{"id":"0d9d8086-910c-463a-876e-f7d135bebe88","arxiv_id":"2607.08978","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A federated Stiefel-ADMM low-rank Koopman method detects multivariate IoT anomalies with competitive F1 and far lower training, communication, and edge latency than neural federated baselines.","lead":"FedKAD learns shared low-rank Koopman dynamics across IoT devices so each edge node can flag abnormal time-series behavior without shipping raw sensors or heavy neural nets. It matches or beats federated deep detectors on standard benchmarks while cutting training time, bandwidth, and inference cost by one to three orders of magnitude.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Accuracy half of the claim rests on unproven expressiveness of fixed random tanh lifting plus shared rank-r subspace for residual scoring under non-IID MVTS.","rationale":"The reader's weakest_assumption correctly isolates the expressiveness of the fixed lifting + shared low-rank Koopman representation as the soft point of the accuracy half of the strongest claim. Efficiency evidence (Figs. 4–6, Raspberry Pi numbers) and the Stiefel-ADMM stationarity analysis (Sec. V, Thm. V.5–V.7) are independently solid and do not require a verdict change. Multi-protocol results already show the accuracy edge is protocol-sensitive, reinforcing that the model class may be only marginally sufficient rather than robustly superior. No internal inconsistency or stronger load-bearing flaw (e.g., in the optimization theory or efficiency accounting) was found. The paper remains a legitimate systems/ML contribution whose joint claim is therefore still CONDITIONAL on public artifacts and direct evidence that the representation fits normal dynamics well enough for residual scoring. The proposed ablation would settle the concern cleanly.","tokens_in":23910,"tokens_out":612,"duration_ms":35216,"concrete_test":"Train two ablations under identical federated envelope, seeds, and primary PA%K (k=0.01) protocol: (1) fully local OMD (each client optimizes its own P_i,Q_i with no consensus/Z), (2) linear no-lift variant (dlift=0). Recompute per-client macro F1 on all four datasets. If either ablation matches or exceeds FedKAD F1 on ≥2 datasets, the shared lifted low-rank model is not load-bearing for the accuracy claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim requires that FedKAD's detection performance maintains or improves on federated neural baselines. That rests on Sec. IV-A–C (Eqs. 7–11): a fixed shared random-feature lifting Φ (tanh, dlift=128) of the windowed state, followed by a shared orthonormal subspace Z of rank r=24–32 and private reduced operators Q_i, with anomalies scored by multi-step prediction residual. No approximation guarantee, no normal-data residual diagnostics, and no ablation of lifting/rank/sharing are provided. Under the documented non-IID skew (Fig. 2), a single low-rank shared subspace can underfit heterogeneous nonlinear modes; residuals then fail to separate anomalies cleanly. The reported 3/4 primary PA%K wins (Fig. 3, Table III) already erode under stricter protocols (Table IV), consistent with marginal expressiveness. If the representation is inadequate, the joint accuracy+efficiency claim collapses even though Stiefel-ADMM converges and wall-clock/comms numbers remain strong.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes FedKAD, a federated anomaly detector for multivariate IoT time series that replaces heavy neural models with sliding-window Koopman dynamics. Each client lifts windowed states with a fixed shared random tanh map, fits a private reduced operator Qi, and participates in consensus on a shared orthonormal subspace Z via a federated Stiefel-ADMM algorithm; anomalies are scored by multi-step prediction residual. The authors give convergence and stationarity analysis under partial participation (Assumptions A1–A5, Theorems V.5 and V.7), evaluate on SMD, PSM, SMAP, and MSL under non-IID partitions and multiple thresholding protocols, and report competitive primary PA%K F1 (best on 3/4 datasets) together with large reductions in training time, communication, and inference latency, including a Raspberry Pi 4 deployment study.","tokens_in":24210,"tokens_out":1282,"duration_ms":26495,"significance":"If the empirical trade-off holds, this is a practically useful contribution for edge IoT anomaly detection: it shows that a low-rank dynamical model with compact subspace consensus can match or beat federated deep baselines on standard MVTS benchmarks while being orders of magnitude cheaper to train and communicate. Strengths include an explicit non-IID characterization (Fig. 2), multi-protocol evaluation motivated by known PA pathologies, a Stiefel-ADMM formulation with partial-participation analysis, and concrete on-device measurements rather than only GPU wall-clock. The work is more systems- and deployment-oriented than a pure theory paper, but the efficiency evidence and honest multi-protocol tables make it a credible candidate for an IoT/edge-learning venue.","major_comments":[{"comment":"Sec. IV-A–C and Eqs. (7)–(11): the accuracy half of the central claim rests on a fixed random tanh lifting plus a shared rank-r subspace with private Qi