{"id":"5744c244-000b-4fc4-b295-dc935a78d4da","arxiv_id":"2504.15724","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"C-SFL splits the model into three parts across weak clients, local aggregators, and a server, enabling per-epoch aggregation of the middle portion, and reports reduced delay and communication with improved accuracy in experiments.","lead":"This paper proposes Collaborative-Split Federated Learning (C-SFL), which splits a neural network into three parts trained by weak clients, strong client-aggregators, and the server, and aggregates the middle part every epoch. The authors report lower training delay and communication overhead than two existing split federated learning baselines, with slightly higher accuracy on three image datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2) uses the weak client's speed for aggregator-side FP instead of the aggregator's speed, overestimating aggregator delay by the heterogeneity ratio; since (h,v) selection and delay-axis plots depend on this model, the delay and same-delay accuracy claims are not yet established.","rationale":"The reader's weakest assumption identifies the delay model in Eqs. (1)-(5) as the hinge for the central claim. I agree: the delay model is load-bearing because it jointly determines the optimal collaborative/cut layers and the x-axis of the accuracy-vs-delay curves. The specific notation error in Eq. (2) is not cosmetic; substituting p_n for p_k introduces an 8x error under the reported heterogeneity, so the optimal (h,v) and the delay comparisons could be materially different after correction. Eq. (3) also mixes indices in ways that obscure the max operation. The absence of any validation against measured round delays, and the absence of an explicit delay model for the SFL and LocSplitFed baselines, compound the issue. However, these are correctable modeling issues rather than a fundamental flaw in the three-way split architecture. The accuracy measurements, while lacking error bars, are presented as simulation outputs and the qualitative accuracy ranking could survive a corrected delay model. For these reasons, the reader's CONDITIONAL verdict should stand, and no further adjustment is needed. The concrete test I propose would settle the concern by checking whether the corrected model preserves the reported (h,v) selections and the delay-based superiority.","tokens_in":10860,"tokens_out":11152,"duration_ms":97413,"concrete_test":"Independently re-derive Eqs. (1)-(5) from the four-phase workflow, fix Eq. (2) by replacing p_n with p_k and the index errors in Eq. (3), then re-run the O(V^2) search over (h,v) for the Table 5 configurations (MNIST/FMNIST/CIFAR-10, gamma=8.5 and gamma=1, R=2 and R=10 Mbps). If the optimal (h,v) changes from Table 5 or the round delays D_round change by more than a few percent, recompute the Fig. 2 accuracy-vs-delay curves with the corrected model and check whether C-SFL still dominates at matched delay.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central delay claim is grounded in the delay model of Section 3.3, Eqs. (1)-(5). That model is used both to choose the collaborative/cut layers (h,v) and to produce the x-axis of the delay-vs-accuracy plots in Fig. 2. Eq. (2) contains a clear scaling error: the aggregator-side FP term writes sum_{j in [h,v]} f_j * sum_{n in N} x_{k,n} / p_n, using the weak client's speed p_n (2 GHz in the experiments) instead of the aggregator's speed p_k (16 GHz). With the reported heterogeneity ratio gamma = 8.5, this overstates the aggregator's computation time by a factor of 8, which can materially shift the exhaustive search for (h,v). Eq. (3) has analogous index problems (x_{k,n} vs x_{n,k}, r_{x_n,n}) and the max is over the wrong free variables. The model is never validated against measured per-phase delays, and no delay model is specified for the SFL and LocSplitFed baselines, so the same-delay comparison in Fig. 2 is not independently reproducible. These are fixable issues, not a fatal flaw, but until the corrected model is re-run and validated, the central delay advantage is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Collaborative-Split Federated Learning (C-SFL), a hierarchical split federated learning scheme that partitions a model into three parts: layers up to a collaborative layer h trained at computationally weak clients, layers between h and a cut layer v trained at stronger 'local aggregator' clients, and the remaining layers trained at the server. The scheme trains and aggregates the aggregator-side and server-side models in parallel per epoch, with clients using a local loss at the cut layer