{"id":"fc13f085-970f-435e-9a9e-7d81c90291c5","arxiv_id":"2505.09704","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A one-time clustering of federated learning clients by label distribution reduces total energy consumption while keeping accuracy comparable to per-round active client selection.","lead":"This paper proposes two ways to pick which devices train a shared AI model in federated learning, grouping devices by their data labels before training starts. The authors report that these one-time groupings cut total energy use while keeping accuracy competitive with more expensive selection methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RepClust may not hold K=10 fixed: with G≠10, selecting one whole partition yields L/G clients, so the reported energy/accuracy advantage could come from sampling fewer clients per round rather than from clustering quality.","rationale":"The reader's verdict is CONDITIONAL, and my read supports that same verdict: the one-time clustering idea is plausible and the energy accounting is unusually detailed, but the experimental support is conditional on resolving participation-count consistency and evaluation-protocol issues. I do not agree that the static-label-distribution assumption is the single most load-bearing weakness, because the paper explicitly states that assumption in footnote 5 and frames the contribution for already-acquired, non-drifting data. A more direct threat is that the RepClust results, as described, may violate the paper's own K=10 setting whenever G is not equal to 10, which would invalidate the energy-budget comparisons in Table II and the headline claim that RepClust is best. This is checkable from the public code, so it is not a reason to reject outright; it is a reason to keep the verdict conditional and require the check or a rerun. The post hoc selection of G on the reported energy metric is a separate methodological concern that reinforces the need for a validation-based or fixed G protocol. My agreement_with_reader is 'partial' because the reader identified the same conditional verdict but a different weakest assumption; the reader's rationale did mention post hoc G selection, which is related to my secondary point, yet the K/G mismatch is a more specific and more consequential flaw.","tokens_in":22496,"tokens_out":5099,"duration_ms":61228,"concrete_test":"Inspect the public repository (github.com/robertomatheuspp/clustering_ee_fl) and count the actual number of clients participating in a RepClust round for each G in {2,5,10,20,25,50}. Then rerun the CIFAR-10 Table II comparison with RepClust forced to exactly K=10 participants per round (for example, by restricting RepClust to G=10 or by specifying a deterministic K-client draw from the chosen partition) and with G selected before seeing test results (for example, by using a validation split or a fixed G=10). If RepClust still matches or beats FedCor and Random within one standard deviation, the concern is resolved; if the advantage shrinks or disappears, the central claim is not supported by the reported experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that one-time label-distribution clustering gives high accuracy at low energy, with RepClust best among data-sharing-free methods. The most load-bearing condition is that all methods are compared under the same per-round participation K=10, and that the reported energy savings come from clustering quality rather than from a different number of active clients. Section III-B describes RepClust as selecting one whole partition per round, and constraint (3) forces every cluster to have size |C_g| = L/G. With L=100 clients, G=20 (the value reported best for RepClust in Fig. 6b) gives clusters of size 5, while G=2 gives clusters of size 50. Section V-A states that K=10 clients are selected in all client selection mechanisms, but the paper never specifies how exactly 10 clients are drawn when G != 10 under RepClust. This is not a cosmetic discrepancy: local training and communication energy both scale with the number of participating clients, so a method that actually trains 5 clients per round has a structural energy advantage over baselines training 10 clients, independent of clustering quality. Table II compares methods under energy budgets defined relative to random sampling with K=10; if RepClust uses fewer clients, its higher sustained accuracy at 60-100% energy may simply reflect lower per-round energy rather than better selection. The concern is compounded by Section V-A's rule that G is chosen as the value minimizing total energy on the reported metric, which tunes the method to the evaluation criterion. A related notation conflict appears in Section III-C, where K denotes total clients, while Section V-A uses K=10 for selected clients, suggesting a possible source of the ambiguity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes two one-time clustering-based client selection schemes for federated learning in AIoT settings: SimClust, which groups clients with similar label distributions and samples across groups, and RepClust, which forms equal-size clusters of maximally diverse clients and selects one entire cluster per round. The authors build an energy model covering pre-processing, local training, and communication, and compare their methods against random sampling, FedCor, PowerD, DELTA, ClustLowVar, and FLIS on F-MNIST, CIFAR-10, and CIFAR-100. The central claim is that one-time clustering achieves accuracy comparable to or better than active selection while consuming less total energy.","tokens_in":22759,"tokens_out":5625,"duration_ms":57147,"significance":"If the results hold, the paper would make a useful contribution: a single pre-training clustering step could replace per-round active client selection and reduce total energy without sacrificing