{"id":"9b470bf9-e23d-4d63-a335-bbaae12f988f","arxiv_id":"2412.17150","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A learned rate-distortion codec trained jointly with a split-federated U-Net compresses transmitted features and gradients, achieving 1e3-1e4x data-transfer reduction at matched accuracy on skin-lesion segmentation, with smaller and inconsistent gains on blastocyst segmentation.","lead":"SplitFedZip adds learned compression codecs at the split points of a split-federated medical image segmentation network, compressing the features and gradients that clients send to the server. On skin-lesion data it cuts transferred data by roughly a thousandfold at matched accuracy; on embryo data accuracy drops at the best evaluated fidelity (MJI 0.847 vs 0.892 without compression).","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'without hindering' claim fails on Blastocyst: no operating point achieves both CR≥1000 and MJI≥0.892; best CR at near-NC accuracy is ~100x, not 1000x.","rationale":"The reader's weakest_assumption focuses on the data-transfer accounting (missing epoch multiplier, equal 40.6 GB NC for different dataset sizes). That is a real concern for the comparison with Ayad et al., but the more load-bearing issue for the paper's headline claim is the internal contradiction on Blastocyst accuracy. The conclusion explicitly states 'at least three orders of magnitude ... without hindering the global model performance' on two datasets. Table 3 and Fig. 2 show that on Blastocyst, the operating points that reach CR ≥ 1000 have MJI well below 0.892, while the points near NC accuracy have only ~100x reduction. This directly falsifies the 'without hindering' conjunct for Blastocyst, independent of any epoch-accounting ambiguity. The DT formula issue is secondary because the per-transmission compression ratio (Fig. 4, CR definition) is what supports the three-orders-of-magnitude claim, and that ratio is epoch-independent. However, both issues point to the same need for a corrected, dataset-specific presentation of results. Since the reader already issued CONDITIONAL, my read does not change the verdict; I agree with CONDITIONAL but for a different primary reason.","tokens_in":9112,"tokens_out":9260,"duration_ms":75155,"concrete_test":"From Table 3, for the Blastocyst two-stage rows, compute CR = 40.6/DT for each row and tabulate against MJI. Verify whether any row with CR≥1000 has MJI ≥ 0.892. If none, the 'without hindering' claim fails for Blastocyst; the conclusion should be restricted to HAM10K or the authors should re-run Blastocyst experiments at higher λ or with retrained codecs to locate an operating point that satisfies both criteria.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 'Comparison with existing SplitFed compression methods' and Table 3 (bottom) undermine the conclusion's central claim that SplitFedZip provides 'at least three orders of magnitude' transfer reduction 'without hindering' on both datasets. For Blastocyst, the NC MJI is 0.892 and NC DT is 40.6 GB. The best two-stage MJI that stays reasonably close is 0.884 at λ=64, but its DT is 0.4 GB, giving a reduction of only 40.6/0.4 ≈ 101x. The only rows with CR≥1000 (DT ≤ 0.0406 GB) have MJI ≤ 0.810 (e.g., λ=1: MJI 0.810, DT 0.03; λ=0.6: MJI 0.799, DT 0.02). Thus on Blastocyst, the method either achieves ~100x reduction with a small accuracy drop, or >1000x reduction with a substantial drop (0.892→0.799). The conclusion as stated is therefore not supported by the paper's own data. This is an internal inconsistency, not a matter of external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SplitFedZip, a method that inserts learned compression codecs (a Ball\\'e-style hyperprior AE and the Cheng2020 attention model) at the two split points of a split U-Net in Split-Federated learning, compressing both forward features and backward gradients. Training minimizes a rate term plus a λ-weighted combination of Dice and MSE distortion. Experiments on Blastocyst and HAM10K segmentation report rate-accuracy curves, BD-MJI/BD-BPP comparisons between codecs, data-transfer (DT) reductions versus no compression, and a comparison with the earlier SplitFed compression method of Ayad et al. The paper claims transferred-data reduction of at least three orders of magnitude on both datasets without hindering final model performance.","tokens_in":9373,"tokens_out":5853,"duration_ms":48344,"significance":"The idea of applying end-to-end rate-distortion optimized learned compression specifically to the features and gradients flowing through SplitFed split points is a meaningful and timely contribution, and the paper is honest in presenting R-A curves as empirical sweeps over λ with bit-rates measured from actual bitstreams. The open-source implementation, use of standard CompressAI codecs, and the BD analysis (showing Cheng AT generally saves bits over the AE) are strengths. However, the central quantitative claim is overstated: the paper's own Table 3 does not support 'at least three orders of magnitude ... without hindering' on the Blastocyst dataset, and the DT accounting used for the comparison with Ayad et al. is not sufficiently specified to verify the reported ratios. These issues affect the main advertised conclusion and require correction before publication.","major_comments":[{"comment":"The conclusion states that SplitFedZip achieves 'transferred data reduction, at least three