to enable parallel backward propagation. The authors derive a delay model (Eqs. (1)-(5)) used both to select the optimal (h,v) by exhaustive search and to produce the x-axis of the delay-vs-accuracy plots, and they report higher accuracy than SFL and LocSplitFed at the same delay and communication overhead on MNIST, FMNIST, and CIFAR-10 under IID and non-IID data, along with a study of the effect of heterogeneity ratio and transmission rate.","tokens_in":11100,"tokens_out":3904,"duration_ms":35352,"significance":"If the reported gains hold, C-SFL is a meaningful contribution to heterogeneous federated learning: the idea of delegating a middle model segment to computationally strong clients and aggregating that segment per epoch is novel, and the claimed communication-overhead reduction by transmitting one aggregated model per aggregator is attractive. The paper also provides a clear system design, a well-motivated delay-based layer selection procedure, and a comparison with two relevant baselines on three datasets. The strengths are the explicit formulation of the (h,v) selection problem and the architectural novelty of parallel aggregator-side and server-side training and aggregation.","major_comments":[{"comment":"The aggregator-side FP delay term writes sum_{j in [h,v]} f_j * sum_{n in N} x_{k,n} / p_n, using the weak client's speed p_n (2 GHz in the experiments) instead of the local aggregator's speed p_k (16 GHz). With the reported heterogeneity ratio gamma = 8.5, this overstates the aggregator's computation time by a factor of 8, which can materially shift the exhaustive search for (h,v). Since the same incorrect term is used to compute the delay-axis values in Fig. 2 and Fig. 4, the central delay and same-delay accuracy claims are not yet established. Please correct the model and re-run the layer selection and experiments.","section":"Section 3.3, Eq. (2)"},{"comment":"The delay model also contains indexing and variable errors: Eq. (2) mixes x_{k,n} and x_{n,k}, uses p_n where p_k is meant, and writes the activation transmission delay as a_h / r_{x_n,n} rather than a_h / r_{k,n} or similar; Eq. (3) has analogous problems, including a max over n and k that does not match the free variables in the inner sums. Because this model determines both the selected layers (Table 5) and the x-axis of the delay plots, the quantitative delay results are not reproducible as written. Please provide a corrected, consistent formulation.","section":"Section 3.3, Eqs. (2)-(3)"},{"comment":"The delay model is never validated against measured per-phase delays on the actual hardware, and no delay model is specified for the SFL and LocSplitFed baselines. The same-delay accuracy comparison in Fig. 2 therefore rests on an unvalidated, model-based x-axis for all three schemes. Please either validate the model with wall-clock measurements of each phase or report the delay axes as computed, with the per-phase assumptions stated for each baseline.","section":"Section 4.1 and Fig. 2"},{"comment":"All accuracy results appear to come from a single run per configuration, with no error bars or statistical significance tests. Many of the claimed gains are small (e.g., 68.91% vs. 66.71% on CIFAR-10 IID, 83.2% vs. 81% on FMNIST IID), so a single run cannot support the assertion that C-SFL 'improves accuracy.' Please report results over multiple seeds with mean and standard deviation.","section":"Section 4.1 and Table 4"}],"minor_comments":[{"comment":"The local-loss MLP used to compute the loss at the cut layer is not specified: no architecture, number of layers, hidden dimensions, or training procedure is given. Since this MLP is central to the parallel training mechanism, please provide its details or a reference to a specific implementation.","section":"Section 3.2"},{"comment":"Phase 3 describes uploads from clients and aggregators to the server, but Eq. (4) uses download rates r_{s,n} and r_{s,k} in the denominators. Please change these to the corresponding upload rates r_{n,s} and r_{k,s}.","section":"Section 3.3, Eq. (4)"},{"comment":"The C-SFL communication-overhead expression contains malformed notation: one sum is written as P_{j in {1..j}} a_j, which should be {1..h}, and the parentheses are unbalanced. Please fix the expression and verify the resulting formula against the described transmission phases.","section":"Table 3"},{"comment":"The definition of B is not given before Eq. (5); the text earlier mentions 'data of B batches' but does not formally introduce B as the number of local batches per epoch. Please introduce B explicitly in Section 3.1 or 3.2 and clarify that D1 and D2 are per-batch delays.","section":"Section 3.3, Eq. (5)"},{"comment":"There are numerous typos and formatting issues: 'commmunication' in the abstract, 'trainind' in Section 2.2, 'the the' in Section 1, 'CIF AR-10' in Table 4, 'Y es' in Table 1, and inconsistent notation for x_{n,k} vs. x_{k,n} across Section 3.1 and Section 3.3. Please proofread carefully.