being expressive enough for residual scoring under the documented non-IID skew (Fig. 2). The manuscript reports no ablation of dlift, r (beyond a PSM validation choice), shared Z vs fully local subspaces, or residual histograms on normal validation data. Without these, it is hard to know when the representation fails (e.g., MSL under the primary protocol, and several losses in Table IV under stricter PA%K/point-wise settings). Please add targeted ablations and a short failure-mode discussion so the joint accuracy–efficiency claim is not tied only to the default hyperparameter point.","section":null},{"comment":"Fig. 3 / Table III vs Table IV: the headline “best F1 on three of four datasets” is specific to PA%K with k=0.01. Under PA%K k=0.10 and strict point-wise no-PA, FedKAD loses on most datasets. The paper already reports these numbers, which is good, but the abstract and conclusion still emphasize “maintains or improves” without qualifying the protocol dependence. Please align abstract/conclusion wording with the multi-protocol evidence and state clearly that the main advantage is segment-level detection under mild PA%K rather than point-level localization.","section":null},{"comment":"Sec. VI-B4 / Fig. 6: on-device inference latency is stated as “0.79 µs/step” in the text but “786 µs” in the figure annotation for FedKAD under the SMAP setting—a 1000× discrepancy. Relative speedups versus baselines may still hold, but absolute numbers and the claimed “79× lower inference latency” need to be reconciled with a single consistent unit system before the efficiency claim can be cited.","section":null}],"minor_comments":[{"comment":"Algorithm 1 is written with full client participation (“for client i = 1, …, N”), while Sec. IV-D and the theory use partial sets Sℓ and Eq. (16). Please make the algorithm match the partial-participation case used in experiments (25% sampling).","section":null},{"comment":"Eq. (10) and the surrounding text: briefly state the condition under which Gi(Pi) is regularized (λ) and whether λ was tuned or fixed across datasets; a one-line sensitivity note would help reproducibility.","section":null},{"comment":"Notation: both FEDKAD and FedKAD appear; pick one spelling. Also “FEDKAD” in Fig. 1 caption vs “FedKAD” elsewhere.","section":null},{"comment":"Related work: a short comparison to other federated MVTS anomaly detectors (e.g., federated VAE / deep federated AD cited as [45]–[47]) on communication volume would situate the 80× claim more clearly.","section":null},{"comment":"Fig. 5: late-round Lagrangian oscillation is attributed to εℓ, ΔQℓ, and partial participation; adding the participation schedule or |Sℓ| on the plot would make that interpretation easier to verify.","section":null},{"comment":"Typos / polish: “It is easily to see that FEDKAD” (Sec. VI-B1); “PAon three of the four datasets” appears truncated in the main-results paragraph; arXiv-style “D”Iot” / encoding artifacts in references should be cleaned for the camera-ready.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid systems paper with unusually careful multi-protocol evaluation and real edge hardware numbers. The main risk is overselling accuracy relative to protocol choice; if the authors fix the µs discrepancy, add compact ablations, and tone the abstract to match Table IV, this is close to accept for an IoT/edge or applied FL venue. Not a pure theory contribution—the ADMM analysis is standard but adequate. I would not reject on expressiveness grounds alone given the empirical package, but I would not accept without the latency unit fix and clearer protocol qualification."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that FedKAD is a practical federated detector built around low-rank Koopman consensus, not another deep model wrapped in FedAvg. They federate only a shared orthonormal subspace Z, keep local reduced operators Q_i and raw snapshots on device, and solve it with Stiefel-ADMM plus partial-participation analysis. That combination is the real contribution.\n\nWhat they do well: the systems case is clean. Training is seconds not hours, communication is tens of KB to low MB, inference is microsecond-scale, and they actually put it on a Raspberry Pi 4. Four public MVTS benchmarks, non-IID partitions (entity splits plus Dirichlet on PSM), multi-protocol evaluation including stricter PA%K and point-wise F1, and paired tests against the strongest baseline. The theory is standard ADMM-style descent/stationarity under A1–A5, not decorative. Citations cover Koopman/OMD, Stiefel, FL, and the metric critique literature without obvious gaps.\n\nSoft spots, in proportion. The accuracy half of the headline is thinner than the efficiency half. Primary PA%K wins on 3/4 datasets (clear on SMD/PSM, soft on SMAP, small loss on MSL); under tighter protocols the ranking erodes. That is consistent with the stress-test concern: fixed random tanh lifting plus shared rank-r subspace is an expressiveness bet with no approximation argument, residual diagnostics, or lifting/rank/sharing ablations. Rank is tuned on validation (r=32 only on PSM). No public code is cited. None of that sinks the paper; it means the joint “as accurate and far cheaper” claim should be read as “competitive accuracy, large systems win.”