orders of magnitude compared to no compression, on two medical image segmentation datasets, without hindering the global model performance.' This is not supported by Table 3 for the Blastocyst dataset. With NC MJI = 0.892 and NC DT = 40.6 GB, the best two-stage operating point near NC accuracy is λ = 64 with MJI = 0.884 and DT = 0.4 GB, which is only a 40.6/0.4 ≈ 101x reduction. The rows with CR ≥ 1000 (e.g., λ = 1: MJI 0.810, DT 0.03 GB; λ = 0.6: MJI 0.799, DT 0.02 GB) have MJI well below 0.892, and the two-phase Blastocyst results never reach 0.892 (max MJI 0.847). The conclusion should be revised to state dataset- and operating-point-specific reductions, e.g., roughly 1e2x on Blastocyst and 1e3x or more on HAM10K for near-NC accuracy, or must be backed by a Blastocyst CR curve showing a ≥1e3x point with MJI equal to NC.","section":"Conclusion; Table 3"},{"comment":"The data-transfer (DT) figures in Table 3 are not verifiable from the formulas given. For SplitFedZip, DT is defined as 'BPP-T × input image resolution × no. of training samples × 2', but training uses 12 local epochs and 10 global epochs, and no epoch multiplier appears in the formula; the same omission applies to the Ayad et al. DT definition. Moreover, NC DT is listed as 40.6 GB for both Blastocyst (about 664 training samples) and HAM10K (about 8513 training samples). If DT scales with the number of training samples, these two values cannot both be 40.6 GB unless the per-sample uncompressed transfer differs by a factor of roughly 12.8, which is not explained anywhere. Please specify the exact computation, including how many epochs, which image resolution is used for each dataset, and whether forward and backward transmissions are both counted, and report the resulting per-sample and per-epoch DT values so the claimed 3.2x–5000x reductions over Ayad et al. can be checked.","section":"Comparison with existing SplitFed compression methods; Table 3"},{"comment":"The generalization of the 'at least three orders of magnitude' claim from Fig. 4 to both datasets is not justified in the text. Fig. 4 shows CR versus λ only for the HAM10K dataset, and the passage immediately after it ('In these experiments, SplitFedZip reduces the transferred data by at least three orders of magnitude without compromising model performance') is made in the context of HAM10K examples, yet the conclusion applies it to both datasets. Please either include the corresponding Blastocyst CR curves or restrict the claim to HAM10K and state the Blastocyst reduction separately.","section":"Results and analysis; Fig. 4"}],"minor_comments":[{"comment":"The phrase 'where data privacy is of utmost performance' should read 'of utmost importance'.","section":"Introduction"},{"comment":"Equations (1)–(3) are not numbered in the text; please add equation numbers for easier reference, and define the superscripts (e.g., L^{S_i,F}_r) explicitly as 'rate loss at split point i for features'.","section":"Training loss"},{"comment":"The sentence 'The horizontal green line represents the MJI during NC (0.892 for both datasets)' should clarify that this is the NC MJI for each dataset separately; both happening to be 0.892 is a coincidence that should be stated explicitly.","section":"Experimental setup"},{"comment":"The description of the threshold gthres in Ayad et al.'s method is confusing: 'gradients are backpropagated without further compression when the loss exceeds an adaptive threshold (gthres; 0 ≤ gthres ≤ ∞)' should state clearly whether gthres = ∞ means gradients are always compressed, and how the threshold is adapted during training.","section":"Comparison with existing SplitFed compression methods"},{"comment":"The caption 'The average MJI for the complete test sets are listed down' is ungrammatical; please rephrase, e.g., 'The average test-set MJI for each setting is listed below the sample images.'","section":"Table 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the empirical methodology is largely transparent, but the headline claim is contradicted by the paper's own Table 3 on Blastocyst, and the DT accounting is underspecified. I believe both issues are fixable within the scope of the manuscript: revise the conclusion to state dataset-specific reduction ranges, and add a clear DT computation including epoch multipliers and per-dataset resolutions. The paper would then be a solid contribution to the SplitFed communication-efficiency literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read SplitFedZip. Bottom line: the idea is a natural and well-executed extension — insert learned rate-distortion codecs at both split points of a SplitFed network, train them jointly with the segmentation model — and on HAM10K the numbers really do show large transfer reductions at maintained accuracy. But the paper's headline claim, that it achieves at least 1000x transfer reduction without hindering performance on both datasets, is not supported by its own Table 3 for Blastocyst. At the best operating point that approaches NC accuracy (two-stage MJI 0.884 vs NC 0.892), the compression ratio is about 101x. To get CR ≥ 1000 you have to drop MJI to 0.810 or lower. So the 'without hindering' qualifier needs to be dataset-specific, and the abstract/conclusion overstate the result.