{"id":"678d31d7-4125-420c-ab17-b51ed05f7cab","arxiv_id":"2412.15536","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cut layer position barely affects SFL-V1 accuracy but strongly changes SFL-V2 accuracy, with early splits usually best and often beating FedAvg.","lead":"This paper studies where to cut a neural network between clients and a server in split federated learning. The cut point barely affects one variant (SFL-V1) but dramatically changes accuracy in another (SFL-V2), where early cuts usually work best.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 analyzes τ server-side updates per round, but Algorithm 1 updates the server once; the invariance proof does not cover the implemented SFL-V1.","rationale":"Reader's weakest assumption already identifies this proof-algorithm mismatch; my reading agrees. Load-bearing because Proposition 1 is the paper's only theoretical contribution and directly supports the headline claim that SFL-V1 is invariant to cut layer. If the proof applies to a different update schedule, the main novelty ('provide theoretical analysis') is reduced to an empirical observation. The concern does not by itself overturn the empirical results: the SFL-V2 cut-layer sensitivity and FedAvg comparisons rest on experiments, and the SFL-V1 invariance pattern in Tables 3-4 is plausibly real. But the central paper claim as stated—'consistent with our theoretical results'—is not justified. The conditional verdict stands, with an additional required revision: either align Algorithm 1 with the proof by updating the server τ times per round, or prove invariance for the actual single-update algorithm. I do not see a basis for REJECT because the experiments can be corrected and likely pattern persists; nor ACCEPT because the theoretical mismatch is substantive.","tokens_in":19755,"tokens_out":5298,"duration_ms":47126,"concrete_test":"Implement Algorithm 1 exactly and, in the same code, a variant in which the Training Server performs τ=ceil(E|D_k|/B_k) gradient steps per client per round before aggregation (matching Appendix B). Run both on CIFAR-10 non-IID with K=10 and a 4-block ResNet for Lc in {1,2,3,4}, 3 seeds. If the single-update Algorithm 1 shows a cross-Lc accuracy spread larger than the within-Lc seed noise (e.g., >1.5%) while the τ-update variant is flat, the Proposition 1 invariance result applies only to the variant, not the experiments. Alternatively, re-derive the server-side bound in Eq. (18) with θ^{t+1}_s - θ^t_s = -η_t Σ_k α_k g^t_{s,k} (one step) and check whether the Lc-independent bound survives.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Appendix B's proof of Proposition 1 does not model Algorithm 1. In Eq. (18) and the surrounding derivation, the server-side update is θ^{t+1}_s - θ^t_s = -η_t Σ_{k=1}^K Σ_{i=0}^{τ-1} α_k g^{t,i}_{s,k}, i.e., τ gradient steps on the server-side model per round. Lemma 1 likewise iterates the local model for τ steps. But Algorithm 1 (lines 19 and 23-26) updates the server-side model exactly once per round, while the client-side model is updated E times with the same cut-layer gradient. Because the number of updates per round differs between the two parts, changing Lc changes which parameters receive τ updates versus one update; the algorithm is not 'equivalently transformed into FedAvg' with a consistent local-update count. Proposition 1 therefore does not establish invariance for the SFL-V1 variant whose results are reported in Tables 3-4. The empirical invariance across Lc may still hold, but the stated theoretical support for that central claim is absent unless the proof is re-derived for the single-server-update schedule or Algorithm 1 is changed to perform τ server-side updates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how the choice of cut layer affects the performance of two split federated learning variants, SFL-V1 and SFL-V2. The authors prove a convergence bound for SFL-V1 that is independent of the cut layer, and support it with experiments across two datasets and two ResNet architectures in IID and non-IID settings. They report that SFL-V1 is largely invariant to cut layer placement, while SFL-V2 is sensitive to it, with early cut layers generally performing best and sometimes beating FedAvg, especially on heterogeneous data. The paper also discusses privacy and