{"id":"129ede75-8672-49cf-bb20-39de635b9dfc","arxiv_id":"2507.14999","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"FedClusAvg is claimed to improve FDIA detection accuracy and communication efficiency under non-IID smart grid data, but the weighting formula contradicts the stated mechanism and the evidence is not reproducible.","lead":"This paper proposes FedClusAvg, a clustered federated learning method for detecting false data injection attacks in smart grids, plus a hierarchical variant called FedClusAvg+. The authors report higher accuracy than FedAvg, FedProx, and FedNova on bus and grid benchmarks, but the supporting equations and experiments have serious inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The aggregation rule is specified inconsistently: Eq. 14 weights clients by deviation from the mean, while the text says large deviations should reduce weight; reported gains depend on an unspecified implementation choice.","rationale":"The reader's weakest-assumption field focuses on the realism of the FDIA-labeled grid datasets and the lack of attack-generation details. That is a serious external-validity concern, but the more immediate and load-bearing problem is internal: the algorithm's own definition contradicts its design rationale. Eq. (14) and Table II assign aggregation weights proportional to parameter deviation from the mean, while Section II.A explicitly says large deviations should reduce a client's weight. If the implemented system followed Eq. 14, the method is not the one described and the claimed robustness under Non-IID data is not explained; if it followed the prose, then the displayed equations and pseudocode are incorrect. Both possibilities mean the experimental results cannot be reliably interpreted without access to code or a corrected specification. A minimal two-variant re-implementation would settle which behavior was used and whether the choice matters. I therefore support the reader's REJECT verdict, though for a slightly different primary reason; hence 'partial' agreement. I do not see a need to adjust the verdict, since the reader already reached REJECT and the internal inconsistency strengthens that conclusion.","tokens_in":14751,"tokens_out":3683,"duration_ms":41494,"concrete_test":"Re-run the Section III.B/C experiments on the IEEE 118-bus and 300-bus datasets with two variants implemented from the same text: (A) weights proportional to ||˜w − w_k|| as in Eq. 14/Table II, and (B) weights inversely proportional to that deviation as the Section II.A prose describes. Compare the FedClusAvg/FedClusAvg+ rows of Tables VII and X. If the two variants differ by more than the reported margins over FedAvg/FedProx, the published specification does not determine the claimed results; if they are nearly identical, the inconsistency is harmless and can be fixed by a clarifying sentence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II.A defines the server-side aggregation twice in incompatible ways. The prose states that 'When the degree of deviation is large, the corresponding drop reduces its weight, and when the degree of deviation is small, the weight increases.' The formula immediately below, Eq. (14), sets η = ||˜w − w_m|| / Σ_m ||˜w − w_m||, i.e., weights are proportional to deviation from the weighted mean ˜w. Table II's server pseudocode uses the same proportional assignment. Thus a client with a large parameter deviation receives a large aggregation weight, the opposite of the stated reliability weighting. Because this aggregation rule is the core mechanism claimed to reduce client drift and to produce the accuracy and communication gains, the algorithm is not well-defined: a reader cannot tell whether Tables VI–VIII and X–XIII used the proportional rule (Eq. 14), the inverse rule described in words, or some third variant. The central experimental claim therefore rests on an unspecified implementation choice. This is a correctness and reproducibility problem, not merely a wording issue, since proportional and inverse weighting can produce materially different global models under the heterogeneous data distributions the paper targets. No code is released, so the ambiguity cannot be resolved post hoc.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedClusAvg and FedClusAvg+, federated learning variants for false data injection attack (FDIA) detection in smart grids. The methods combine client-side sample clustering into sub-clients, deviation-based weighted aggregation, and, in FedClusAvg+, a three-tier client–sub-server–server communication hierarchy. The authors claim improved detection accuracy, robustness under Non-IID data, and reduced communication cost compared with FedAvg, FedProx, and FedNova, based on experiments on IEEE 118-bus, IEEE 300-bus, and two in-house regional/provincial grid datasets. The central technical mechanism is a weighted aggregation rule intended to down-weight clients whose local parameters deviate strongly from the global average.","tokens_in":15030,"tokens_out":4169,"duration_ms":49916,"significance":"If the claims were fully supported, the paper would address a practically important problem: privacy-preserving FDIA detection under heterogeneous smart-grid data with communication constraints. The hierarchical architecture and client-side clustering are reasonable ideas, and the comparison against FedAvg, FedProx, and FedNova across multiple system scales is a useful evaluation design. However, the current manuscript has load-bearing ambiguities and