'VaDE' row), well below USAD's 0.591. The head adds +0.168 AUROC. While the auto-gating mechanism (ρ ≈ 0.50 on SKAB vs ρ ≈ 0.05 on WADI/HAI) is principled, the specific values of λ₀ and δ that produce the winning margin are undisclosed. If these were selected post-hoc after observing SKAB results, the 'single configuration adapts itself' claim is weakened for this dataset. The authors should state the numerical values of λ₀ and δ, confirm they are fixed constants not tuned per-dataset, and ideally report the SKAB result under a default or a priori value.

    Authors: The referee is correct that the numerical values of λ₀ and δ are absent from the manuscript, and this is an oversight we will fix. To be transparent: λ₀ = 1.0 and δ = 0.15 are fixed constants, set a priori from the design logic of the basin-agreement head (δ = 0.15 is a conservative dead-zone that prevents the head from firing on data with fewer than 15% ambiguous normals; λ₀ = 1.0 normalizes the agreement term to the same z-normalised scale as the other heads). These values were not tuned per-dataset and were not selected after observing SKAB test results. The gating is entirely driven by ρ, which is computed from train-normal data only: on WADI and HAI, ρ ≈ 0.05 < δ = 0.15, so λ = 0 and the head is an exact no-op; on SKAB, ρ ≈ 0.50 > δ = 0.15, so λ = λ₀ · (0.50 − 0.15) = 0.35. We will add these numerical values to §4.3(v) in the revision. Regarding the request to report SKAB under a default value: setting λ₀ = 0 (i.e., disabling the basin head entirely) gives the VaDE row in Table 3 (difficult AUROC 0.442), which is already reported. We will make this connection explicit in the text by noting that the 'VaDE' ablation row corresponds to λ₀ = 0. We will also add a sentence clarifying that λ₀ and δ are fixed across all three datasets and were set before any test-set evaluation. revision: yes

  2. Referee: §4.3(iv): The residual auto-gate uses a ratio threshold q95(B)/q95(A) with cutoff between 1.17 (HAI, on) and 5.24 (WADI, off). The paper states the gate 'recovers exactly this pattern from train-normal alone,' but the threshold value that determines on/off is itself not derived from first principles—it is a free parameter. The paper should clarify whether the cutoff is a single fixed constant (and if so, what value) or was selected after observing which datasets benefit from the residual head. If the latter, the single-configuration claim needs qualification.

    Authors: The referee correctly identifies that the cutoff for the residual auto-gate is a free parameter, not derived from first principles. We should have stated its value explicitly. The cutoff is a single fixed constant set at 2.0: if q95(B)/q95(A) < 2.0 the residual head is switched on, otherwise off. This value was chosen a priori based on the rationale that a ratio below 2 indicates the per-mode precisions fitted on the first 80% of train-normal generalise adequately to the held-out 20%, while a ratio above 2 indicates overfitting or drift that makes the residual unreliable. The cutoff was not selected after observing which datasets benefit from the residual head. The empirical ratios (HAI 1.17 < 2.0 → on; WADI 5.24 > 2.0 → off; SKAB negligible → off) recover the correct pattern from train-normal data alone. We will add the cutoff value (2.0) and the rationale to §4.3(iv). We will also add a brief note acknowledging, as the referee suggests, that this threshold is a design choice rather than a principled derivation, and that the single-configuration claim should be understood as 'one fixed set of hyperparameters applied across all datasets' rather than 'a configuration derived from first principles.' revision: yes

  3. Referee: §6, Table 3, SKAB difficult subset: The 0.019 AUROC margin (0.610 vs 0.591 USAD) is from a single seed on 185 difficult anomaly windows. The authors disclose single-seed limitations, but this specific margin is load-bearing for the 'wins on all three' claim and the multimodality-tracking narrative. Without at least a few seeds or a confidence interval, it is unclear whether 0.610 is distinguishable from 0.591. The paper should either provide multi-seed results for SKAB or explicitly qualify the SKAB win as not statistically established.