non-IID challenges as future work.","tokens_in":20009,"tokens_out":2276,"duration_ms":21320,"significance":"If the claims hold, this would be a useful contribution to the SFL literature: cut layer selection is usually treated as a system-level or privacy-related design choice, and this paper provides the first systematic study of its effect on model accuracy, together with a convergence-style theoretical statement for SFL-V1. The empirical scope is reasonable for an initial study: four datasets, two architectures, and both IID and non-IID partitions. The claimed invariance of SFL-V1 is an interesting architectural observation, and the SFL-V2 versus FedAvg comparison addresses a practically relevant question. However, the theoretical result as written does not cover the implemented SFL-V1 algorithm, and the FedAvg comparison is confounded by different optimizers and learning rates; these issues currently limit the strength of the central claims.","major_comments":[{"comment":"Proposition 1's proof analyzes a schedule in which the server-side model receives τ gradient updates per communication round (Eq. (18) sums gradients over i = 0, ..., τ-1 and the subsequent bound uses τ server-side steps), but Algorithm 1 updates each server-side model exactly once per round. Therefore the invariance result is proved for a different algorithm than the one whose results appear in Tables 3 and 4. The authors need to either re-derive the bound for the single-server-update schedule or change Algorithm 1 and the experiments to implement τ server-side updates; otherwise the paper's central theoretical claim does not support its empirical SFL-V1 results.","section":"Appendix B, Eq. (18); Algorithm 1 (line 19)"},{"comment":"The comparison between SFL and FedAvg is confounded: SFL-V1 and SFL-V2 use Adam with learning rate 0.001, while FedAvg uses SGD with learning rate 0.01. Since both optimizer and learning rate differ, the observed SFL-V2 advantage over FedAvg (e.g., 92.30% vs. 85.25% on IID CIFAR-10) cannot be attributed to the split architecture or cut layer selection. The authors should either use the same optimizer and a comparable learning rate for FedAvg, or provide additional experiments isolating the effect of the optimizer choice.","section":"Table 2 and Tables 3-4"},{"comment":"Several non-IID conclusions rest on a small number of runs with large variance. For example, on Tiny ImageNet, SFL-V2 (Lc=1) is reported as 30.14 ± 8.58 versus FedAvg's 28.33 ± 0.28, so the claimed superiority is not supported by the overlap of the confidence ranges; similarly, SFL-V2 (Lc=2) on non-IID CIFAR-10 is 59.98 ± 11.99. With only three runs and no significance testing, the statement that SFL-V2 with an appropriate cut layer 'significantly outperforms' FedAvg on heterogeneous data is too strong. More seeds and a paired or corrected significance test are needed for the cross-condition claims.","section":"Tables 4 and Section 5.2, non-IID results"}],"minor_comments":[{"comment":"The phrase 'significant performance variations respect to with cut layer placement' contains a typo; it should read 'with respect to cut layer placement.'","section":"Section 1"},{"comment":"The claim that SFL-V1 is 'relatively invariant' is supported by small performance ranges on CIFAR-10 and CIFAR-100, but on non-IID Tiny ImageNet the SFL-V1 results (around 12.8-13.9%) are far below FedAvg (28.33%) and have non-negligible variability; the paper should comment on this discrepancy instead of only discussing CIFAR results.","section":"Section 5.2, Tables 3-4"},{"comment":"The definition of τ as ⌈E D_k / B_k⌉ is fine, but the proof consistently treats τ as an integer number of client updates per round; the paper should clarify how E local epochs in Algorithm 1 translate into exactly this τ, since the pseudocode does not explicitly sample multiple batches per epoch.","section":"Proposition 1, Eq. (6)"},{"comment":"The client backward pass in Algorithm 1 (lines 23-26) uses the same gradient ∇a_k(t) for E epochs, which is unusual; this point should be explained, since it differs from the standard SFL description in the text and from the proof's local-update model.","section":"Section 3 and Appendix A"},{"comment":"The conclusion states that SFL-V2 outperforms FedAvg 'in both IID and non-IID settings', but Table 4 shows several configurations where FedAvg is competitive or better; this should be qualified to reflect the actual experimental conditions.