reproducibility gaps that prevent the experimental claims from being interpreted: the aggregation formula contradicts the verbal description, the FDIA-labeled datasets and attack-generation protocol are not described, and no uncertainty quantification is provided. The paper therefore cannot currently be accepted as evidence for the claimed improvements, despite the plausibility of the underlying approach.","major_comments":[{"comment":"The aggregation rule is defined inconsistently. The text states that a large parameter deviation should reduce the client's weight and a small deviation should increase it, but Eq. (14) sets η = ||w̃ − w_m|| / Σ_m ||w̃ − w_m||, which makes the aggregation weight proportional to the deviation. Table II's server pseudocode uses the same proportional rule, and Eq. (11) does the same at the client level. Since proportional and inverse weighting can produce materially different global models under Non-IID data, the algorithm as written is not well-defined. The authors must state which rule was actually used in Tables VI-VIII and X-XIII, correct the equations or the prose, and rerun the experiments with a single unambiguous rule.","section":"Section II.A, Eqs. (13)-(15) and Table II"},{"comment":"The data-generation process is not described. Reference [30] is MATPOWER, which does not provide FDIA labels, yet the paper reports accuracy and AUC on 'IEEE 118-Bus dataset,' 'IEEE 300-Bus,' 'regional power grid,' and 'provincial power grid' without explaining how attack samples were synthesized, what attack models were used, or how the in-house regional/provincial grids were constructed. The claim that FedClusAvg+ 'achieves an average Accuracy of 94.63% and AUC of 0.9425' across these systems is therefore not reproducible, and the generalization analysis cannot be evaluated. A complete description of the attack-injection procedure and dataset construction is required.","section":"Section III.A and Section III.C, Tables IX-XIII"},{"comment":"No statistical significance or variability information is reported. The tables list minimum, first quartile, median, mean, and maximum, but it is unclear whether these statistics are computed across clients, across training rounds, or across independent runs. There are no error bars, standard deviations, or confidence intervals, and no indication of how many random seeds were used. The reported performance gaps of 1-3 percentage points in accuracy may be within run-to-run variance. The experiments should be repeated with multiple seeds and the variability reported.","section":"Tables VI-VIII and Section III.B"},{"comment":"The local detection model is only identified as 'the Rec-AD model we proposed in reference [31],' a self-cited preprint. No architecture, loss function, or hyperparameter details are given in this paper, and no public implementation is provided. Because all experimental results depend on this model, the specification is incomplete. The model should either be described in sufficient detail or the code should be released so that the experiments can be reproduced.","section":"Section III.A and reference [31]"}],"minor_comments":[{"comment":"The standardization formula x*_ki = (x_ki − x_ki)/√s_ki is dimensionally inconsistent: it appears to use the mean and variance of a single sample rather than feature-wise statistics over the dataset. Please rewrite the formula to standardize each feature across samples.","section":"Section II.A, Eq. (7)"},{"comment":"The distance D12 = √((Z1−Z2)^2) is not a proper vector norm; for vectors it should be written as ||Z1−Z2||, with the norm defined appropriately.","section":"Section II.A, Eq. (9)"},{"comment":"The pseudocode is labeled 'SpectralClust,' but the described procedure is a longest-distance or complete-linkage-style clustering heuristic, not spectral clustering. The naming is misleading and should be corrected.","section":"Table I and Section II.A"},{"comment":"The method is called 'Federated Cluster Average' in the abstract but 'Federated Clustered Averaging' in the main text. Please use one consistent name throughout.","section":"Abstract and Section II"},{"comment":"The metric name 'Precison' is misspelled; it should be 'Precision.' Also, Figures 4-6 are captioned as 'Model KS values' but appear to show ROC curves; the captions should match the content.","section":"Table VI"},{"comment":"The phrase 'the samples of each client are encrypted and aligned' suggests secure computation, but no encryption or alignment protocol is specified anywhere in the paper. The privacy claim should be stated precisely, or this sentence should be revised.","section":"Section I, Fig. 1 and Section II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a companion to reference [31] and has the feel of a work in progress. The most serious problem is that the core aggregation rule is defined contradictorily, and the FDIA datasets used for the generalization tables are generated in-house without any description. If the authors can clarify the rule, release the data-generation code, provide multi-seed results, and specify the local model, the paper might become publishable; as it stands, the experimental evidence cannot be trusted. I would lean toward rejection if these issues are not fully addressed in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: the FedClusAvg idea—cluster each client's data into sub-clients, weight local updates by deviation, and add a sub-server tier—is a reasonable combination of known parts, and the reported gains over FedAvg, FedProx, and FedNova are consistent across Tables VI–VIII and X–XIII. That consistency earns some credit. The communication latency numbers in Fig. 6 are also concrete.