{"id":"89e73fef-f4f6-4ebf-83f0-186f2612cb06","arxiv_id":"2507.08177","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper is a position piece advocating causal graph learning as the basis for interpretable, drift-robust anomaly detection in cyber-physical systems, with a small comparison table as supporting evidence.","lead":"This paper argues that security systems for critical infrastructure should detect cyberattacks by learning the cause-effect structure of the system, not just by spotting statistical outliers. It lays out three research directions for causal graph-based anomaly detection and compares published causal methods with deep learning baselines on water treatment testbeds.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's F1 scores come from different protocols and the two asserted SWaT numbers have no reproducibility detail, so the central 'match or exceed' claim is not yet evidence-backed.","rationale":"The reader's weakest assumption focuses on the faithfulness of learned causal graphs under violated assumptions (acyclicity, faithfulness, stationarity, causal sufficiency), which the paper itself acknowledges in Sections 4.2 and 4.3. That is a real and central risk for the causal mechanism, but it is an open research challenge rather than a checkable flaw in the paper's reported evidence. In contrast, the empirical claim in Section 3 is presented as current evidence and is directly checkable. Table 1 mixes F1 numbers from multiple sources without controlling for protocol, and the two new numbers are asserted without any experimental detail. This is the most load-bearing concern because the reader cannot verify whether the headline 'match or exceed' is a fair comparison or an artifact of differing evaluation setups. The reader's rationale does note that the F1 evidence is 'too thinly documented,' so there is partial overlap, but the formal weakest assumption was about graph faithfulness rather than evidence reproducibility. My recommended verdict is unchanged: the paper remains UNVERDICTED. A vision paper can legitimately propose a research direction without full experimental validation, but its central empirical assertion needs a reproducible protocol before it can be accepted as evidence. The proposed test would settle this by quantifying whether the reported F1 differences survive a controlled comparison. This does not require rejecting the paper; it requires treating the empirical claim as unverified, which is consistent with the reader's UNVERDICTED verdict.","tokens_in":9408,"tokens_out":4876,"duration_ms":56543,"concrete_test":"Re-evaluate all six models on the standard SWaT and WADI splits under one protocol: identical preprocessing, same attack windows, threshold picked by best F1 on a validation segment, and one scoring rule (e.g., segment-wise or point-adjusted F1). Report mean and standard deviation over five random seeds, and require the authors to release the configuration for the asserted F1-0.84 and F1-0.85 results. If Entropy-CGAD's SWaT advantage over USAD (0.82 versus 0.79) falls within one standard deviation, or CausAnom's WADI 0.35 is not significantly above USAD's 0.25, the 'match or exceed' claim loses quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3's central empirical claim—'causal methods match or exceed deep learning performance, while providing interpretable structural alerts'—rests entirely on Table 1, which aggregates F1 scores from separate publications (SMV-CGAD [17], CausAnom [30], Entropy-CGAD [11]) against deep baselines (USAD, VAE-LSTM, DAGMM) without a unified protocol. F1 on SWaT and WADI is highly sensitive to threshold selection, attack-segment definition, and point-adjusted versus event-based scoring; none of these choices are reported for the aggregated entries. The two asserted new numbers (F1-0.84 for divergence-aware profiling and F1-0.85 for Continual-CGAD on SWaT) are given without model configuration, hyperparameters, training/attack split, or code, so they cannot be checked. Moreover, the 'interpretable structural alerts' part of the claim is never measured: the paper explicitly defers human-in-the-loop and root-cause localization studies to future work. If the F1 comparison is not reproducible under a common protocol, the primary evidence for preferring causal methods over deep baselines is unsupported, and the paper's call to pivot to causality-driven detection lacks its empirical anchor.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a vision/position paper that argues for replacing black-box deep anomaly detectors in cyber-physical systems with causality-driven methods. It organizes the proposal around three directions: divergence-aware causal graph profiling, multi-view causal graph fusion, and continual causal graph learning. Section 3 claims that causal methods match or exceed deep learning performance on the SWaT and WADI testbeds and provide interpretable structural alerts, based on Table 1 and two additional unreferenced F1 scores. The paper then discusses challenges such as scalability, hidden confounding, and assumption violations, and outlines future directions involving multi-modality, generative AI, and streaming causal learning.","tokens_in":9551,"tokens_out":6811,"duration_ms":70506,"significance":"If the central claim could be backed by a reproducible, common-protocol comparison, the paper would make a useful contribution by shifting the anomaly-detection conversation from accuracy to interpretability and drift