acceptodds
Under review as a conference paper at ICLR 2027

EDSR: Extreme Domain Shift Resilience via Causality-Guided Zero-Shot Anomaly Detection

Abstract

Anomaly detection is significantly challenging when multivariate time series are subjected to extreme domain shifts. These shifts can lead to magnitude variations of up to 100 between the training phase and deployment, occurring without any labeled anomaly examples for reference, in a zero-shot context. Existing models struggle with substantial domain shifts because they rely on statistical correlations that break down when major changes occur in the distribution. Therefore, we propose a hierarchical causality-guided framework (EDSR) that exploits the invariance of causal relationships; the physical laws remain constant regardless of magnitude scaling. EDSR integrates causal discovery, causality-aware contrastive encoding, gated attention refinement, adversarial manifold alignment, memory-based edge screening, and conditional domain adaptation into a unified pipeline that is resilient to 100 domain shifts. This hierarchical edge-cloud framework filters 85% of normal data at the edge in under 2 ms, and with end-to-end latency under 26 ms. Empirical evaluation demonstrates performance on synthetic benchmark datasets: VAR-100 and Lorenz-96. The EDSR under 100 magnitude shifts achieved F1 scores of 35.86% and 45.86%, respectively, outperforming baseline CAROTS method by 186% and 341%. Ablation studies revealed the critical role of adversarial refinement and attention mechanisms, particularly for nonlinear chaos. This study establishes causality as a foundational primitive for zero-shot anomaly detection in safety-critical deployments.

Primary area:
causal reasoning
Keywords:
Anomaly Detection, Causal Discovery, Zero-Shot Learning, Domain Adaptation, Extreme-Scale Systems

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_YYF8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes EDSR, a causality-guided framework for zero-shot multivariate time-series anomaly detection under extreme domain shifts. The main idea is to leverage causal relationships that are assumed to remain stable when the magnitude of the input signals changes significantly. The framework consists of six training phases, including causal discovery, causality-aware contrastive learning, spatio-temporal attention refinement, adversarial representation learning, memory-based screening, and domain adaptation. During inference, it adopts an edge-cloud architecture and combines multiple anomaly scores for detection and root-cause attribution. Experiments on two synthetic datasets, VAR-100 and Lorenz-96, show that EDSR achieves better detection performance than the selected baselines under 100× magnitude shifts.

Critical strengths
  1. The paper addresses a practically relevant problem, as time-series anomaly detectors may encounter substantial distribution shifts between training and deployment. Exploring structural relationships that may remain stable across operating conditions is a potentially interesting direction.

  2. The proposed framework also considers deployment efficiency through an edge-cloud architecture, aiming to reduce computational overhead by filtering normal samples before more expensive anomaly detection.

Critical weaknesses
  1. The paper argues that normalization may destroy useful information under extreme domain shifts. However, it is unclear what information would be lost, especially when the shift is simply a 100× magnitude scaling. This claim needs further justification.

  2. The proposed framework is overly complex, involving six training phases. The necessity of each component is not clearly justified, and the ablation studies are insufficient to support the overall design.

  3. The final anomaly score combines five weighted scores. Although the weights are selected through grid search, their sensitivity and the contribution of each score are not adequately investigated.

  4. The experiments are limited to two synthetic datasets. Evaluation on real-world benchmarks, such as SMD, SWaT, and PSM, would help demonstrate the practical effectiveness of the proposed method.

Questions for authors

See weaknesses.

Reviewer_Pw6tOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EDSR, a multivariate time-series anomaly detection framework targeting uniform amplitude shifts between training and deployment. It combines causal discovery, causality-aware contrastive learning, attention, adversarial reconstruction, and memory-based scoring within an edge–cloud architecture. Experiments on synthetic VAR-100 and Lorenz-96 datasets compare EDSR with USAD and CAROTS under original and 100× test amplitudes, reporting improved detection performance under the scaling shift.

Critical strengths
  1. The evaluation includes both linear VAR and nonlinear Lorenz systems, providing a controlled test of robustness across different dynamics. EDSR maintains comparable AUROC before and after 100× amplitude scaling on both datasets, offering initial evidence that its detection performance can withstand the specific scaling shift studied.
Critical weaknesses
  1. The evaluated shift is narrow, and the need for the proposed solution is not established. The experiments consider only a uniform multiplication of all test variables by 100 on two synthetic systems. This is a restricted form of domain shift, and its connection to the motivating real-world scenarios is insufficiently justified. Moreover, uniform positive scaling can be canceled by normalization using appropriately estimated window or recent-history statistics. Without comparisons against existing detectors equipped with such normalization, the results do not establish that the proposed causal framework is needed to address the evaluated challenge.
  2. The central contribution and the value of combining its components remain unclear. Robustness to amplitude scaling and edge–cloud deployment address largely separate objectives: detection robustness and computational efficiency. The paper does not establish a specific technical connection between them or demonstrate additional benefits from their integration. Furthermore, the ablations remove attention, memory, and the adversarial critic, but do not isolate the contribution of causal information through a non-causal or randomized-graph control. Consequently, it remains unclear whether the reported robustness comes from causal modeling, other network components, or their combination, weakening the justification for the overall framework.
Questions for authors
  1. Clarification and supporting evidence: How does EDSR compare with existing detectors equipped with window-based or recent-history normalization under the same scaling protocol? What evidence shows that causal modeling provides benefits beyond removing the uniform scale change?