\n\nWhat's genuinely new: to my knowledge this is the first end-to-end RD-optimized compression of both features and gradients in SplitFed, with a comparison against the only prior SplitFed compression method (Ayad et al., reimplemented). The R-A curves come from actual bitstreams, and the BD-BPP/MJI analysis is a reasonable way to compare codecs. Code is released. That's real empirical work.\n\nThe main soft spots, in order:\n1. The Table 3 comparison shows NC DT of 40.6 GB for both Blastocyst (781 samples) and HAM10K (10,015 samples). That is almost certainly wrong; the two datasets differ by an order of magnitude in size. The DT formulas lack an epoch multiplier, so it's unclear whether the 3.2x–5000x reductions over Ayad et al. are apples-to-apples. This needs a corrected accounting with explicit training schedule.\n2. The 'without hindering' claim is not dataset-agnostic. The paper should say HAM10K achieves 3+ orders with no loss; Blastocyst achieves ~100x with a small drop or 1000x with a substantial drop.\n\nMinor: NC MJI being 0.892 for both datasets looks suspicious even though it's possible; worth a note.\n\nI don't think this is a desk-reject. The mechanism is solid, the evaluation is mostly transparent, and the flaws are correctable with a revised comparison and more careful claims. I'd send it to a serious referee, with the expectation of major revisions. I wouldn't cite the transfer-reduction numbers until the accounting is fixed.","headline":"Sound SplitFed compression study with a strong HAM10K result, but the conclusion overclaims on Blastocyst and the comparison table has accounting problems.","tokens_in":9900,"tokens_out":2831,"would_cite":false,"duration_ms":25132,"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":"Learned rate-distortion codecs at the split points of a split-federated network reduce transferred data by at least three orders of magnitude while preserving segmentation accuracy.","keywords":["split federated learning","learned compression","rate-distortion optimization","medical image segmentation","communication efficiency","gradient compression","feature compression","Bjøntegaard Delta"],"falsifier":"Recompute the data transfer for both methods from a single explicit formula that includes the number of training samples, local epochs, global epochs, and the number of transmissions per sample, using identical settings for SplitFedZip and the baseline; if the reduction ratio does not remain in (or near) the 3.2x-5000x range, the central communication-savings claim is refuted.","tokens_in":8907,"feed_emoji":"📉","tokens_out":8958,"duration_ms":71792,"temperature":0.7,"pith_summary":"SplitFedZip is a way to compress the tensors that travel between client and server in split-federated learning. The paper places learned neural compressors at the two split points of a split U-Net and trains them together with the segmentation model, so that both forward feature maps and backward gradients are compressed before transmission and reconstructed on arrival. On two medical image segmentation datasets, Blastocyst and HAM10K, the paper reports that this reduces transferred data by at least three orders of magnitude relative to no compression while keeping the mean Jaccard index at the uncompressed baseline (0.892 on both datasets). A single weighting parameter $\\lambda$ controls the trade-off between bit-rate and accuracy, and a suitable value reaches baseline accuracy at compression ratios of roughly $10^{3.6}$ to $10^{4.4}$. The authors describe this as the first end-to-end rate-distortion optimized compression scheme for SplitFed learning.","feed_headline":"Learned codecs cut split-federated traffic 1000x","feed_subtitle":"Rate-distortion compressors at split points preserve segmentation accuracy while cutting transferred data.","key_machinery":"The machinery is a pair of learned codecs attached at the split points of the network, with separate codecs for forward features and backward gradients. Each codec is trained jointly with the segmentation model using the loss $\\mathcal{L} = \\mathcal{L}_r + \\lambda(\\mathcal{L}_{\\text{Dice}} + \\mathcal{L}_{\\text{mse}})$, where $\\mathcal{L}_r$ is the rate estimate from the codec's entropy model and $\\lambda$ trades bit-rate against a distortion term that includes both the Dice segmentation loss and reconstruction mean squared error. Federated averaging is applied to codec parameters as well as model parameters, so the compressors adapt to the changing statistics of the tensors they compress over global epochs. Two codec architectures are tested: a hyperprior autoencoder with generalized divisive normalization and an attention-based model with discretized Gaussian mixture likelihoods, both borrowed from learned image compression.","core_discovery":"The central claim is that learned rate-distortion codecs can be inserted into the SplitFed training loop without disturbing convergence: the codecs are updated by the same federated averaging that updates the network, and they learn to match the statistics of features and gradients as training proceeds. The paper demonstrates that both a feature-only scheme and a feature-plus-gradient scheme let the final model match the no-compression segmentation accuracy, with the feature-plus-gradient scheme compressing traffic in both directions. It also shows that an attention-based codec with Gaussian mixture entropy modeling saves roughly 69% bit-rate over a hyperprior autoencoder in the feature-only scheme and 24% in the feature-plus-gradient scheme, measured by Bjøntegaard $\\Delta$ on weighted average, with only about 