\n\nThe problem is the core aggregation rule. Section II.A says large deviations should reduce a client's weight and small deviations should increase it. Eq. 14 (and the pseudocode in Table II) does the opposite: η is proportional to ||˜w − w_m||, so a client farther from the weighted mean gets a larger weight. Same issue in the local sub-client aggregation (Eqs. 10–12). This is not a wording nit—it means the algorithm is not well-defined, and a reader cannot tell whether the tables used the proportional rule, the inverse rule, or something else. No code is released, so the ambiguity can't be resolved. Also, the local detection model is a self-cited prior preprint (ref. 31) that isn't described, which adds another layer of non-reproducibility.\n\nAlso worth noting: the 'SpectralClust' in Table I is a farthest-point heuristic with a distance threshold, not spectral clustering. The generalization experiments use in-house generated FDIA-labeled versions of the IEEE 118/300 and regional/provincial grids, but the attack generation process is never described, so the labels' realism is untestable. And the abstract's claim about reducing communication rounds is not actually shown; Fig. 6 only compares per-round latency between FedClusAvg and FedClusAvg+, not rounds-to-convergence.\n\nI'd tell an editor to desk reject for now: the central mechanism is internally contradictory, and the empirical claims depend on an implementation choice the paper doesn't specify. If the authors clarify the aggregation rule (or fix the prose) and release code and attack-generation details, a resubmission would be worth a proper review. As it stands, I wouldn't want to referee it.","headline":"The FedClusAvg pipeline is a sensible combination of known ideas and the tables are consistent, but Eq. 14 contradicts the prose on the core aggregation rule, leaving the central mechanism undefined and the empirical claims unreproducible.","tokens_in":15499,"tokens_out":3271,"would_cite":false,"duration_ms":31850,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that clustered, hierarchical federated averaging gives 94.63% average FDIA detection accuracy on non-IID smart grid data while cutting communication.","keywords":["false data injection attack","federated learning","non-IID data","smart grid security","clustered aggregation","hierarchical communication","FedClusAvg"],"falsifier":"Retrain FedClusAvg+ and FedAvg+ on the IEEE 300-bus system using FDIA labels produced by a documented attack generator, holding the same label-skewing protocol; if the accuracy gap over FedAvg+ is no longer around 2.5 points on average, the reported generalization advantage rests on the paper's undisclosed label generation rather than on the clustering and hierarchy.","tokens_in":14578,"feed_emoji":"⚡","tokens_out":7926,"duration_ms":75622,"temperature":0.7,"pith_summary":"Smart grid measurement data is non-IID across regions, which degrades standard federated FDIA detectors. The paper's claim is that a clustered, hierarchical federated design fixes this: FedClusAvg partitions each client's local samples into sub-clients, trains them in parallel, and merges them with deviation-based weights; FedClusAvg+ inserts sub-servers between clients and the central server. The authors report that FedClusAvg+ reaches 94.63% average accuracy and 0.9425 AUC across IEEE 118-bus, IEEE 300-bus, regional, and provincial grid benchmarks, beating FedAvg, FedProx, and FedNova on every metric while consuming fewer communication rounds and less bandwidth. A sympathetic reader should care because the design addresses three practical obstacles at once: privacy, non-IID heterogeneity, and communication cost.","feed_headline":"Clustered federated learning lifts FDIA detection to 94.63%","feed_subtitle":"Hierarchical clustering and sub-servers beat FedAvg, FedProx, FedNova on non-IID grid data with fewer rounds.","key_machinery":"The load-bearing mechanism is client-side sample clustering followed by deviation-weighted aggregation at two levels. Each client with more than 300 samples is partitioned into sub-clients by a greedy farthest-point clustering routine (labeled SpectralClust in the pseudocode): pick a threshold, choose the farthest sample from the first center as the second center, keep adding centers while any sample's minimum distance to existing centers exceeds the threshold, then assign every sample to its nearest center. Each sub-client trains a copy of the model in parallel, and the client merges the sub-models with weights inversely proportional to the distance of each sub-model's gradient from the client's average gradient. The server applies the same deviation-weighting rule to the client models, and in FedClusAvg+ an intermediate sub-server tier performs this aggregation for a subset of clients before forwarding results to the central server; that extra tier is what the paper credits for reduced communication rounds and bandwidth.","core_discovery":"The central discovery is that local model divergence under non-IID data can be reduced by clustering samples inside each client before local training, and then weighting the aggregation of both sub-clients and clients by how far their updates deviate from the average. On the paper's four benchmarks, FedClusAvg+ attains an average accuracy of 94.63% and AUC of 0.9425, outperforming