accuracy. The strengths of the paper are its careful energy decomposition, the inclusion of several recent baselines, multi-seed experiments, code availability, and a differential-privacy appendix. However, the main quantitative conclusions currently rest on two unresolved points: the number of clients actually trained per round under RepClust is not fixed to K=10, and the number of clusters G is selected per scenario on the evaluation metric. These issues must be resolved before the energy-efficiency claim can be accepted.","major_comments":[{"comment":"In Section III-B, RepClust is described as selecting one whole partition per round, and constraint (3) imposes |C_g| = L/G for all clusters. With L=100 clients and G=20, the value reported as best for RepClust in Fig. 6(b), this yields 5 participating clients per round, not the K=10 stated in Section V-A and claimed as constant K=10 in the Fig. 6 caption. The manuscript does not describe how exactly 10 clients are selected for RepClust when G differs from 10. Because local training and communication energy both scale with the number of participating clients, the reported energy advantage of RepClust could be a structural artifact of training fewer clients per round rather than evidence of better selection. Please specify the sampling rule for every G and ensure the number of active clients is identical across all compared methods, or analyze per-client energy separately.","section":"Section III-B, Section V-A, Section V-D"},{"comment":"For the proposed clustering solutions, G is chosen as the value in {2,5,10,20,25,50} that minimizes the overall energy costs for each scenario. This is equivalent to tuning the free parameter on the test scenario using the evaluation metric, so the reported energy savings and accuracy results are partly an artifact of this test-set selection. The experiments should use a validation-based or fixed G across scenarios, or the results should be reported as a sensitivity analysis rather than as the expected performance of the method.","section":"Section V-A and Tables I-II"},{"comment":"The statement that, among methods not sharing local data, RepClust achieves the highest accuracy across all datasets is not borne out by Table II for the CIFAR-10 final accuracy column: SimClust reaches 42.79% versus 41.80% for RepClust. If the claim is intended to apply only under the 60%, 80%, and 100% energy budgets, it should be stated explicitly; as written, it overstates the result.","section":"Section V-E and Table II"}],"minor_comments":[{"comment":"The text says 'we perform 10 rounds of local training' but elsewhere refers to epochs; this should be corrected to '10 local epochs' for consistency.","section":"Section V-A"},{"comment":"The complexity analysis uses K for the total number of clients and G for the number of groups, but K was defined earlier as the per-round number of selected clients; either reuse L for the total number of clients or define the notation locally.","section":"Section III-C"},{"comment":"The communication-energy equation has mismatched parentheses and the sentence describing P_down and P_up appears to swap the roles of uplink and downlink; please check the notation against the formulas in Refs. [4] and [48].","section":"Section IV-B"},{"comment":"The Gaussian noise added to label distributions is not calibrated to sensitivity or to a stated privacy budget, so the formal differential-privacy guarantee is not established; either provide the privacy analysis or soften the wording to an empirical robustness study.","section":"Appendix C"},{"comment":"Reference [28] contains an odd date field 'oct 5555' that should be corrected to the actual publication date.","section":"References"},{"comment":"The caption states 'constant K=10' but, given the RepClust selection rule and constraint (3), it is not clear how K=10 is realized for G values other than 10; this relates to the first major comment and should be clarified in the caption or the experiment setup.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of an AIoT/sustainable-FL venue, and the empirical comparison is more thorough than many client-selection papers. The reader's concern about RepClust training fewer than K clients per round is well-founded and should be resolved before publication. The test-set tuning of G is also a serious methodological issue that needs a validation protocol. I see no problem with the authors citing their own prior work [29]; it is directly relevant. The public code release and multi-seed reporting are positive features that should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious empirical study, and the energy accounting is more careful than most client-selection papers. But the main comparison is compromised by an ambiguity in how RepClust selects clients when the number of clusters is not 10, and by choosing G on the reported metric. I'd send it to referees, but the RepClust protocol needs to be fixed before the energy claims can be trusted.\n\nWhat's new: the paper reframes client selection as a one-time pre-training clustering step rather than per-round active selection, and it measures the full energy footprint—local training, communication, server-side preprocessing—with CodeCarbon. The comparison includes six baselines, actual hardware energy, and a serviceable communication model. That is more than most FL selection papers do. The finding that clustering-informed selection can match FedCor's accuracy while avoiding the GP training cost is interesting, and if it holds, it is useful for AIoT designers.