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper's topic is timely and the empirical study is broadly useful, but the proof-algorithm mismatch and the optimizer confound are serious enough that the current version overstates its theoretical and empirical support. I do not see grounds for rejection: the empirical invariance of SFL-V1 could still be demonstrated after fixing the proof, and the SFL-V2 comparison could be made fair. The paper is better suited to a venue where empirical and algorithmic claims in distributed learning are evaluated; the contribution is incremental but meaningful if the open issues are addressed. I would ask the authors to provide the missing proof for the implemented schedule and to rerun or reanalyze the FedAvg baseline with matched optimization settings before reconsidering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is the empirical picture. It is the first systematic look at how cut layer placement affects SFL-V1 and SFL-V2 accuracy across four datasets and two ResNet sizes, and the main pattern is clear: SFL-V2 accuracy swings with cut layer (early cuts usually best, by up to 10 points on non-IID CIFAR-100), while SFL-V1 is roughly flat. If you are choosing a cut layer in an SFL-V2 deployment, this gives you a concrete rule of thumb. The authors also report mean and std over three runs and honestly flag the non-IID CIFAR-10 outlier.\n\nThe soft spots are concentrated in the theory and the comparison. The convergence proof in Appendix B analyzes a schedule where the server-side model receives tau gradient steps per round per client (Eq. 18 sums over i=0 to tau-1). But Algorithm 1 updates each client's server model exactly once per round, then runs E local epochs on the client. So Proposition 1's invariance result is proved for a different algorithm than the one in the experiments. The empirical invariance of SFL-V1 may be true, but the stated theoretical support does not apply as written. That is a load-bearing gap, not a typo; the proof needs to be re-derived for the single-server-update schedule or the algorithm changed. The proof also \"mainly follows\" Han et al. 2024b with overlapping authorship, which makes the lack of a fresh derivation more visible.\n\nThe FedAvg comparison is also not apples-to-apples: SFL uses Adam at LR 0.001, FedAvg uses SGD at LR 0.01. That weakens the claim that \"SFL-V2 outperforms FedAvg\" — part of the gap could be optimizer and tuning. And the non-IID results have large variance with only three runs and no significance testing; the CIFAR-10 non-IID numbers (std up to 11.99) put some cut-layer orderings within noise.\n\nOverall: send this to a serious referee, but require major revision. The empirical core will likely survive a fairer comparison, and the question is worth answering. The theory needs to match the algorithm, the baseline needs to be controlled, and the variance needs to be addressed. For readers working on SFL, the empirical result is worth knowing; for theory readers, the proof needs work before it supports the invariance claim.","headline":"Useful empirical map of cut-layer effects in split federated learning, but the invariance proof doesn't cover the algorithm that was actually run.","tokens_in":20526,"tokens_out":3503,"would_cite":true,"duration_ms":31051,"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":"Cut-layer selection is provably neutral for SFL-V1 but performance-critical for SFL-V2, where an early split beats FedAvg on heterogeneous data.","keywords":["split federated learning","cut layer selection","federated learning","split learning","convergence analysis","data heterogeneity","non-IID data","FedAvg"],"falsifier":"Run SFL-V1 on non-IID CIFAR-10 with fixed seeds across cut layers $L_c=1,\\dots,4$ while logging the number of server-side gradient updates per communication round. If test accuracy shifts by more than the reported run-to-run spread (about 1.9 points) as the cut layer moves, the claimed invariance is empirically false; if accuracy stays flat but the log shows one server update per round, the invariance holds for the implemented algorithm while Proposition 1's proof, which sums $\\tau$ server updates per round, does not