  2. Clarification of causal contribution: Does the robustness persist when the learned causal graph is removed or replaced with a randomized graph, while keeping the remaining architecture and training procedure unchanged?

  3. Clarification of scope and integration: What deployment scenario motivates uniformly scaling all variables by the same factor, and what specific benefit does the edge–cloud architecture provide for handling this shift? Is its filtering effectiveness maintained under 100× scaling?

Reviewer_3FH2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EDSR, a multi-phase, causality-guided framework for zero-shot time-series anomaly detection under extreme domain shifts, particularly 100× magnitude scaling. The framework integrates causal discovery, contrastive representation learning, spatiotemporal attention, adversarial manifold alignment, memory-based screening, and KL-based domain gating within an edge-cloud architecture. Experiments on VAR-100 and Lorenz-96 report improved F1 scores over CAROTS under 100× scaling, along with ablation studies and latency evaluations.

Critical strengths

The paper addresses an important and relatively under-explored problem: zero-shot anomaly detection under extreme magnitude shifts, with an emphasis on exploiting causal information to improve robustness.

The framework covers multiple aspects of anomaly detection, from causal discovery and representation learning to edge-cloud inference, and provides a range of experimental evaluations.

The ablation studies highlight the contributions of adversarial refinement and spatiotemporal attention, particularly for nonlinear dynamics such as Lorenz-96. The paper also provides detailed descriptions of the proposed pipeline.

Critical weaknesses

The motivation and methodological narrative are insufficiently developed. The paper combines several established techniques, but the central novelty and the need for this particular combination remain unclear. The connections between individual components are not sufficiently justified, making the overall framework appear more like an integration of existing methods than a coherent methodological contribution.

The theoretical justification for causal invariance under magnitude scaling is insufficient. The rate of change of a homogeneous function is not generally invariant to scaling; additional assumptions, such as degree-one homogeneity, are required. Moreover, the Lorenz-96 system contains quadratic, linear, and constant terms, so uniformly scaling the test data by 100× does not directly establish the invariance of the learned causal representations. The authors should clarify the underlying assumptions and explain how they align with the experimental setting.

The empirical comparisons are limited. The empirical comparisons are limited, as EDSR is not compared with several relevant baselines, including ContrastAD, CGSTA, DACAD, DIVAD, Anomaly Transformer, and CANDI. Moreover, EDSR incurs substantially higher inference latency, raising concerns about whether its performance gains justify the additional computational cost.

The zero-shot evaluation protocol requires clarification. The anomaly-score weights are selected by maximizing AUROC on validation windows containing both normal and anomalous samples. This raises concerns about the use of anomaly labels, given the abstract's claim that labeled anomaly examples are unavailable. The authors should clarify whether these labels are restricted to the source domain and whether any target-domain anomaly labels are used during model selection.

The experimental setup is inconsistent across the paper, making the evaluation protocol confusing. The main text specifies an 80/20 train-test split and 5% test contamination, whereas Algorithm 4 specifies a 30% test split and a 1% anomaly ratio. The main text describes VAR-100 anomalies as causal-coefficient swaps, but Algorithm 2 only describes nominal VAR generation, while Algorithm 4 describes the anomaly injection procedure.

The manuscript contains numerous typographical and presentation issues. These include inconsistent naming, such as “GAROTS” versus “CAROTS,” and the use of “ATUAD” before it is defined. The adversarial objective is also described inconsistently in terms of embeddings and reconstruction. These issues, together with minor formatting problems, reduce the clarity of the presentation.

Questions for authors

Can EDSR recover the causal graph and maintain comparable detection performance when graph sparsity is selected without access to the ground-truth edge count?

Are anomalous validation labels used to optimize the anomaly-score weights? If so, are these labels restricted to the source domain, and how is the zero-shot setting defined?

Which exact anomaly-generation protocol, integration step size, data split, contamination ratio, and threshold percentile were used to obtain the reported results?

Does the reported latency represent average end-to-end inference with conditional routing, or the latency of processing each sample through the full pipeline?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.