{"id":"772cedb1-11a9-435c-ad64-875dc14cf545","arxiv_id":"2411.17833","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ACSP-FL combines below-average client selection, a decay schedule, and layer sharing with per-client personalization to reduce federated learning communication cost on HAR tasks.","lead":"ACSP-FL is a federated learning method that selects only underperforming clients each round, gradually reduces their number, and shares only part of the neural network to cut communication. In tests on three activity-recognition datasets it reports up to 99% fewer transmitted bytes and similar or better accuracy than standard baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy advantage is not yet established: ACSP-FL is evaluated with personalized per-client models while the baselines are evaluated as a single global model; a controlled personalization baseline is needed.","rationale":"The reader's weakest_assumption identifies exactly the same soft spot: ACSP-FL is compared with personalization while the baselines are not. This is the most load-bearing concern because the paper's headline claim couples communication reduction with 'good convergence' and 'at least 10%' accuracy gains. The communication savings are real and independently visible in TX-byte counts, but the convergence claim depends wholly on the accuracy metric being a fair comparison of like-for-like models. It is not, because ACSP-FL's distributed accuracy averages over personalized per-client composites and the baselines' accuracy uses a single global model. The proposed FedAvg+FT control would settle whether the accuracy advantage remains when baselines receive the same personalization step. I would keep the paper's verdict at CONDITIONAL: the method is promising and the communication result is plausible, but the central accuracy claim needs a controlled baseline before it can be accepted. Secondary issues such as single runs without error bars and inconsistent percentage claims (95%, 90%, 97-99%) reinforce the condition, but the asymmetric evaluation is the decisive methodological gap.","tokens_in":19600,"tokens_out":3092,"duration_ms":30871,"concrete_test":"Add a FedAvg+FT arm to the Section 4.5 comparison: run FedAvg exactly as configured, then let each client fine-tune the final global model on its local data and choose the better of local and global model via Equation 8, and recompute the distributed accuracy metric. If FedAvg+FT matches ACSP-FL DLD's 0.92/0.75/0.92 accuracy on the three datasets, the accuracy gap in Table 4 is attributable to personalization rather than to adaptive client selection or layer sharing, and the convergence claim must be restated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The fragile premise is the accuracy comparison in Table 4 and Figure 8. Section 4.3 defines 'distributed accuracy' as the average over per-client evaluations, and Algorithm 2's Evaluate function constructs each client's model as w_i = [w_g, w_l_i], the composite of shared and local layers. Section 3.4 Equation 8 additionally lets ACSP-FL clients choose whichever of local or global model performs better on their own data. Thus the ACSP-FL accuracy reported in Table 4 is not the accuracy of one collaboratively trained global model; it is the average over personalized per-client predictors. In contrast, FedAvg, POC, OORT, and DEEV are evaluated only as the single global model, with no fine-tuning or local-model selection. The Conclusion's claim that ACSP-FL 'increases the model's performance by at least 10%' is therefore not supported: the comparison conflates personalization with improved convergence of the shared model. The communication-reduction claim is directionally supported by the TX-byte numbers and is not the weak point; the load-bearing issue is that 'good convergence' is asserted on an asymmetric evaluation protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ACSP-FL combines adaptive client selection (based on client accuracy relative to the mean), a decay function that reduces the number of participating clients over rounds, and partial model sharing with per-client personalization. The authors evaluate it on three HAR datasets (UCI-HAR, MotionSense, ExtraSensory) against FedAvg, POC, OORT, and their own DEEV, claiming communication reductions up to 95-99%, faster convergence, and higher accuracy. The communication-reduction direction is plausible and supported by TX-byte measurements, but the accuracy and convergence claims rest on an asymmetric evaluation protocol in which ACSP-FL uses personalized per-client models while the baselines are evaluated as a single global model.","tokens_in":19812,"tokens_out":7056,"duration_ms":62250,"significance":"If the reported communication reductions hold under a fair evaluation, the contribution is practically useful: it directly attacks the two dominant costs in federated learning (round count and per-round bytes), and the paper ships open-source code, a container-based testbed, and measurements on three real datasets rather than only synthetic ones. The TX-byte accounting is a strength. However, the headline accuracy and convergence claims are currently not supported because the comparison confounds personalization with improved shared-model convergence; the significance is therefore conditional on the authors providing a controlled personalization baseline and on resolving the suspiciously high initial accuracy values in the