cover that algorithm.","tokens_in":19525,"feed_emoji":"✂️","tokens_out":13008,"duration_ms":92430,"temperature":0.7,"pith_summary":"Split federated learning divides a neural network at a cut layer, putting early layers on client devices and the rest on a training server. This paper asks whether the placement of that cut layer changes what the model learns, and how the resulting accuracy compares with standard federated averaging. The answer it argues for is asymmetric: SFL-V1, which keeps a separate server-side model per client, is provably and experimentally insensitive to the cut layer, while SFL-V2, which shares one server-side model across clients, depends strongly on cut depth, with early cuts performing best. In experiments across four datasets and two ResNet architectures, an early cut in SFL-V2 outperformed FedAvg on both IID and non-IID data, with the largest gains on the most skewed distributions. The practical upshot is that cut-layer selection can be treated as a free design choice in SFL-V1 but as a performance-critical hyperparameter in SFL-V2.","feed_headline":"Split point: a free knob for SFL-V1, a decisive one for SFL-V2","feed_subtitle":"Proof plus experiments show cut placement is neutral for SFL-V1 but decides whether SFL-V2 beats FedAvg on skewed data.","key_machinery":"The central object is the cut layer $L_c$, the index at which the network is split so that layers $1,\\dots,L_c$ live on clients and layers $L_c+1,\\dots,L$ live on the training server. The argument's load-bearing piece is Proposition 1, a convergence bound for SFL-V1 derived under standard non-convex smoothness, bounded-variance, and bounded-heterogeneity assumptions; the proof decomposes each round's descent into client-side ($\\theta_C$) and server-side ($\\theta_S$) updates and shows the resulting update sequence is identical to FedAvg for every $L_c$, which is why the bound contains no dependence on the cut layer. On the empirical side, the machinery is the comparison grid: the same four cut positions ($L_c=1,2,3,4$) are applied to ResNet-18 and ResNet-50 across CIFAR-10, CIFAR-100, Tiny ImageNet, and HAM10000, under IID and label-skewed (Dirichlet $\\mu=0.1$) partitions, isolating cut-layer effects from architecture and data effects. The contrast between the two SFL variants is explained structurally: SFL-V1 preserves per-client independence through separate server-side models, while SFL-V2's shared server-side model learns from all clients' activations without weight averaging.","core_discovery":"The paper's central claim is that the two variants of split federated learning respond to cut-layer selection in opposite ways. For SFL-V1, where the training server keeps a separate server-side model per client, the authors prove (Proposition 1) that the convergence bound is independent of the cut layer $L_c$ for any $L_c \\in \\{1,\\dots,L-1\\}$ under standard assumptions (non-convex smooth losses, bounded gradient variance and heterogeneity), because any such configuration is equivalent to FedAvg with identical model updates. For SFL-V2, where a single shared server-side model processes all clients sequentially, no such invariance holds: test accuracy shifts substantially with cut depth, with $L_c=1$ (the earliest cut) giving the best accuracy on three of four datasets, and SFL-V2 at $L_c=1$ beating FedAvg by margins up to 9.78 points on non-IID data (52.38% vs 42.60% on CIFAR-100). The paper reads the cut layer as an interpolation knob between centralized learning ($L_c=0$) and FedAvg ($L_c=L$), and leaves a convergence proof for SFL-V2 to future work.","pith_inferences":["One consequence the paper leaves implicit: SFL-V1's invariance makes deeper cut layers a free privacy upgrade, since moving the split toward the server hides more input structure from the server-side model without measurable accuracy loss.","The proof-implementation mismatch over server update counts suggests a concrete test: re-running SFL-V1 with the server updated once per local step, as the proof assumes, could make the invariance slightly less clean in practice, separating an architectural fact from a proof artefact.","The paper's single counterexample (non-IID CIFAR-10, where $L_c=4$ beat $L_c=1$) hints that \"early cut is best\" is dataset-dependent; a cheap per-dataset probe across cut layers could turn