acceptodds
Under review as a conference paper at ICLR 2027

Streaming Anomaly Detection in Contaminated Manifold-Structured High-Dimensional Data

Abstract

High-dimensional observations, such as material spectra and industrial process measurements, often concentrate near low-dimensional manifolds and arrive as streams during continuous scanning and process monitoring. Detecting anomalies in these streams requires learning emerging normal patterns without absorbing sparse anomalies into the background. Both emerging normal patterns and anomalies can receive high anomaly scores from a model trained on earlier observations, so its scores alone cannot determine which samples should guide subsequent learning. We propose Evolving Score Fields (ESF), a streaming anomaly detection framework for manifold-structured observations. By learning gradients of the log density, its score model implicitly represents the background distribution and its local geometry. ESF distinguishes compatibility with the learned background from support within the current block. Score-direction consistency measures departure from the learned field, while neighborhood support allows well-supported new patterns to enter adaptation even when the existing field assigns them high anomaly scores. Selected samples update the field through local denoising score matching with LoRA adapters and historical replay. The adapted field scores the current block, and a sample-aware merge carries the update into subsequent blocks. This connects anomaly detection and background learning within one evolving field, without anomaly labels or full-model retraining. Experiments on HyperAD and sparse-fault TEP streams demonstrate ESF's strong detection performance across spectral and industrial data.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral Anomaly Detection, Streaming Anomaly Detection, Score-based Generative Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NZ9cOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers streaming anomaly detection when the normal distribution changes over time and incoming data are contaminated by anomalies. The main difficulty is deciding which observations can safely be used for updating the detector. In particular, samples from a new normal component may look anomalous under the current model, while updating on all incoming samples risks incorporating genuine anomalies into the estimated background. The authors propose Evolving Score Fields (ESF). The method uses a neural score field to characterize the background distribution and measures anomalies through the consistency of score directions under perturbations. For adaptation, it combines this score-based measure with neighborhood support computed from the current block. Selected observations are then used for local denoising score matching with LoRA adapters and historical replay. A sample-aware merge transfers part of the local update to the model used for subsequent blocks. The experiments cover three HyperAD scenes and sparse-fault streams derived from the Tennessee Eastman Process (TEP). ESF performs particularly well on the hyperspectral benchmarks and also obtains strong results on TEP.

Critical strengths

The problem addressed in this paper is relevant to practical streaming anomaly detection. I agree with the authors that rejecting all high-scoring observations can prevent a detector from learning new normal patterns. Conversely, updating on every incoming observation can gradually reduce sensitivity to anomalies. This tension is important and deserves more attention.

In Table 1, ESF achieves an AUPRC of 0.8749 on SNP and 0.4111 on Beach, both substantially higher than the reported baselines. Table 3 also provides useful evidence for the importance of selecting adaptation samples: removing confidence-based selection reduces the mean AUPRC from 0.5630 to 0.1700. This suggests that the improvement is not simply due to continuously updating a score model.

Critical weaknesses
  1. My main concern is the assumption that neighborhood support provides sufficient evidence for admitting unfamiliar observations into adaptation. In Section 3.2, support is estimated using the distance to the K-th nearest neighbor. The authors argue that choosing K above the anticipated anomaly count prevents isolated anomaly groups from receiving high support. This reasoning seems plausible for sparse, isolated anomalies, but I am less convinced when abnormal observations form a persistent or structured cluster. For instance, a repeated sensor fault could produce a compact group of observations. If the group becomes sufficiently prevalent, it may receive high neighborhood support even though it does not represent a new normal operating condition. Moreover, Eqs. (11)–(12) increase the weight assigned to neighborhood support when unfamiliar observations become widespread. I wonder whether this mechanism could reinforce the selection of such abnormal clusters.

  2. The HyperAD experiments are convincing in terms of detection performance, and the TEP experiment provides evidence beyond spectral observations. However, I found it difficult to determine how much of the improvement comes from learning new normal components, as opposed to adapting to relatively smooth changes in the observed background. The controlled A-to-B experiment in Appendix D.2 helps explain how newly appearing normal samples can enter adaptation. Still, the paper would benefit from a more systematic comparison of gradual drift, abrupt shifts, and the appearance of multiple new components within a short period.