learning curves.","major_comments":[{"comment":"The distributed-accuracy metric is computed differently for ACSP-FL and the baselines. For ACSP-FL, Algorithm 2 (Evaluate) constructs each client model as w_i = [w_g, w_l_i], and Eq. (8) lets the client select whichever of the local composite model and the global model has lower loss on its own data; for FedAvg, POC, OORT, and DEEV, the single global model is evaluated without fine-tuning or local-model selection. Consequently, the higher accuracies in Table 4 and Figure 8 confound personalization with better convergence of the shared model. Please add a controlled personalization baseline (e.g., per-client fine-tuning of the global model or the same local/global selection rule for all methods) and also report the accuracy of the pure shared global model for ACSP-FL.","section":"Section 4.3 and Table 4"},{"comment":"The accuracy curves start at implausibly high values at round 0. For example, on ExtraSensory the first plotted point shows FedAvg around 0.84 and ACSP-FL DLD around 0.92, although Algorithm 1 initializes the model randomly (RANDOMINIT) and a randomly initialized MLP should give near-chance accuracy on a multiclass HAR task. Please clarify whether the x-axis starts at the initial random model, and state the local training hyperparameters (number of epochs, learning rate, batch size, optimizer settings). If the models are warm-started or pre-trained, this must be disclosed, since it changes the interpretation of both the convergence and communication-savings claims.","section":"Figures 6 and 8; Section 4.1"},{"comment":"All reported results are single runs, and both the ACSP-FL hyperparameters (decay = 0.005 in Section 4.5; the PMS mapping in Eq. (9)) and the baseline hyperparameters (k = 50% for OORT and POC) were chosen via exploratory analysis without stating whether the tuning was done on test data. Without multiple seeds, error bars, or a sensitivity analysis for decay and the layer-sharing threshold, the magnitude of the claimed reductions (e.g., 97-99% fewer TX bytes versus FedAvg) cannot be assessed for robustness. Please report means and standard deviations over at least three runs and describe how the exploratory hyperparameter selection is separated from the reported test results.","section":"Tables 3-4 and Section 4.5"}],"minor_comments":[{"comment":"The reported savings are inconsistent: the abstract and conclusion claim up to 95% communication reduction, while Table 4 shows 97-99% TX-byte reductions versus FedAvg; similarly, Section 4.5 says accuracy increases up to 8%, while the conclusion says at least 10%. Please reconcile these numbers.","section":"Abstract, Section 4.5, Conclusion"},{"comment":"The dynamic layer-sharing function states that '4 is the total number of layers of the model' but the MLP is described as having three hidden layers; clarify which layers are counted (input, hidden, output) and explain the behavior of the formula at the accuracy = 0.25 threshold, where ceil(1/0.25) = 4 yields no change in shared layers.","section":"Equation (9) and Section 4.2"},{"comment":"The x-axis of Figure 5 is labeled 'Client Accuracy (%)' but the plotted values range from 0.0 to 1.0; use fraction or percent consistently to avoid confusion.","section":"Figure 5"},{"comment":"The comment in Algorithm 2 says 'combines the local model with the local one'; this should read 'combines the global piece with the local piece' to match the w_i = [w_g, w_l_i] construction.","section":"Algorithm 2, line 2"},{"comment":"The same paper by Vaizman et al. appears as both reference [33] and reference [34]; the duplicate should be merged.","section":"References [33] and [34]"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the authors' own DEEV work [9], adding personalization and layer sharing. The journal should weigh whether the novelty is sufficient for a full-length article. The main technical risk is the evaluation protocol: if a personalization-controlled baseline closes the accuracy gap, the remaining contribution would be primarily the communication savings, which are already directionally supported. The implausibly high round-0 accuracies also require an explicit explanation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, short version first: ACSP-FL is a sensible engineering combination of adaptive client selection (their own DEEV), a decay schedule, partial layer sharing, and per-client personalization. The reported TX-byte reductions (97–99% vs FedAvg) are directionally credible, the code is public, and the experimental setup is reproducible. But the headline accuracy claim does not hold as stated: ACSP-FL clients get to pick between their localized composite model and the global model (Eq. 8, Algorithm 2), while FedAvg, POC, OORT, and DEEV are evaluated only as a single global model. So Table 4's 'distributed accuracy' for ACSP-FL is not the accuracy of the collaboratively trained global model; it's the average over personalized per-client predictors. The conclusion's 'at least 10%' improvement is not supported by that comparison. The stress-test note is right on this.\n\nWhat the paper does well: it is transparent about extending DEEV; the dynamic layer-count rule (Eq. 9) is ad hoc but clearly stated; the docker/Flower evaluation infrastructure is a real asset; and the communication-reduction mechanism—fewer clients over rounds plus fewer layers per message—is mechanically sound. The authors also correctly identify the fixed-k limitation of POC and OORT.