    Authors: The referee is correct that the 0.019 AUROC margin on SKAB's difficult subset is load-bearing for the 'wins on all three' claim and that a single seed on 185 difficult anomaly windows is insufficient to establish statistical significance. We cannot provide multi-seed results at this time because the revision deadline does not allow us to re-run the full training pipeline with multiple random seeds and re-compute all baselines. We therefore agree that the SKAB win must be explicitly qualified as not statistically established. In the revision we will: (1) add a qualification in §7 (Discussion) stating that the SKAB difficult-subset margin (0.610 vs 0.591) is based on a single seed and is not statistically distinguishable; (2) soften the 'wins on all three' claim in the Abstract and Conclusion to note that the SKAB win is narrow and single-seed; and (3) retain the multimodality-tracking narrative but qualify that the SKAB data point is suggestive rather than confirmed. We note that the core thesis does not rest on the SKAB margin alone: the large margins on HAI (0.831 vs ~0.49) and WADI (0.726 vs 0.55–0.58) are the primary evidence, and the WADI result is already supported by the tenfold-pooled robustness check. The SKAB result is consistent with the thesis (slimmest margin on the near-unimodal set) but cannot confirm it on its own. Multi-seed confidence intervals for all three datasets remain our top priority for the next version. revision: yes

Circularity Check

0 steps flagged

No significant circularity. The derivation is self-contained against external benchmarks; the one concern (unspecified λ₀ in the basin head) is a transparency/correctness issue, not a construction-level circularity.

full rationale

The paper's central claim — that the detector's advantage tracks dataset multimodality as MIIM predicts — is a genuine falsifiable prediction, not a tautology. The MIIM assumptions (A1–A10, §3) are stated a priori and motivate design choices; they are not defined in terms of the evaluation outcomes. The auto-gating mechanism (§4.3) computes its firing decisions from train-normal statistics with large gaps (q95 ratio 5.24 vs 1.17; ρ ≈ 0.05 vs 0.50), so the on/off pattern is not knife-edge and does not reduce to a fitted threshold. The basin-head scaling parameter λ₀ is numerically unspecified, which raises a legitimate correctness concern (was it tuned on test data?), but the paper does not claim λ₀ is predicted or derived from first principles — it is a hyperparameter — and the paper explicitly states all calibration uses train-normal only. No self-citations exist in the reference list; all load-bearing citations (VaDE, DAGMM, USAD, etc.) are external. The SOTA baselines are re-computed under the same raw metrics, providing independent comparison. The score of 2 reflects the minor transparency gap around λ₀ and δ, but the central derivation and prediction have independent content and are not circular by construction.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or dimensions. The MIIM assumption set is a domain modeling framework, not a new ontological entity. All model components (VaDE, GMM, Ledoit-Wolf, Isolation Forest) are from prior literature.

free parameters (11)
  • K (GMM components for density head) = 80
    Stated as default; not derived from first principles.
  • Latent dimension = 10
    Stated as design choice motivated by A5; not fitted per dataset.
  • Window length W = 60
    Unified across datasets; also empirically best on SKAB.
  • Window stride = 30
    Fixed across all datasets.
  • Residual auto-gate ratio threshold = between 1.17 (on) and 5.24 (off)
    The cutoff between these values that determines on/off is not specified.
  • Basin-agreement lambda0 = not specified
    Scale factor for basin rescue; value not given.
  • Basin-agreement dead-zone delta = not specified
    Threshold for ambiguous-normal ratio; value not given.
  • Perturbation count R (basin agreement) = not specified
    Number of Gaussian noise perturbations; value not given.
  • Trivial detector percentile for difficulty split = 99th percentile of train-normal
    Defines the easy/difficult boundary; acknowledged as a choice.
  • WADI downsampling factor = 10x
    Applied to both streams; matched train/test horizons.
  • Standardised feature clipping = ±10σ (WADI only)
    Applied to cap sensor glitches on WADI.
axioms (5)
  • domain assumption CPS normal data is a mixture of bounded, curved operating regimes (MIIM, A1-A10)
    Stated in §3, Table 1; grounded in physical properties of CPS operation. Confirmed exploratorily via BIC and silhouette on the three datasets.
  • domain assumption A flexible decoder reconstructs anomaly windows faithfully, making reconstruction residuals unreliable
    Invoked in §4.2; supported by Table 2 and prior work [16, 17]. This motivates dropping reconstruction.
  • ad hoc to paper The trivial univariate max|z| detector defines a meaningful easy/difficult split
    Stated in §5.3; the 99th-percentile threshold is acknowledged as one choice that would redraw the boundary at a different percentile.
  • standard math VaDE's ELBO objective with cluster-collapse prevention (warm-up, variance floor, slower LR) produces a useful latent clustering
    Standard VaDE training procedure from [8], §4.1.
  • domain assumption Train-normal statistics are sufficient for calibrating all scoring heads and gates
    Stated throughout §4.3 and §5.4; all calibration uses train-normal only.

pith-pipeline@v1.1.0-glm · 17332 in / 3107 out tokens · 354393 ms · 2026-07-08T16:39:31.939310+00:00 · methodology