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes a genuinely interesting architecture and the accuracy advantage is plausible, but the load-bearing delay model has clear errors and is unvalidated, and the experiments are single runs without error bars. I cannot recommend acceptance until the delay model is corrected, the (h,v) selection and delay-axis results are re-computed, and the experiments are re-run with multiple seeds. The local-loss MLP specification also needs to be added for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: C-SFL is a genuine architectural variant, not just a parameter tweak. Splitting the model into three segments and letting strong clients act as local aggregators that coalesce the middle segment before it reaches the server is new relative to SFL, SFL & Helpers, and LocSplitFed, and the accuracy gains in Table 4 are consistent across the three datasets. The communication-overhead analysis is the cleanest contribution: Table 3 makes explicit why the scheme saves bytes, and Fig. 3 backs it up.\n\nThe paper is well organized and the workflow is clearly explained. The idea of aggregating the aggregator-side models per epoch, in parallel with the server-side updates, is simple and reasonable, and the authors' positioning against the existing helper-based and local-loss schemes is fair. The empirical accuracy numbers move in the right direction, with a few percentage points improvement over LocSplitFed on MNIST and CIFAR-10.\n\nThe soft spot is the delay model. Eq. (2) divides the aggregator-side forward-pass workload by the weak client's speed p_n instead of the aggregator's speed p_k. Given the reported heterogeneity ratio gamma = 8.5, that overstates the aggregator's computation time by almost an order of magnitude. Eq. (3) has similar index problems (x_{k,n} vs x_{n,k}, r_{x_n,n}). Since the model is used both to select (h,v) and to draw the x-axis of the delay-vs-accuracy plots in Fig. 2, the headline same-delay comparisons are not yet established. There is also no delay model specified for the SFL and LocSplitFed baselines, so Fig. 2 cannot be reproduced independently. And the experiments are single runs, with no error bars or multiple seeds, no released code, and the local-loss MLP is left unspecified.\n\nThese are fixable issues, not a fatal flaw. The accuracy-vs-communication results in Fig. 3 do not depend on the delay model and support the communication-overhead advantage. The architecture itself is coherent and worth examining.\n\nRecommendation: send to peer review, but with a major-revision request targeted at the delay model — correct the indexing and heterogeneity errors, validate the model against measured per-phase delays, and report results across multiple seeds. This paper deserves referee time; the current delay claims just need to be rebuilt on a sounder base.","headline":"A genuine three-way split with per-epoch aggregator-side aggregation is new and worth referee time, but the delay model driving the headline gains has indexing and heterogeneity errors and no empirical validation, so the delay-axis plots are not yet credible.","tokens_in":11654,"tokens_out":2696,"would_cite":false,"duration_ms":24056,"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":"A three-way split of the model at clients, helper devices, and the server cuts delay and communication while raising accuracy.","keywords":["split federated learning","client heterogeneity","local aggregator","parallel training and aggregation","communication overhead","training delay","local loss","model splitting"],"falsifier":"Run C-SFL on a testbed or simulator with 100 clients, 10 local aggregators, and the same MNIST, FMNIST, and CIFAR-10 settings, measure the actual per-round delay and the optimal layers by grid search on measured times, and compare with the values predicted by Eqs. (1)–(5); if the predicted round delays deviate by more than the margins shown in the paper, or if the measured optimal layers differ from the model-selected ones, the paper's delay-gain claim is falsified.","tokens_in":10645,"feed_emoji":"🤖","tokens_out":5821,"duration_ms":51129,"temperature":0.7,"pith_summary":"The paper proposes a split federated learning scheme, C-SFL, that divides the global model into three parts: a