robustness. The qualitative case is coherent: black-box detectors are brittle under distribution shift and cannot attribute causes, while causal graphs offer a principled alternative. The taxonomy of three research directions and the candid list of open problems in Section 4 are valuable for the community. However, the paper's empirical anchor, Table 1 and the two F1 values in Section 3, is not reproducible, and interpretability is never measured. As written, the contribution is a research agenda rather than a demonstrated result, and the strength of the claims exceeds the evidence. No code or artifacts are provided, which is acceptable for a position paper but not for the empirical statements it makes.","major_comments":[{"comment":"The sentence 'causal methods match or exceed deep learning performance' is supported only by Table 1, which aggregates F1 scores from references [17], [30], [11], [4], [15], and [44]. These scores come from separate publications with no common evaluation protocol. F1 on SWaT and WADI is highly sensitive to threshold selection, point-adjusted versus event-based scoring, and attack-segment definitions, none of which are reported for the aggregated entries. A cross-paper summary cannot support the comparative claim. Please either run a unified benchmark with a shared protocol and release the configuration, or explicitly recharacterize Table 1 as an anecdotal cross-paper summary and soften the claim accordingly.","section":"Section 3, Table 1"},{"comment":"The two additional F1 values, 0.84 for divergence-aware causal graph profiling and 0.85 for Continual-CGAD on SWaT, are unreferenced and come with no model configuration, hyperparameters, training/attack split, or code. Without reproducibility detail, these numbers cannot be verified and cannot serve as evidence for the practical viability of causal methods. They should be removed or replaced by a full experimental description.","section":"Section 3, 'Perspectives from Causal Paradigms'"},{"comment":"The second half of the central claim, 'while providing interpretable structural alerts,' is never evaluated. The paper explicitly defers human-in-the-loop and root-cause localization studies to future work. Since interpretability is the primary argued advantage over black-box detectors, the paper should either include a concrete interpretability assessment, such as localization accuracy or expert validation of discovered graph changes, or clearly present interpretability as a hypothesis rather than a demonstrated benefit.","section":"Section 3"},{"comment":"The detection principle relies on structural shifts in a learned causal graph being a reliable anomaly signal. Sections 4.2 and 4.3 concede that causal sufficiency, acyclicity, faithfulness, and stationarity are violated in real CPS. The paper provides no evidence that structural-shift detection remains reliable under these violations, despite the 'demonstrate' language in Sections 3 and 6. For a vision paper this can be reframed as an open research challenge, but as written the central promise is unsupported. Please either add a concrete stress-test under assumption violations or explicitly label robustness under these violations as an open question.","section":"Sections 4.2 and 4.3"}],"minor_comments":[{"comment":"In the final paragraph, 'advancesscalable' is a typo; it should read 'advances scalable.'","section":"Section 6"},{"comment":"References [41] and [42] are the same paper, 'Multi-modal Causal Structure Learning and Root Cause Analysis,' and should be merged into a single entry.","section":"References"},{"comment":"The heading 'Perspectives from Causal Paradigms' is unnumbered and appears without a corresponding subsection style; it should be formatted consistently with the rest of the paper.","section":"Section 3"},{"comment":"The phrase 'our own work' in the paragraph on divergence-aware profiling is vague; if it refers to the authors' prior or current research, it should be cited explicitly and consistently with the unreferenced F1 scores in Section 3.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper's empirical core relies substantially on the authors' own prior work, references [17], [33], and [34], and on two unreferenced F1 numbers in Section 3 that appear to be unpublished results from the same group. This should be disclosed clearly. The journal should also consider whether a vision paper with this level of empirical support meets its bar; if the authors cannot provide a common-protocol comparison, the comparative claims should be removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this as a position paper, not a research preprint. What's genuinely new: it organizes the causal anomaly detection space into three directions (divergence-aware profiling, multi-view fusion, continual graph learning) and argues for causal graph structure as the detection signal. That synthesis is coherent and the qualitative critique of black-box detectors is fair. The authors show good command of the relevant literature, including pointing to work like Sarfraz et al. that questions what deep models actually add.