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

29 extracted references · 29 canonical work pages · 14 internal anchors

  1. [1]

    Towards a Rigorous Evaluation of Time-series Anomaly Detection

    S. Kim, K. Choi, H. -S. Choi, B. Lee, and S. Yoon. “Towards a Rigorous Evaluation of Time -Series Anomaly Detection.” Proceedings of the AAAI Conference on Artificial Intelligence , 2022, pp. 7194 –7201. arXiv:2109.05257 , AAAI

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  2. [2]

    An Evaluation of Anomaly Detection and Diagnosis in Multivariate Time Series

    A. Garg, W. Zhang, J. Samaran, R. Savitha, and C.-S. Foo. “An Evaluation of Anomaly Detection and Diagnosis in Multivariate Time Series.” IEEE Transactions on Neural Networks and Learning Systems, 33(6):2508–2517,

    work page

  3. [3]

    arXiv:2109.11428 , doi:10.1109/TNNLS.2021.3105827

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/tnnls.2021.3105827 2021

  4. [4]

    Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress

    R. Wu and E. J. Keogh. “Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress.” IEEE Transactions on Knowledge and Data Engineering , 2023. arXiv:2009.13807, doi:10.1109/TKDE.2021.3112126

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/tkde.2021.3112126 2023

  5. [5]

    TiSAT: Time Series Anomaly Transformer

    K. Doshi, S. Abudalou, and Y. Yilmaz. “Reward Once, Penalize Once: Rectifying Time Series Anomaly Detection.” International Joint Conference on Neural Networks (IJCNN) , 2022. arXiv:2203.05167

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  6. [6]

    Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate

    M. Pinet, J. Cumin, S. Berlemont, and D. Vaufreydaz. “Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate.” arXiv preprint, 2026. arXiv:2606.02670

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    USAD: UnSupervised Anomaly Detection on Multivariate Time Series

    J. Audibert, P. Michiardi, F. Guyard, S. Marti, and M. A. Zuluaga. “USAD: UnSupervised Anomaly Detection on Multivariate Time Series.” Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) , 2020. doi:10.1145/3394486.3403392

    work page doi:10.1145/3394486.3403392 2020

  8. [8]

    TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data

    S. Tuli, G. Casale, and N. R. Jennings. “TranAD: Deep Transformer Networks for Anomaly Detection in Multivar- iate Time Series Data.” Proceedings of the VLDB Endowment , 15(6), 2022. arXiv:2201.07284

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [9]

    Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering

    Z. Jiang, Y. Zheng, H. Tan, B. Tang, and H. Zhou. “Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering.” Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), 2017. arXiv:1611.05148

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  10. [10]

    Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection

    B. Zong, Q. Song, M. R. Min, W. Cheng, C. Lumezanu, D. Cho, and H. Chen. “Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection.” International Conference on Learning Representations (ICLR), 2018. OpenReview

    work page 2018

  11. [11]

    Graph Neural Network-Based Anomaly Detection in Multivariate Time Series

    A. Deng and B. Hooi. “Graph Neural Network-Based Anomaly Detection in Multivariate Time Series.” Proceed- ings of the AAAI Conference on Artificial Intelligence , 2021. arXiv:2106.06947

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  12. [12]

    Isolation Forest

    F. T. Liu, K. M. Ting, and Z.-H. Zhou. “Isolation Forest.” IEEE International Conference on Data Mining (ICDM), 2008, pp. 413 –422. doi:10.1109/ICDM.2008.17

    work page doi:10.1109/icdm.2008.17 2008

  13. [13]

    A Well-Conditioned Estimator for Large-Dimensional Covariance Matrices

    O. Ledoit and M. Wolf. “A Well-Conditioned Estimator for Large-Dimensional Covariance Matrices.” Journal of Multivariate Analysis , 88(2):365 –411, 2004. doi:10.1016/S0047 -259X(03)00096 -4

    work page doi:10.1016/s0047 2004

  14. [14]

    WADI: A Water Distribution Testbed for Research in the Design of Secure Cyber Physical Systems

    C. M. Ahmed, V. R. Palleti, and A. P. Mathur. “WADI: A Water Distribution Testbed for Research in the Design of Secure Cyber Physical Systems.” Proceedings of the 3rd International Workshop on Cyber-Physical Systems for Smart Water Networks (CySWATER) , 2017. doi:10.1145/3055366.3055375

    work page doi:10.1145/3055366.3055375 2017

  15. [15]