\n\nThe soft spots are real, and one is load-bearing. RepClust is described as selecting one whole partition per round, and constraint (3) forces every partition to have L/G clients. With L=100 and G=20—the value reported best in Fig. 6—that is 5 clients per round, not the K=10 used by every baseline and by SimClust. The paper never explains how the K=10 fixed client count is achieved under RepClust. If fewer clients are actually trained, the energy savings are partly a structural artifact of doing less work per round, not a consequence of clustering quality. This is not a cosmetic gap.\n\nSecond, G is selected post hoc as the value that minimizes total energy on the test scenarios. Since energy is the evaluation metric, the method is tuned to the outcome. The search cost over G is not included in the reported energy, and for SimClust the paper even says the optimal G varies with the seed.\n\nThird, some accuracy claims are stronger than the tables support. The paper says RepClust achieves the highest accuracy across all datasets among methods that do not share data, but in Table II FedCor beats RepClust on CIFAR-10 (42.57 vs 41.80). And \"accelerates convergence\" is not uniformly true: Table I shows FedCor reaching target accuracies in fewer rounds for the (1,1)-Dir scenarios.\n\nThe static-label-distribution assumption is stated in footnote 5, so at least it is explicit.\n\nWho this is for: researchers working on energy-efficient FL client selection. They should read it for the energy measurement methodology and the one-time clustering idea, and should be careful with the RepClust numbers. I would send it to peer review with a request for a clear protocol description and an honest treatment of G. As it stands, I would not cite the energy numbers without verification.","headline":"A serious energy-measurement study of one-time client clustering for FL, but RepClust's per-round client count is unclear and G is tuned on the metric.","tokens_in":23345,"tokens_out":6476,"would_cite":false,"duration_ms":60345,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that one-time clustering of clients by label distribution makes federated learning reach target accuracy with lower total energy than per-round adaptive selection.","keywords":["federated learning","client selection","clustering","energy efficiency","label distribution","AIoT","non-IID data","repulsive clustering"],"falsifier":"Run the same protocol on a dataset whose local label distributions change every 50 rounds, or compute the initial clustering from deliberately corrupted label estimates, and compare energy-to-accuracy against a per-round adaptive selector such as FedCor; if the clustering methods lose their advantage whenever the initial distributions become stale, the one-time-clustering claim is falsified.","tokens_in":22285,"feed_emoji":"⚡","tokens_out":4989,"duration_ms":51814,"temperature":0.7,"pith_summary":"This paper proposes replacing the per-round active client selection of federated learning with a single clustering step performed before training starts. Clients are grouped by their label distributions so that each round's participants jointly represent the whole data distribution. The authors argue that local training, not communication, dominates the energy budget of FL on AIoT devices, so selecting representative clients that converge faster saves more energy than merely reducing communication rounds. Their two clustering methods, SimClust and RepClust, reach comparable or higher accuracy than recent selection approaches while spending far less on pre-processing, with RepClust the best among methods that do not share local data.","feed_headline":"One-time client clustering cuts federated learning energy","feed_subtitle":"Grouping devices by label distribution once, before training, matches accuracy of adaptive selection at lower energy.","key_machinery":"The machinery is a one-time clustering of the label-distribution vectors of all clients, computed before federated training and then held fixed. SimClust clusters similar label distributions so that stratified sampling from clusters yields diversity across the data space; RepClust solves a multi-objective dispersion problem, maximizing intra-cluster pairwise distance while minimizing inter-cluster distance with equal cluster sizes, using a heuristic swap algorithm. This transforms client selection from an online optimization repeated every round into a fixed partition, moving the cost into a negligible pre-processing step while ensuring each round's sampled clients collectively cover the label space.","core_discovery":"The central discovery is that the most energy-consuming part of federated learning on constrained AIoT devices is local training, and that a one-time, pre-training clustering of clients by label distribution can make every communication round's selected clients jointly approximate the global label distribution. SimClust groups similar clients, using k-means with symmetrized KL divergence, and draws one client per group per round; RepClust forms equal-sized groups that are internally diverse and mutually similar, then engages a whole group per round. The paper reports that these methods need fewer rounds to reach target accuracy in heterogeneous settings than random sampling or FedCor's Gaussian-process selection, while spending almost nothing on pre-processing, so total energy falls. Among baselines that do not share local data, RepClust achieves the highest accuracy on F-MNIST, CIFAR-10, and CIFAR-100 under equal energy budgets.","pith_inferences":["If local training dominates the energy budget, the relative benefit of one-time clustering should grow as models get shallower or local epochs increase; on deeper models the communication fraction rises, so clustering's advantage may narrow.","The fixed-cluster assumption suggests a testable extension: re-clustering periodically or when drift is detected could extend the method to non-stationary AIoT data at a small additional pre-processing cost.","A practical deployment could have each device send a differentially private label