\n\nSoft spots, in proportion: the missing controlled baseline is load-bearing. Adding per-client fine-tuning to FedAvg (or comparing global-model-only accuracy for ACSP-FL) would settle whether the accuracy gain comes from personalization or from better collaborative training. The paper also has no error bars or repeated runs, baseline hyperparameters were chosen via exploratory analysis on the test data, and the reported numbers drift: abstract says 95% communication reduction, introduction says 90%, Table 4 shows 99%, and the conclusion claims 'at least 10%' accuracy improvement. These are fixable but need addressing.\n\nWho is this for: anyone building FL systems on heterogeneous mobile/edge clients who cares about communication cost. The communication-efficiency story is worth a serious look; the accuracy story needs another experimental pass.\n\nRecommendation: send to peer review, but the reviewers should demand a personalized baseline comparison and consistency in reported numbers. If the accuracy advantage survives that, this becomes a useful systems paper.","headline":"The communication savings are probably real; the accuracy advantage is not yet supported by the asymmetric evaluation.","tokens_in":20401,"tokens_out":2017,"would_cite":false,"duration_ms":19338,"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":"ACSP-FL shows that adapting the number and choice of federated-learning clients and sharing only part of each model can cut communication by up to 99% while keeping or raising accuracy on activity-recognition data.","keywords":["federated learning","client selection","model personalization","partial model sharing","communication efficiency","human activity recognition","non-IID data"],"falsifier":"Re-run the benchmarks evaluating every method two ways—once on the global model alone and once with the same best-of-local/global personalization applied to all baselines; if ACSP-FL's accuracy advantage disappears or reverses when baselines get the personalization, then the claimed convergence gain is an artifact of the evaluation protocol rather than of the adaptive client selection.","tokens_in":19406,"feed_emoji":"📉","tokens_out":10395,"duration_ms":83620,"temperature":0.7,"pith_summary":"ACSP-FL is a federated learning scheme built around one idea: the clients that need training most are the ones whose personalized models perform worst, and the network should only hear from those clients, in shrinking numbers, and never about the whole model. The paper argues that this triple mechanism—adaptive client selection, a decay that prunes the participant set as the model converges, and partial model sharing with client-side personalization—cuts the bytes exchanged between server and devices by up to 99% compared to standard federated averaging, on three human activity recognition datasets, while the distributed accuracy is as good as or better than the baselines. That matters because communication is the main bottleneck in federated learning: if the number of participating devices and the model size can both shrink during training without hurting accuracy, federated learning becomes practical on bandwidth-limited mobile networks.","feed_headline":"Picking only weak clients cuts federated learning traffic 99%","feed_subtitle":"Training only below-average devices and sharing partial models keeps accuracy on three activity-recognition datasets.","key_machinery":"The argument runs on three coupled mechanisms. The selection filter $\\pi(i, \\mathcal{A})$ (Equation 4) admits a client only if its accuracy $\\mathcal{A}_i$ is at or below the mean accuracy of all clients; the decay function $\\phi(\\mathcal{S}, t) = \\lceil |\\mathcal{S}| \\cdot (1 - \\text{decay})^t \\rceil$ (Equation 6) then keeps only the first $\\phi(\\mathcal{S},t)$ of those below-average clients, so the number of participants falls as the model improves. Communication is cut by the layer-sharing operator $\\mathcal{S}(w, L)$, which sends only the chosen layers of the model in both directions, and by the dynamic rule (Equation 9) that shares all layers only while accuracy is low and progressively shares fewer layers as a client's accuracy rises. Finally, the personalization function $\\mathcal{I}(w_i^l, w^g)$ (Equation 8) lets each client keep the better of its fully local model and the global model at inference, which the paper argues is what keeps accuracy high even though most of the model is trained locally. Together these form a loop: better per-client models raise the mean accuracy, which shrinks the selected set and the number of shared layers, which in turn cuts bytes and latency.","core_discovery":"The central claim is that ACSP-FL, by coupling performance-based client selection with model personalization and layer sharing, achieves communication-efficient convergence without the accuracy loss usually associated with aggressive communication reduction. Selection is driven by each round's distributed evaluation: clients whose accuracy falls at or below the mean become the training set for the next round, and a decay function then shrinks that set further as rounds progress, so participation drops automatically at convergence instead of staying fixed. In parallel, each client keeps a fully local model and receives only the shared layers of the global model; at evaluation time it keeps whichever of the two—local or global—performs better on