0.002-0.009 change in mean Jaccard index. The discovery, as the authors state it, is that SplitFed communication can be cut by orders of magnitude without sacrificing task performance by making compression a learned part of the federated model.","pith_inferences":["The same split-point codec insertion should transfer to other architectures that shuttle intermediate tensors, such as split transformers or pipeline-parallel models, where the tensors are even larger and the potential savings correspondingly greater; this is an extrapolation the paper does not test.","The occasional accuracy improvements seen on HAM10K at moderate compression suggest a possible regularizing effect, but the paper does not isolate it; a controlled ablation with varying $\\lambda$ schedules could clarify when compression helps rather than merely preserves accuracy.","The comparison with the prior SplitFed compression method rests on data-transfer formulas that omit explicit epoch multipliers; an independent re-computation using identical training-configuration assumptions would test whether the claimed 3.2x-5000x advantage holds under a fully consistent accounting.","A practical deployment would need to weigh the added compute and memory of the codecs themselves; the paper reports bit-rate and accuracy but not wall-clock time or energy, so the end-to-end efficiency gain remains to be quantified."],"forward_implications":["SplitFed training becomes feasible on very low-bandwidth links by choosing a $\\lambda$ that gives a 1000x or larger reduction in bytes transferred, at a modest, tunable cost in accuracy.","The compression codecs are part of the federated aggregation, so communication savings require no changes to the SplitFed privacy model or to the split-network training loop.","Compressing gradients as well as features addresses the backward pass, which is often the dominant communication cost in split learning.","The comparison metrics (BD-BPP, BD-MJI) give a concrete way to rank future codec designs for SplitFed, with the attention-based codec setting a new efficiency bar on the two test datasets."],"supporting_citations":[{"why":"Supplies the hyperprior autoencoder codec architecture used as the simpler learned compressor in the experiments.","marker":"Ballé et al. 2018"},{"why":"Supplies the attention-based codec with discretized Gaussian mixture likelihoods that achieves the larger bit-rate savings.","marker":"Cheng et al. 2020"},{"why":"Provides the federated averaging algorithm used to aggregate both split-network weights and codec weights across clients.","marker":"McMahan et al. 2017"},{"why":"Introduces the SplitFed framework that SplitFedZip optimizes.","marker":"Thapa et al. 2022"},{"why":"The prior SplitFed compression method whose data-transfer and accuracy numbers SplitFedZip is compared against.","marker":"Ayad, Renner, and Schmeink 2021"},{"why":"Defines the Bjøntegaard Delta metric used to compute average bit-rate and accuracy differences between codec rate-accuracy curves.","marker":"Bjontegaard 2001"}],"fun_headline_variants":["Learned codecs cut SplitFed traffic 1000x, keep accuracy","SplitFedZip: learned compression for leaner federated splits","Rate-distortion codecs shrink SplitFed data transfer 10x","SplitFedZip: compress to cut split-federated communication cost","Learned codecs preserve accuracy while trimming SplitFed traffic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the data-transfer figures for SplitFedZip and for the comparison method count the same number of training samples, local/global epochs, and forward/backward transmissions, so that the reported 3.2x-5000x reduction is an apples-to-apples comparison.","fun_headline_variants_meta":{"raw":{"variants":["Learned codecs cut SplitFed traffic 1000x, keep accuracy","SplitFedZip: learned compression for leaner federated splits","Rate-distortion codecs shrink SplitFed data transfer 10x","SplitFedZip: compress to cut split-federated communication cost","Learned codecs preserve accuracy while trimming SplitFed traffic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1480,"prompt_tokens":942,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":558,"tokens_out":538,"duration_ms":5859,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:22:27.268645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the data transfer for both methods from a single explicit formula that includes the number of training samples, local epochs, global epochs, and the number of transmissions per sample, using identical settings for SplitFedZip and the baseline; if the reduction ratio does not remain in (or near) the 3.2x-5000x range, the central communication-savings claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the attention-based codec with discretized Gaussian mixture likelihoods that achieves the larger bit-rate savings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the federated averaging algorithm used to aggregate both split-network weights and codec weights across clients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the SplitFed framework that SplitFedZip optimizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior SplitFed compression method whose data-transfer and accuracy numbers SplitFedZip is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Bjøntegaard Delta metric used to compute average bit-rate and accuracy differences between codec rate-accuracy curves."}],"review_version":1}