\n\nThe empirical support is the weak spot. Table 1 aggregates F1 scores from different papers with no common protocol—thresholds, point-adjusted vs. event-based scoring, attack segment definitions—none of it is reported, and SWaT/WADI F1 is known to be sensitive to those choices. The two new numbers in Section 3 (0.84, 0.85) have no configuration, code, data, or protocol, so they function as assertions, not evidence. Just as important, the \"interpretable structural alerts\" part of the central claim is never measured; the paper itself defers human-in-the-loop and root-cause localization to future work. So the claim that causal methods match or exceed deep learning while giving interpretable alerts is only half-supported: the F1 half isn't reproducible, and the interpretability half isn't evaluated.\n\nAlso worth noting: three of the four central references (SMV-CGAD, the incremental causal graph learning papers) are the authors' own prior work. Self-citation isn't a flaw by itself, but here the paper's own frameworks constitute the main empirical evidence, and in this setting that amplifies the reproducibility problem. The paper is honest about the load-bearing assumption: Sections 4.2 and 4.3 concede that causal sufficiency, acyclicity, faithfulness, and stationarity are violated in real CPS. If learned graphs are corrupted by those violations, the central promise collapses—and the paper has no evidence, no simulation, no stress test, that the approach survives them.\n\nDespite that, I'd give it a serious referee. It's a well-written vision statement in an area where a research agenda is useful to the community, and its stated directions aren't vacuous. The right outcome would be a major revision: either present the F1 comparison under a common protocol with full settings, or drop the empirical claim and make the paper explicitly a position piece. As it stands, the empirical section should not be taken as evidence.","headline":"A coherent, well-written vision paper whose central empirical claim isn't supported by reproducible evidence; worth refereeing as a position piece, not as a research result.","tokens_in":10176,"tokens_out":1355,"would_cite":false,"duration_ms":15069,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Causal graph models can match black-box deep learning on cyber-physical anomaly detection while also explaining why alerts fire.","keywords":["causal graph learning","anomaly detection","cyber-physical systems","spatio-temporal time series","structural causal models","continual causal graph learning","root cause attribution","SWaT and WADI testbeds"],"falsifier":"Run the same SWaT attack scenarios against the causal detectors while also injecting benign configuration changes (setpoint adjustments, valve maintenance, scheduled load changes) and compare alarm rates; if graph-divergence flags benign changes as often as statistical detectors do, the central promise of drift robustness collapses. A sharper version would publish per-attack detection latency of graph-shift alarms versus reconstruction-error alarms on the SWaT ground-truth attack timetable.","tokens_in":9141,"feed_emoji":"🛡️","tokens_out":10214,"duration_ms":96245,"temperature":0.7,"pith_summary":"This paper argues that anomaly detection in cyber-physical infrastructures should be built around learned cause-effect graphs instead of black-box deep models. Its central claim, supported by F1 scores on the SWaT and WADI water-treatment testbeds, is that causal methods match or exceed deep learning accuracy while also returning interpretable structural alerts and root-cause attribution. The paper organizes this program around three directions: divergence-aware causal graph profiling, multi-view causal graph fusion, and continual causal graph learning. The value, if the claim holds, is that operators would learn why a system is failing, not merely that it is failing, and detectors would keep working when sensor distributions drift.","feed_headline":"Causal graph methods match or beat deep learning in anomaly detection","feed_subtitle":"On water-treatment testbeds they also give root-cause alerts black-box models cannot.","key_machinery":"The machinery is a Structural Causal Model (SCM) cast as a directed graph whose nodes are sensors and actuators and whose edges encode directed cause-effect influences learned from streaming observations by constraint-based, score-based, or neural causal discovery. Anomalies are flagged by divergence between the graph learned under current operation and a reference graph learned during normal operation, measured by the appearance or disappearance of edges or by changes in edge strength. The three instantiations differ in how they obtain and update the graph: SMV-CGAD fuses domain knowledge, CausAnom learns flexible graphs with a graph neural network encoder, and Entropy-CGAD uses transfer entropy with weighted graph networks. Continual causal graph learning extends the same idea by updating the graph incrementally so that drift does not force full retraining.","core_discovery":"The central claim is that an anomaly in a cyber-physical system is better modeled as a structural change in the causal graph describing how sensors and actuators influence one another than as a statistical deviation in the observed time series. The paper compares three causal detectors (SMV-CGAD, CausAnom, and Entropy-CGAD) against deep baselines on SWaT and WADI and reports that the causal methods match or beat the deep models on F1, for example Entropy-CGAD at 0.82 on SWaT and SMV-CGAD at 0.79 on WADI versus 0.79 and 0.25 for USAD, while additionally providing interpretable, topology-sensitive alerts. It further reports that a divergence-aware profiling variant reaches 0.84 and a continual-update variant reaches 0.85 on SWaT. The paper does not claim to introduce a new method; it claims that this