weak-side model trained by computationally weak clients, an aggregator-side model trained by stronger client devices acting as local aggregators, and a server-side model trained by the server. The central claim is that training and aggregating the aggregator-side and server-side parts in parallel, rather than sequentially as in standard SFL, reduces per-round training delay and total communication while improving test accuracy. The paper grounds this claim in experiments on MNIST, FMNIST, and CIFAR-10, where C-SFL reports higher accuracy than SFL and LocSplitFed at the same training delay and at the same communication overhead. The two split points, collaborative layer h and cut layer v, are chosen by exhaustive search over the $O(V^2)$ candidates using a delay model. If the claim holds, C-SFL gives a practical way to keep weak IoT-style clients in federated training without forcing them to train large models.","feed_headline":"Three-way model split speeds federated learning and lifts accuracy","feed_subtitle":"C-SFL trains weak clients, strong helpers, and the server in parallel, beating SFL and LocSplitFed on three benchmarks.","key_machinery":"The central object is the split of the deep network into three parts at two named layers: the collaborative layer h and the cut layer v. The mechanism that carries the argument is local-loss parallel training: a small auxiliary network above the aggregator-side model computes the local loss at the cut layer, so weak clients can perform backward propagation without waiting for server gradients, while the local aggregator runs forward and backward propagation on the aggregator-side models and aggregates them per epoch in parallel with the server's update and aggregation. This per-epoch aggregator-side aggregation is what the paper credits for the accuracy gain, and the single aggregated model per aggregator is what cuts communication overhead.","core_discovery":"The paper's central claim is that inserting a tier of local aggregators between weak clients and the server yields all three benefits at once: lower training delay, lower communication, and higher accuracy. In C-SFL, each weak client trains only layers $1 \\ldots h$, its assigned local aggregator trains layers $h \\ldots v$ for that client and aggregates those aggregator-side models every local epoch, and the server trains layers $v \\ldots V$. Because the aggregator-side model is aggregated per epoch rather than per round, the paper argues, accuracy improves over LocSplitFed without the extra communication that frequent client-side aggregation normally costs; and because each aggregator sends one aggregate model to the server, communication drops by roughly a factor of the number of clients per aggregator. Reported accuracies on MNIST, FMNIST, and CIFAR-10 are consistently above both baselines, for example 92.83% versus 91.18% and 83.06% on IID MNIST, and 68.91% versus 66.71% and 65.7% on CIFAR-10.","pith_inferences":["A direct test of the delay model's realism would be to compare measured per-phase times on real hardware with Eqs. (1)–(5); if the measured optimal layers differ from the model-selected ones, the delay-axis gains would need revision, although the accuracy and communication comparisons may still hold.","The per-epoch aggregation of the middle segment suggests a broader principle: aggregating a shallow shared representation more often can reduce gradient staleness without extra uplink cost, and the same idea could be tested in hierarchical federated learning with more than two intermediate tiers.","The scheme exposes weak clients' activations at the collaborative layer to local aggregators, so a privacy analysis beyond 'raw data is not shared' would be needed before deployment in sensitive applications; this is not addressed in the paper.","A testable scaling question is how accuracy, delay, and communication behave as the aggregator fraction and the number of clients per aggregator grow, since the paper fixes one configuration with lambda = 0.1 and equal group sizes."],"forward_implications":["On MNIST, FMNIST, and CIFAR-10, C-SFL reports accuracy one to ten percentage points above SFL and LocSplitFed at matched training delays, under both IID and non-IID data splits.","At a fixed communication budget, C-SFL reaches materially higher accuracy: about 69% versus 57% to 60% on CIFAR-10 at 0.60 TB, and about 90% versus 77% and 59% on MNIST at 0.06 TB.","The gains are largest when client heterogeneity is high and transmission rates are low, and they shrink when all clients have equal computing speed and high-rate links.","The optimal collaborative and cut layers