histogram before clustering; the paper's Appendix C results imply RepClust would withstand more noise than SimClust before cluster structure is lost."],"forward_implications":["In heterogeneous settings with location-dependent partitions, the clustering methods reach target accuracy in the same or fewer rounds than FedCor, making them the lowest-energy option.","Because clustering is executed only once, pre-processing energy is negligible compared with active selection methods that train a Gaussian process or rank clients every round.","RepClust's optimal number of clusters stays relatively stable across random seeds, so its energy profile is less sensitive to the choice of cluster count than SimClust's.","Under differential privacy noise added to label distributions, RepClust retains a non-random clustering solution even when SimClust degrades toward random assignment, suggesting the approach can tolerate privacy masking.","The authors note the clustering pre-processing is compatible with further efficiency techniques such as over-the-air computing, quantization, and pruning, which could reduce server-side energy on top of the sampling savings."],"supporting_citations":[{"why":"Supplies the FedCor baseline, a Gaussian-process active client selector whose per-round GP training drives high pre-processing energy.","marker":"[16]"},{"why":"Supplies the ClustLowVar baseline, a clustered sampling method compared in energy and accuracy in Table II.","marker":"[18]"},{"why":"Supplies the DELTA baseline, a gradient-diversity client sampling method whose performance is compared in Table II.","marker":"[17]"},{"why":"Supplies the FLIS baseline, which clusters by inference similarity and shares local data, making it the strongest CIFAR-10 competitor.","marker":"[21]"},{"why":"Supplies the PowerD baseline, which ranks clients by local loss and is compared in Table II.","marker":"[14]"},{"why":"Supplies the stratified-sampling principle underlying SimClust's one-client-per-cluster selection.","marker":"[38]"},{"why":"Supplies the repulsive-clustering heuristic that RepClust adapts for its swap-based group construction.","marker":"[39]"},{"why":"Supplies the IEEE 802.11ax communication energy model used to estimate uplink and downlink energy in the comparisons.","marker":"[4]"}],"fun_headline_variants":["Cluster once, save energy in federated learning","Pre-training clustering cuts FL energy costs","Label-based clusters make FL lean and fast","Static clustering beats adaptive selection on energy","Group clients by labels to cut FL energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The clusters are computed once from label distributions known before training, and the paper assumes those distributions do not change over the 500 training rounds; if the data drifts or the estimates are wrong, the sampled clients are no longer representative and the energy and accuracy results would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Cluster once, save energy in federated learning","Pre-training clustering cuts FL energy costs","Label-based clusters make FL lean and fast","Static clustering beats adaptive selection on energy","Group clients by labels to cut FL energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1312,"prompt_tokens":874,"completion_tokens":438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":373}},"tokens_in":490,"tokens_out":438,"duration_ms":4866,"temperature":1.0,"reasoning_tokens":373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:26:42.686896+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same protocol on a dataset whose local label distributions change every 50 rounds, or compute the initial clustering from deliberately corrupted label estimates, and compare energy-to-accuracy against a per-round adaptive selector such as FedCor; if the clustering methods lose their advantage whenever the initial distributions become stale, the one-time-clustering claim is falsified.","supporting_citations":[{"cited_title":"FedCor: Correlation-Based Active Client Selection Strategy for Het- erogeneous Federated Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedCor baseline, a Gaussian-process active client selector whose per-round GP training drives high pre-processing energy."},{"cited_title":"Clustered Sampling: Low-Variance and Improved Representativity for Clients Selection in Federated Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the ClustLowVar baseline, a clustered sampling method compared in energy and accuracy in Table II."},{"cited_title":"Delta: Diverse client sam- pling for fasting federated learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the DELTA baseline, a gradient-diversity client sampling method whose performance is compared in Table II."},{"cited_title":"FLIS: Clustered Fed- erated Learning Via Inference Similarity for Non-IID Data Distribution ,","cited_arxiv_id":null,"evidence_quote":"Supplies the FLIS baseline, which clusters by inference similarity and shares local data, making it the strongest CIFAR-10 competitor."},{"cited_title":"Neyman, On the Two Different Aspects of the Representative Method: the Method of Stratified Sampling and the Method of Purposive Selec- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the stratified-sampling principle underlying SimClust's one-client-per-cluster selection."},{"cited_title":"Repulsive clustering based pilot assignment for cell-free massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the repulsive-clustering heuristic that RepClust adapts for its swap-based group construction."},{"cited_title":"The Cost of Training Machine Learning Models over Distributed Data Sources,","cited_arxiv_id":null,"evidence_quote":"Supplies the IEEE 802.11ax communication energy model used to estimate uplink and downlink energy in the comparisons."}],"review_version":1}