  3. The paper motivates score fields through their ability to represent local manifold geometry without explicitly parameterizing the manifold. This is reasonable, and the direction-consistency measure in Eq. (9) provides a specific way to use that geometry for anomaly detection. However, the current experiments make it difficult to separate the benefit of score-field modeling from the benefit of the overall adaptation strategy. For example, Table 3 shows that confidence-based sample selection is important, but it does not establish whether a score field is necessary to obtain this improvement. Could a comparable selection and replay strategy also work with an autoencoder or another learned representation?

Questions for authors
  1. In the persistent-contamination experiment, how are anomalous samples distributed across blocks? In particular, does the experiment include coherent anomaly clusters that remain present over several consecutive blocks?

  2. What happens when the anomaly proportion approaches or exceeds the neighborhood-size assumption used to choose K? Is there a contamination level at which the selection mechanism begins to admit a substantial number of anomalies?

  3. Can the authors provide a controlled experiment comparing gradual background evolution with an abrupt shift? It would be useful to report both the detection performance immediately after the shift and the number of blocks required for adaptation.

  4. How much of the performance gain comes specifically from score-direction consistency? A comparison using the same sample-selection and adaptation procedure with an alternative representation would help clarify the role of score-based modeling.

Reviewer_43LVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies an anomaly detection problem in high-dimensional data in a streaming manner.

The key idea is to first score the samples by directional consistency and then choose samples from a combination of the score and the NN support, involving a LoRA for model updates.

Experiments are done in three hyperspectral scenes and show its effectiveness.

Critical strengths
  1. The paper is relatively easy to follow. The figure clearly shows the key components of the paper.

  2. The idea is not too complicated and is shown to be useful. The mechanism sounds intuitive to me.

  3. The paper is thorough evaluation metrics, seed variation analysis and a controlled synthetic study.

Critical weaknesses
  1. It is not clear how much of the gain comes from the so-called adaptation procedure. Looking at the results like Table 3, without any online update the results is already well above other baselines, and other baselines are mainly Gaussian RX variances, which are too weak. There is a concern of overengineering, and a concern that the baseline is too weak. Consider a more substantial baseline evaluation like neural and tree based streaming detectors.

  2. The assumption shall be more carefully validated on the data. The selection has an assumption that the anomalies being scattered <30% of the blocks, this is much stronger than what the Eq. 3 appears to claim. Also the synthetic strees test shall go further to support the claim.

  3. No runtime information; we need a fair comparison across methods with all info faithfully reported. The sensitivity study is also a concern, no enough tuning information to let people understand the generalizability.

  4. The paper, though well motivated, has a relatively modest contribution to the field, as the ideas are similar to direction-consistency score (ScoreAD), active sample selection (CANDI), and standard LoRA fitting and weight merging. It would be helpful to support the necessity of the proposed component with more targeted analysis.

Questions for authors

Please see the weakness section.

Reviewer_hpNkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Evolving Score Fields (ESF), a streaming anomaly detector that scores each sample by the consistency of normalized denoising-score-matching directions at Gaussian perturbations and adapts the same score field block by block with LoRA on samples selected by mixing field-based confidence with within-block k-NN support, followed by replay and a partial merge into the base weights. The selection is designed so that well-supported emerging normal patterns are learned while sparse anomalies are not absorbed. ESF is evaluated on three HyperAD hyperspectral line-scan scenes and 28 sparse-fault Tennessee Eastman streams, supported by an affine-manifold analysis of the score statistic and a synthetic diagnostic stream.