the cut layer into a tunable knob rather than a fixed default.","SFL-V2's edge over FedAvg on skewed data echoes mechanisms studied in personalized and split learning; combining SFL-V2's shared server-side model with control variates or proximal correction, as the discussion suggests, is a natural next experiment."],"forward_implications":["In SFL-V1, cut-layer placement can be chosen purely to minimize communication, client computation, or privacy risk; test accuracy will not move with the choice.","In SFL-V2, the cut layer is effectively a hyperparameter connecting centralized learning ($L_c=0$) to FedAvg ($L_c=L$); early cuts are the empirically best default.","On heterogeneous data, SFL-V2 with an early cut is a viable replacement for FedAvg: a 9.78-point gain on non-IID CIFAR-100 and a 7.05-point gain on IID CIFAR-10 in the reported runs.","Deployments that need per-client server-side models can use SFL-V1 and expect FedAvg-like accuracy regardless of where the network is split.","Since SFL-V1 behaves like FedAvg across cut layers, the architectural difference between V1 and V2, not the cut position, is what explains the accuracy gap between the two variants."],"supporting_citations":[{"why":"Introduces SFL-V1 and SFL-V2, the two architectures whose cut-layer sensitivity is compared.","marker":"(Thapa et al. 2022)"},{"why":"Supplies the convergence-analysis template that Proposition 1 extends to cut-layer-specific client and server updates.","marker":"(Han et al. 2024b)"},{"why":"Defines FedAvg, the aggregation rule used for model synchronization and the FL baseline SFL is compared against.","marker":"(McMahan et al. 2017)"},{"why":"Provides the Dirichlet non-IID data partition used to generate the heterogeneous client datasets.","marker":"(Hsu, Qi, and Brown 2019)"},{"why":"Establishes the privacy-energy tradeoff of cut-layer placement that motivates studying its performance impact.","marker":"(Lee et al. 2024)"},{"why":"Defines the ResNet residual-block boundaries used as the cut positions ($L_c=1,\\dots,4$).","marker":"(He et al. 2016)"}],"fun_headline_variants":["Cut layer choice: indifferent for SFL-V1, pivotal for SFL-V2","Where to split? For SFL-V1 it's a no-op, for SFL-V2 it decides vs FedAvg","Split federated: cut layer matters only for shared-server variant","SFL-V2 beats FedAvg when cut early; SFL-V1 doesn't care","Cut depth dictates SFL-V2's edge over FedAvg, not SFL-V1's"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The invariance result rests on the assumption that the server-side model is updated on every local training step, while Algorithm 1 as written updates the server-side model only once per round, so if the experiments follow Algorithm 1 the proof does not cover them.","fun_headline_variants_meta":{"raw":{"variants":["Cut layer choice: indifferent for SFL-V1, pivotal for SFL-V2","Where to split? For SFL-V1 it's a no-op, for SFL-V2 it decides vs FedAvg","Split federated: cut layer matters only for shared-server variant","SFL-V2 beats FedAvg when cut early; SFL-V1 doesn't care","Cut depth dictates SFL-V2's edge over FedAvg, not SFL-V1's"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3366,"prompt_tokens":993,"completion_tokens":2373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2256}},"tokens_in":609,"tokens_out":2373,"duration_ms":11925,"temperature":1.0,"reasoning_tokens":2256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:19:54.815816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SFL-V1 on non-IID CIFAR-10 with fixed seeds across cut layers $L_c=1,\\dots,4$ while logging the number of server-side gradient updates per communication round. If test accuracy shifts by more than the reported run-to-run spread (about 1.9 points) as the cut layer moves, the claimed invariance is empirically false; if accuracy stays flat but the log shows one server update per round, the invariance holds for the implemented algorithm while Proposition 1's proof, which sums $\\tau$ server updates per round, does not cover that algorithm.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces SFL-V1 and SFL-V2, the two architectures whose cut-layer sensitivity is compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the privacy-energy tradeoff of cut-layer placement that motivates studying its performance impact."}],"review_version":1}