its own data, and the number of shared layers is reduced as client accuracy rises. On the UCI-HAR, MotionSense, and ExtraSensory datasets, ACSP-FL is reported to send roughly 99%, 99%, and 97% fewer bytes than FedAvg respectively, cut latency by about 90%, and achieve higher distributed accuracy, including on the non-IID ExtraSensory data.","pith_inferences":["Because ACSP-FL's selection rule reacts to personalized client accuracy, the natural ablation—running the same pipeline without the Equation 8 personalization—would separate how much of the byte reduction comes from adaptive selection and layer sharing versus from the accuracy boost that personalization itself provides.","The selection loop it creates (personalization raises accuracy, which shrinks the below-average set, which cuts bytes further) leaves open whether the chosen clients stay representative on strongly heterogeneous data; logging the data distribution of selected clients across rounds would test this.","Each client now maintains both a local and a global model and runs an extra evaluation pass per round, so the scheme trades per-device storage and computation for communication; reporting those costs alongside the byte savings would show where the tradeoff lands."],"forward_implications":["If the reported reductions hold, federated learning can run with roughly 1% of its usual uplink and downlink traffic on the same model architecture, with no fixed per-round client count to tune.","The automatic shrink of both the participant set and the shared layer count means the scheme needs no k-value or compression schedule set in advance, which removes a common hyperparameter burden for deployment.","Because clients with above-average personalized accuracy stop being called to train, devices save both bandwidth and compute, which directly translates into longer battery life and lower network cost in mobile settings.","The same or better accuracy is reported on a non-IID benchmark (ExtraSensory), suggesting the approach is not limited to identically distributed client data."],"supporting_citations":[{"why":"Introduced FedAvg, the baseline whose transmitted bytes and latency serve as the reference for every comparison in the paper.","marker":"[10]"},{"why":"Proposed Power-of-Choice, a main comparison baseline and the source of the convergence analysis for partial client participation that ACSP-FL builds on.","marker":"[7]"},{"why":"Introduced OORT, the utility- and delay-aware selection baseline included in the accuracy and overhead evaluations.","marker":"[22]"},{"why":"Introduced DEEV, the earlier adaptive client selection work that ACSP-FL extends with layer sharing and personalization, and a direct comparison target.","marker":"[9]"},{"why":"Surveys and motivates model personalization, underpinning ACSP-FL's local-plus-global model design and the client-drift problem it addresses.","marker":"[8]"}],"fun_headline_variants":["Adaptive client picker cuts federated learning bytes by 99%","Personalized federated learning reduces communication 99%","Weak-client selection trims federated traffic 99% without accuracy loss","ACSP-FL: adaptive selection and personalization cut FL costs 95%+","Adaptive selection and personalization cut FL traffic to 1%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy comparison is asymmetric: ACSP-FL clients pick the better of their local and global model at evaluation time, while FedAvg, POC, OORT, and DEEV are judged only on the shared global model, so if the baselines were given the same personalization the gap might shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive client picker cuts federated learning bytes by 99%","Personalized federated learning reduces communication 99%","Weak-client selection trims federated traffic 99% without accuracy loss","ACSP-FL: adaptive selection and personalization cut FL costs 95%+","Adaptive selection and personalization cut FL traffic to 1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3194,"prompt_tokens":949,"completion_tokens":2245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2151}},"tokens_in":565,"tokens_out":2245,"duration_ms":14333,"temperature":1.0,"reasoning_tokens":2151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:47:19.096141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the benchmarks evaluating every method two ways—once on the global model alone and once with the same best-of-local/global personalization applied to all baselines; if ACSP-FL's accuracy advantage disappears or reverses when baselines get the personalization, then the claimed convergence gain is an artifact of the evaluation protocol rather than of the adaptive client selection.","supporting_citations":[{"cited_title":"McMahan, E","cited_arxiv_id":null,"evidence_quote":"Introduced FedAvg, the baseline whose transmitted bytes and latency serve as the reference for every comparison in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced OORT, the utility- and delay-aware selection baseline included in the accuracy and overhead evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced DEEV, the earlier adaptive client selection work that ACSP-FL extends with layer sharing and personalization, and a direct comparison target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Surveys and motivates model personalization, underpinning ACSP-FL's local-plus-global model design and the client-drift problem it addresses."}],"review_version":1}