Critical strengths
  • A ScoreAD-style score field [1] is continually re-fitted by block-wise local DSM and rescores each block, while a persistent gate (h_t, Eqs. 10–12) shifts sample selection from field compatibility to within-block k-NN support when unfamiliar samples become widespread, so a well-supported new component can train the model before it is scored as normal. The principle mirrors cohesion-based stream novelty detection [2,3] and CANDI [4], but its use to steer continual adaptation of a generative score model is, to my knowledge, new; prior streaming detectors on HyperAD use recursive mean/covariance or PCA statistics [5].
  • The problem setup (Eqs. 3–6), Algorithm 1, all settings (Table 4), metric definitions (App. C.2) and limitations (App. F) are explicit, making the method precise and reproducible. The controlled stream of App. D.2 makes the selection mechanism easy to inspect: 3 of 1,224 trained samples are anomalies, against 230 under random selection.

References

[1] J. Sheng, Y. Shi, S. Xiang, X. Li, and S. Chen. Utilizing the Score of Data Distribution for Hyperspectral Anomaly Detection. arXiv:2601.12379, 2026. [2] E. J. Spinosa, A. C. P. L. F. de Carvalho, and J. Gama. OLINDDA: A Cluster-Based Approach for Detecting Novelty and Concept Drift in Data Streams. In Proc. ACM Symposium on Applied Computing (SAC), pp. 448–452, 2007. [3] E. R. de Faria, A. C. P. L. F. de Carvalho, and J. Gama. MINAS: Multiclass Learning Algorithm for Novelty Detection in Data Streams. Data Mining and Knowledge Discovery, 30(3):640–680, 2016. [4] H. Kim, J. Mok, H. Lee, J. Shin, and S. Yoon. CANDI: Curated Test-Time Adaptation for Multivariate Time-Series Anomaly Detection Under Distribution Shift. In Proc. AAAI Conference on Artificial Intelligence, 40(17):15018–15026, 2026. [5] S. Garske, B. Evans, C. Artlett, and K. C. Wong. ERX: A Fast Real-Time Anomaly Detection Algorithm for Hyperspectral Line Scanning. IEEE Transactions on Geoscience and Remote Sensing, 63:1–17, 2025. HyperAD code: https://github.com/WiseGamgee/HyperAD

Critical weaknesses
  • Novelty is limited and partly unattributed: the anomaly score (Eq. 9) is ScoreAD's statistic [1] divided by M, with the same Gaussian perturbations and perturbation time t = 0.05, yet [1] is cited only for intuition and never compared against, and Props. A.1–A.2 reduce to the mean resultant length of a projected normal [6] plus an elementary second-moment identity. The adaptation parts are also established (score-gated updates against memory poisoning [7], LoRA merge-and-reinitialize [8] and fractional merging [9] in continual learning, replay [10]), and because m_t is fixed by the block size, the "sample-aware" merge ρ_t (Eq. 17) is a data-independent schedule converging to 1−η, i.e., an EMA of LoRA increments.
  • On HyperAD, the frozen initial field without any online update already obtains about half of ESF's three-scene AUPRC margin over the per-scene best baseline (0.153 of 0.296; Table 3), yet ScoreAD [1], the streaming baselines used on TEP (Window-kNN, Online-iForest [11], xStream [12], ARCUS [13]) and external adaptation strategies such as score-gated updates [7] are not run on HyperAD, so the contribution of the streaming mechanism is not isolated. Baseline scores also appear to keep HyperAD's per-line z-normalization (App. C.3; normalise_md: True in the code of [5]) while ESF's raw scores are pooled, although only 112/2301, 245/2400 and 596/2973 evaluation columns of Synthetic, SNP and Beach contain any anomaly, an asymmetry that likely depresses the baselines' AUPRC.
  • The central claim is never stressed where it matters: no HyperAD evaluation column has more than 23% anomalous pixels (released masks; below ESF's own 0.25 bound and K_t = 0.3 n_t), replay has no measurable effect (AUPRC 0.5630 vs. 0.5619), and TEP faults are inserted as 74 isolated i.i.d. samples instead of the persistent post-onset faults of standard TEP evaluation [14,15], so the persistent, well-supported anomalies that App. F concedes ESF would absorb are examined only on a 16-D toy stream. Results also rest on three seeds and single-seed ablations whose gaps are far below the across-seed spread (Beach AUPRC s.d. 0.053), with ≈30 hyperparameters (Table 4) but no sensitivity analysis and a TEP input scale "fixed during earlier HyperAD Synthetic development, using whole-cube normalization" (App. C.4.1).