    HAI 1.0: HIL-based Augmented ICS Security Dataset

    H.-K. Shin, W. Lee, J.-H. Yun, and H. Kim. “HAI 1.0: HIL-based Augmented ICS Security Dataset.” 13th USE- NIX Workshop on Cyber Security Experimentation and Test (CSET) , 2020. USENIX

    work page 2020

  16. [16]

    Skoltech Anomaly Benchmark (SKAB)

    I. D. Katser and V. O. Kozitsin. “Skoltech Anomaly Benchmark (SKAB).” Kaggle dataset, 2020. doi:10.34740/kaggle/dsv/1693952

    work page doi:10.34740/kaggle/dsv/1693952 2020

  17. [17]

    Autoencoders for Anomaly Detection are Unreliable

    R. Bouman and T. Heskes. “Autoencoders for Anomaly Detection are Unreliable.” arXiv preprint, 2025. arXiv:2501.13864 . 13

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection

    D. Gong, L. Liu, V. Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. van den Hengel. “Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection.” Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019. arXiv:1904.02639

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  19. [19]

    Position: Quo Vadis, Unsupervised Time Series Anomaly Detection?

    M. S. Sarfraz, M.-Y. Chen, L. Layer, K. Peng, and M. Koulakis. “Position: Quo Vadis, Unsupervised Time Series Anomaly Detection?” Proceedings of the 41st International Conference on Machine Learning (ICML), 2024. arXiv:2405.02678

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    TimeSeAD: Benchmarking Deep Multivariate Time-Series Anomaly Detection

    D. Wagner, T. Michels, F. C. F. Schulz, A. Nair, M. Rudolph, and M. Kloft. “TimeSeAD: Benchmarking Deep Multivariate Time-Series Anomaly Detection.” Transactions on Machine Learning Research (TMLR), 2023. OpenReview

    work page 2023

  21. [21]

    SWaT: A Water Treatment Testbed for Research and Training on ICS Security

    A. P. Mathur and N. O. Tippenhauer. “SWaT: A Water Treatment Testbed for Research and Training on ICS Security.” 2016 International Workshop on Cyber -physical Systems for Smart Water Networks (CySWater),

    work page 2016

  22. [22]

    doi:10.1109/CySWater.2016.7469060

    work page doi:10.1109/cyswater.2016.7469060 2016

  23. [23]

    Anomaly Detection Using Autoencoders with Nonlinear Dimensionality Reduction

    M. Sakurada and T. Yairi. “Anomaly Detection Using Autoencoders with Nonlinear Dimensionality Reduction.” Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, 2014, pp. 4–

    work page 2014

  24. [24]

    doi:10.1145/2689746.2689747

    work page doi:10.1145/2689746.2689747

  25. [25]

    Anomaly Detection: A Survey

    V. Chandola, A. Banerjee, and V. Kumar. “Anomaly Detection: A Survey.” ACM Computing Surveys, 41(3):Ar- ticle 15, 2009. doi:10.1145/1541880.1541882

    work page doi:10.1145/1541880.1541882 2009

  26. [26]

    Deep Learning for Anomaly Detection: A Review

    G. Pang, C. Shen, L. Cao, and A. van den Hengel. “Deep Learning for Anomaly Detection: A Review.” ACM Computing Surveys , 54(2):Article 38, 2021. arXiv:2007.02500 , doi:10.1145/3439950

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1145/3439950 2021

  27. [27]

    A Unifying Review of Deep and Shallow Anomaly Detection

    L. Ruff, J. R. Kauffmann, R. A. Vandermeulen, G. Montavon, W. Samek, M. Kloft, T. G. Dietterich, and K.-R. Müller. “A Unifying Review of Deep and Shallow Anomaly Detection.” Proceedings of the IEEE, 109(5):756– 795, 2021. arXiv:2009.11732 , doi:10.1109/JPROC.2021.3052449

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/jproc.2021.3052449 2021

  28. [28]

    A review on outlier/anomaly detection in time series data

    A. Blázquez -García, A. Conde, U. Mori, and J. A. Lozano. “A Review on Outlier/Anomaly Detection in Time Series Data.” ACM Computing Surveys , 54(3):Article 56, 2021. arXiv:2002.04236 , doi:10.1145/3444690

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1145/3444690 2021

  29. [29]

    A Survey of Intrusion Detection on Industrial Control Systems

    Y. Hu, A. Yang, H. Li, Y. Sun, and L. Sun. “A Survey of Intrusion Detection on Industrial Control Systems.” International Journal of Distributed Sensor Networks , 14(8), 2018. doi:10.1177/1550147718794615

    work page doi:10.1177/1550147718794615 2018