FedAvg+, FedProx, and FedNova on every reported metric; the accuracy gap over FedAvg+ grows from roughly 1.8 points on the IEEE 118-bus system to 2.5 points on the provincial grid, and reaches 8.5 to 9.9 points under very high client heterogeneity. The authors attribute this to the combination of hierarchical communication and clustered aggregation, which mitigates model drift, regional misclassification, and communication latency.","pith_inferences":["A next step the paper leaves implicit is to compare total bytes exchanged and rounds-to-target-accuracy, not just per-round latency, since that is what determines operational bandwidth savings.","Because the deviation-weighting rule is a form of robust aggregation, the same mechanism could be tested against Byzantine clients that send poisoned updates; large-deviation weighting may either filter them out or, if an adversary controls many clients, amplify harm.","The reported benchmarks use synthetic attacks generated without a stated procedure; a public FDIA benchmark with documented attack construction would clarify whether the clustering advantage transfers to coordinated adversaries that craft stealthy vectors."],"forward_implications":["If the reported margins hold, moving from a flat federated architecture to a client-subserver-server one should give the largest accuracy gains on the largest grids, since the gap over FedAvg+ grows with system scale.","At 1:9 attack:normal class imbalance, the method is claimed to lift precision by 8.2 points, recall by 6.3 points, and F1 by 7.3 points over FedAvg+, meaning rare attacks are caught without a proportional jump in false alarms.","The hierarchical tier cuts per-round communication delay by roughly 15.6% under WiFi and by 15% under 5G, so the design is meant to run on bandwidth-limited substation links.","Since only model parameters are shared, the scheme is positioned as compatible with data-sharing restrictions between regional operators, which centralized FDIA training cannot satisfy.","Under very high heterogeneity, accuracy stays above 89.7% on the 300-bus system and 85.2% on the provincial grid, indicating the method is designed for the regime where standard FedAvg collapses."],"supporting_citations":[{"why":"Supplies the FedAvg algorithm that FedClusAvg modifies and is directly compared against in the experiments.","marker":"[29]"},{"why":"Provides the IEEE 118-bus system dataset used for training and testing the FDIA detection models.","marker":"[30]"},{"why":"Supplies the Rec-AD detection model that is integrated inside the federated framework as the classifier.","marker":"[31]"},{"why":"FedProx is a comparison baseline that adds a proximal term to limit client drift under heterogeneity.","marker":"[32]"},{"why":"FedNova is a comparison baseline that normalizes local updates to stabilize heterogeneous federated optimization.","marker":"[33]"}],"fun_headline_variants":["FedClusAvg lifts FDIA detection to 94.63% on non-IID grids","Clustered FL outperforms FedAvg+ on heterogeneous grid attacks","Smart grid FDIA detection improved via clustered federated learning","Non-IID FDIA detection enhanced by hierarchical clustered aggregation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic FDIA labels on the IEEE 118/300, regional, and provincial grids faithfully mimic real attacks, but the paper never describes how attack samples are generated, so if those labels are unrepresentative, every reported accuracy and AUC advantage is an artifact of the benchmark rather than the algorithm.","fun_headline_variants_meta":{"raw":{"variants":["FedClusAvg lifts FDIA detection to 94.63% on non-IID grids","Clustered FL outperforms FedAvg+ on heterogeneous grid attacks","Smart grid FDIA detection improved via clustered federated learning","Non-IID FDIA detection enhanced by hierarchical clustered aggregation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000343,"raw_usage":{"total_tokens":1884,"prompt_tokens":942,"completion_tokens":942,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":865}},"tokens_in":558,"tokens_out":942,"duration_ms":9198,"temperature":1.0,"reasoning_tokens":865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:43:41.258525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FedClusAvg+ and FedAvg+ on the IEEE 300-bus system using FDIA labels produced by a documented attack generator, holding the same label-skewing protocol; if the accuracy gap over FedAvg+ is no longer around 2.5 points on average, the reported generalization advantage rests on the paper's undisclosed label generation rather than on the clustering and hierarchy.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg algorithm that FedClusAvg modifies and is directly compared against in the experiments."},{"cited_title":"Matpower-a matlab power system simulation package: User","cited_arxiv_id":null,"evidence_quote":"Provides the IEEE 118-bus system dataset used for training and testing the FDIA detection models."},{"cited_title":"Rec-AD: An Efficient Computation Framework for FDIA Detection Based on Tensor Train Decomposition and Deep Learning Recommendation Model","cited_arxiv_id":"2507.14668","evidence_quote":"Supplies the Rec-AD detection model that is integrated inside the federated framework as the classifier."},{"cited_title":"Federated optimization in heterogeneous networks,","cited_arxiv_id":null,"evidence_quote":"FedProx is a comparison baseline that adds a proximal term to limit client drift under heterogeneity."}],"review_version":1}