References

[6] B. Presnell and P. Rumcheva. The Mean Resultant Length of the Spherically Projected Normal Distribution. Statistics & Probability Letters, 78(5):557–563, 2008. [7] S. Bhatia, A. Jain, S. Srivastava, K. Kawaguchi, and B. Hooi. MemStream: Memory-Based Streaming Anomaly Detection. In Proc. ACM Web Conference (WWW), pp. 610–621, 2022. [8] X. Wei, G. Li, and R. Marculescu. Online-LoRA: Task-Free Online Continual Learning via Low Rank Adaptation. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 6634–6645, 2025. [9] I. E. Marouf, S. Roy, E. Tartaglione, and S. Lathuilière. Weighted Ensemble Models Are Strong Continual Learners. In Proc. European Conference on Computer Vision (ECCV), 2024. [10] D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, and G. Wayne. Experience Replay for Continual Learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. [11] F. Leveni, G. Weigert Cassales, B. Pfahringer, A. Bifet, and G. Boracchi. Online Isolation Forest. In Proc. International Conference on Machine Learning (ICML), PMLR 235:27288–27298, 2024. [12] E. Manzoor, H. Lamba, and L. Akoglu. xStream: Outlier Detection in Feature-Evolving Data Streams. In Proc. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2018. [13] S. Yoon, Y. Lee, J.-G. Lee, and B. S. Lee. Adaptive Model Pooling for Online Deep Anomaly Detection from a Complex Evolving Data Stream. In Proc. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), pp. 2347–2357, 2022. [14] E. L. Russell, L. H. Chiang, and R. D. Braatz. Fault Detection in Industrial Processes Using Canonical Variate Analysis and Dynamic Principal Component Analysis. Chemometrics and Intelligent Laboratory Systems, 51(1):81–93, 2000. [15] C. Reinartz, M. Kulahci, and O. Ravn. An Extended Tennessee Eastman Simulation Dataset for Fault-Detection and Decision Support Systems. Computers & Chemical Engineering, 149:107281, 2021.

Questions for authors
  • Can you report ScoreAD [1] (both frozen after initialization and refit on a sliding window), Window-kNN, Online-iForest [11], xStream [12], ARCUS [13] and MemStream [7] on HyperAD, with all baselines re-scored without per-line z-normalization (or ESF with it)? Please also give the frozen field's per-scene, multi-seed results and pre- vs. post-update scores (Ã_t vs. A_t), to quantify what streaming adaptation and same-block refitting add.
  • How does ESF behave in three settings: TEP with persistent post-onset faults [14,15]; HyperAD with implanted group anomalies whose per-column share is swept past K_t/n_t = 0.3; and a new normal material occupying 5–20% of every column, which has fewer than K_t members, so its k-NN radius always reaches into other components? Please report detection delay and the fraction of anomalies admitted to T_t and the replay bank over time.
  • In your own affine model the score norm ‖s⁰_ν(x)‖ = dist(x, M)/ν² is a noise-free monotone function of distance, whereas Eqs. (24), (27) and (28) imply a detection limit Δ* ≈ σ⋆(2k)^{1/4}M^{−1/2} for A, so how does A compare with the score norm [16], the local DSM loss and the denoising residual σ⋆²‖s‖ [17] under identical adaptation? Please also provide ≥10-seed results, sensitivity to τ⋆, M, κ and α_init, and a tuning protocol that does not use the evaluation scenes.

References

[16] A. Mahmood, J. Oliva, and M. Styner. Multiscale Score Matching for Out-of-Distribution Detection. In Proc. International Conference on Learning Representations (ICLR), 2021. [17] B. Efron. Tweedie's Formula and Selection Bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.