\n\nThis is for people who care about deployable federated anomaly detection on constrained IoT hardware, not for pure dynamical-systems theory. It deserves a serious referee. I would engage: cite the efficiency and formulation, push for ablations and artifacts in revision.","headline":"Solid edge-FL systems paper: the efficiency story is real and well measured; the accuracy claim is competitive, not dominant, and rests on a fixed random lifting that is under-justified.","tokens_in":24825,"tokens_out":518,"would_cite":true,"duration_ms":11386,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"FedKAD learns shared low-rank Koopman dynamics so edge IoT devices can detect multivariate time-series anomalies without shipping raw data or training large neural models.","keywords":["Internet of Things","Federated Learning","Anomaly Detection","Multivariate Time Series","Koopman Operator","Stiefel Manifold","Edge Intelligence","ADMM"],"falsifier":"On the same four federated partitions, replace the fixed random lifting and rank-r shared subspace with a client-specific full-rank operator or a learned neural lifting of equal communication budget; if residual-based F1 under PA%K collapses relative to FedKAD while communication stays comparable, the expressiveness claim fails.","tokens_in":24835,"feed_emoji":"📡","tokens_out":648,"duration_ms":11419,"temperature":0.7,"pith_summary":"Distributed IoT sensors produce multivariate time series that must be monitored for faults and attacks, but the data sit on weak devices with non-IID streams and tight bandwidth. This paper argues that normal temporal behavior can be captured by a lightweight sliding-window Koopman model: a shared low-rank orthonormal subspace plus private reduced operators, rather than by federated deep networks. Only the compact subspace matrices are exchanged; raw streams and local dynamics stay on device. A federated Stiefel-ADMM procedure aligns those subspaces under partial client participation, with convergence and stationarity guarantees. On four standard benchmarks the method matches or beats federated neural detectors on segment-level F1 while cutting training time, communication volume, and inference latency by one to three orders of magnitude, including on a Raspberry Pi 4.","feed_headline":"IoT anomaly detection without heavy neural federated models","feed_subtitle":"Shared low-rank Koopman subspaces match neural F1 while slashing train time, bandwidth, and Pi-class latency.","key_machinery":"Federated Stiefel-ADMM on the OMD-style factorization Φ(z(t+1)) ≈ P Q Pᵀ Φ(z(t)), with P on the Stiefel manifold and Q client-private: clients exchange only compact subspace matrices P, the server forms a consensus Z, and dual variables enforce agreement under partial participation.","core_discovery":"The authors claim that federated multivariate time-series anomaly detection for constrained IoT can be solved by low-rank Koopman consensus: clients keep snapshot matrices and reduced operators private, agree only on a shared orthonormal subspace of the lifted windowed states, and flag anomalies by multi-step prediction residual. This formulation yields detection quality competitive with federated deep baselines while making training, communication, and on-device inference practical for edge hardware.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Federated low-rank Koopman for IoT multivariate anomaly detection","Shared Koopman subspaces enable local IoT anomaly detection","Resource-efficient federated anomaly detection via Koopman dynamics","Low-rank federated Koopman consensus for edge IoT time series","IoT anomalies flagged by private Koopman residual prediction"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A fixed random tanh lifting plus one shared low-rank orthonormal subspace (with private reduced operators) is expressive enough that prediction residual reliably separates normal from abnormal dynamics across non-IID clients.","fun_headline_variants_meta":{"raw":{"variants":["Federated low-rank Koopman for IoT multivariate anomaly detection","Shared Koopman subspaces enable local IoT anomaly detection","Resource-efficient federated anomaly detection via Koopman dynamics","Low-rank federated Koopman consensus for edge IoT time series","IoT anomalies flagged by private Koopman residual prediction"]},"model":"grok-4.5","effort":"low","cost_usd":0.00719,"raw_usage":{"total_tokens":1833,"prompt_tokens":856,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":71900000,"prompt_tokens_details":{"text_tokens":856,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":886,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":856,"tokens_out":91,"duration_ms":8175,"temperature":1.0,"reasoning_tokens":886,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:16:14.473380+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same four federated partitions, replace the fixed random lifting and rank-r shared subspace with a client-specific full-rank operator or a learned neural lifting of equal communication budget; if residual-based F1 under PA%K collapses relative to FedKAD while communication stays comparable, the expressiveness claim fails.","supporting_citations":[],"review_version":1}