family of approaches deserves to be the default research direction for spatially distributed critical infrastructure.","pith_inferences":["Editorial inference: the strongest testable consequence left implicit is temporal lead, that structural graph shifts should be detectable before sensor values cross statistical thresholds, a hypothesis that could be checked directly against the SWaT attack timetable.","Editorial inference: if causal structure transfers across similar plants, a graph trained on one water-treatment configuration could bootstrap detection on another without new labels, which the paper does not test.","Editorial inference: the WADI column, where causal methods far outscore deep baselines (0.79 versus 0.25), hints that the real advantage appears under distribution shift, but a single testbed cannot separate that mechanism from dataset-specific effects.","Editorial inference: a practical deployment could combine graph divergence with an LLM-based explanation layer that narrates the changed edges; the paper lists LLMs as a future direction but does not make this concrete."],"forward_implications":["If the central claim is correct, causal detectors can be deployed where black-box accuracy was previously the only acceptable trade: SWaT and WADI results show parity or better F1 alongside structural explanations.","Root-cause attribution becomes a by-product of detection, because the edges that changed point at the sensors or actuators involved instead of requiring a separate explainer.","Drift robustness follows from the mechanism: if normal variability shifts marginal distributions but not causal structure, graph-divergence alarms fire less often than reconstruction-error alarms.","Continual causal graph learning removes the retraining bottleneck for evolving systems, provided the proposed online update methods scale to high-dimensional sensor networks.","Anomaly-detection benchmarks should start reporting interpretability and localization metrics alongside F1, because the paper's contribution is precisely that structural alerts carry information opaque scores hide."],"supporting_citations":[{"why":"Supplies the SWaT water-treatment testbed on which the central F1 comparison is run.","marker":"[18]"},{"why":"Supplies the WADI water-distribution testbed for the second half of the comparison.","marker":"[2]"},{"why":"SMV-CGAD, the causal baseline with domain-knowledge pruning, contributes its reported F1 scores.","marker":"[17]"},{"why":"CausAnom, the GNN-based causal encoder baseline, contributes its reported SWaT and WADI F1 scores.","marker":"[30]"},{"why":"Entropy-CGAD, the transfer-entropy weighted-GNN causal baseline, contributes the strongest causal F1 on SWaT.","marker":"[11]"},{"why":"USAD, a deep autoencoder baseline, anchors the comparison on the non-causal side.","marker":"[4]"},{"why":"VAE-LSTM, a deep hybrid baseline, anchors the comparison on the non-causal side.","marker":"[15]"},{"why":"DAGMM, a deep unsupervised baseline, completes the non-causal comparison.","marker":"[44]"},{"why":"Incremental causal graph learning supplies the continual-update variant reported at F1 0.85 on SWaT.","marker":"[33]"}],"fun_headline_variants":["Causal graph detectors rival deep learning and expose root causes","For cyber anomalies, causal graphs beat black-box on accuracy and insight","Anomaly detection: causal methods match deep nets and give root causes","Causality-driven anomaly detection: interpretable and competitive","Root-cause alerts from causal graphs match deep learning's accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach depends on the learned cause-effect graph being accurate enough that a real cyberattack shows up as a change in the graph itself, rather than being hidden by the unmeasured confounders, feedback loops, and non-stationary sensor behavior the paper itself says are common in cyber-physical systems.","fun_headline_variants_meta":{"raw":{"variants":["Causal graph detectors rival deep learning and expose root causes","For cyber anomalies, causal graphs beat black-box on accuracy and insight","Anomaly detection: causal methods match deep nets and give root causes","Causality-driven anomaly detection: interpretable and competitive","Root-cause alerts from causal graphs match deep learning's accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000841,"raw_usage":{"total_tokens":3677,"prompt_tokens":972,"completion_tokens":2705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2619}},"tokens_in":588,"tokens_out":2705,"duration_ms":20346,"temperature":1.0,"reasoning_tokens":2619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:24:54.590961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same SWaT attack scenarios against the causal detectors while also injecting benign configuration changes (setpoint adjustments, valve maintenance, scheduled load changes) and compare alarm rates; if graph-divergence flags benign changes as often as statistical detectors do, the central promise of drift robustness collapses. A sharper version would publish per-attack detection latency of graph-shift alarms versus reconstruction-error alarms on the SWaT ground-truth attack timetable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CausAnom, the GNN-based causal encoder baseline, contributes its reported SWaT and WADI F1 scores."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"USAD, a deep autoencoder baseline, anchors the comparison on the non-causal side."}],"review_version":1}