shift with heterogeneity and transmission rate, indicating that the two split points are tunable system parameters rather than fixed architectural choices."],"supporting_citations":[{"why":"Defines the standard SFL baseline that C-SFL extends and whose publicly available code is reused in the experiments.","marker":"[15]"},{"why":"Introduces LocSplitFed, the parallel-training baseline where clients compute a local loss at the cut layer, which C-SFL adopts and modifies.","marker":"[3]"},{"why":"Establishes split learning as the foundational client-server model splitting approach that SFL and C-SFL build on.","marker":"[18]"},{"why":"Shows that training with local error signals can degrade accuracy, which motivates the paper's per-epoch aggregator-side aggregation design.","marker":"[10]"},{"why":"Demonstrates the trade-off between aggregation frequency and communication overhead in adaptive SFL, supporting the claim that per-epoch aggregation improves accuracy without extra cost.","marker":"[7]"},{"why":"Shows that individual cut layers per client degrade accuracy, supporting the paper's choice of common collaborative and cut layers across clients.","marker":"[14]"},{"why":"Proposes helper-based parallel split learning, providing the offloading-to-stronger-clients idea that C-SFL generalizes with local aggregators.","marker":"[17]"},{"why":"Supplies the FedAvg aggregation algorithm used to combine weak-side and aggregator-side models in C-SFL.","marker":"[1]"}],"fun_headline_variants":["Three-way split reduces federated delay, overhead, and improves accuracy","Parallel aggregation in three-part FL cuts delay and lifts accuracy","C-SFL: three-way parallel training beats baselines on speed and accuracy","Weak clients, strong clients, and server train in parallel for faster, more accurate FL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the delay formulas in Eqs. (1)–(5) accurately describe the real training pipeline, since those formulas select the collaborative and cut layers and produce the delay-axis comparisons; the paper gives no measurement-based validation and some equations contain notation slips, such as Eq. (2) mixing indices and using $p_n$ where the aggregator's speed is meant.","fun_headline_variants_meta":{"raw":{"variants":["Three-way split reduces federated delay, overhead, and improves accuracy","Parallel aggregation in three-part FL cuts delay and lifts accuracy","C-SFL: three-way parallel training beats baselines on speed and accuracy","Weak clients, strong clients, and server train in parallel for faster, more accurate FL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002146,"raw_usage":{"total_tokens":8334,"prompt_tokens":963,"completion_tokens":7371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":7292}},"tokens_in":579,"tokens_out":7371,"duration_ms":50469,"temperature":1.0,"reasoning_tokens":7292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:19:01.308844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run C-SFL on a testbed or simulator with 100 clients, 10 local aggregators, and the same MNIST, FMNIST, and CIFAR-10 settings, measure the actual per-round delay and the optimal layers by grid search on measured times, and compare with the values predicted by Eqs. (1)–(5); if the predicted round delays deviate by more than the margins shown in the paper, or if the measured optimal layers differ from the model-selected ones, the paper's delay-gain claim is falsified.","supporting_citations":[{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"Defines the standard SFL baseline that C-SFL extends and whose publicly available code is reused in the experiments."},{"cited_title":"In: ICML 2021 workshop on federated learning for user privacy and data confidentiality","cited_arxiv_id":null,"evidence_quote":"Introduces LocSplitFed, the parallel-training baseline where clients compute a local loss at the cut layer, which C-SFL adopts and modifies."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"Shows that training with local error signals can degrade accuracy, which motivates the paper's per-epoch aggregator-side aggregation design."},{"cited_title":"IEEE INFOCOM 2024-IEEE Conference on Computer Communications pp","cited_arxiv_id":null,"evidence_quote":"Proposes helper-based parallel split learning, providing the offloading-to-stronger-clients idea that C-SFL generalizes with local aggregators."},{"cited_title":"Advances in Neural Information Processing Systems 35, 10572–10586 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation algorithm used to combine weak-side and aggregator-side models in C-SFL."}],"review_version":1}