{"id":"f9516b43-8750-4a80-87ca-31a403b9c913","arxiv_id":"2412.13589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SemiDFL is a semi-supervised decentralized federated learning method that uses neighborhood pseudo-labels, consensus-based diffusion-generated data, and adaptive aggregation to handle mixed labeled and unlabeled clients.","lead":"SemiDFL is a new algorithm that lets decentralized federated learning work when some clients have few labels, some only unlabeled data, and others both. It improves image classification accuracy by combining neighbor-based pseudo-labeling, a shared diffusion data generator, and adaptive aggregation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The consensus data space is assumed, not shown: a_j in Eq. (12) is measured on each client's own generated D_j, so if the diffusion models diverge, the adaptive weights are not a shared benchmark and the central mechanism is circular.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the paper's two headline mechanisms—consensus data space and adaptive aggregation—stand or fall together, and both rely on an unproven claim that all locally generated datasets D_i share a similar distribution. I agree with that assessment. The paper does provide broad experimental support: consistent improvements over three datasets, multiple label ratios, non-IID degrees, and communication topologies, plus ablations isolating each component. Those results are real evidence that the overall recipe works empirically. But the recipe's internal logic is circular unless the distributional convergence is actually demonstrated. Eq. (12) uses accuracies computed on different local generated sets; the only justification for comparing them is the assertion that the sets are similar, which is exactly what the diffusion consensus is supposed to guarantee. Without a diagnostic of the generated distributions, the adaptive weights could be rewarding clients whose diffusion models produce easy or overfit samples, rather than clients with genuinely better classifiers. This would not necessarily invalidate the empirical gains—the method might still work through regularization or implicit ensemble effects—but it would change the interpretation from 'consensus established' to 'heuristic that works for unexplained reasons.' Because the concern is substantive but testable, and because the paper already reports ablations that could be extended, conditional acceptance with a request for this diagnostic is the right posture. The unresolved reproducibility items (code availability, unreported statistical tests) are secondary; they do not bear directly on the correctness of the central mechanism. My concrete test is designed to settle the distributional-convergence question directly and to expose whether the adaptive weights are actually evaluating on a common benchmark.","tokens_in":16746,"tokens_out":4463,"duration_ms":43930,"concrete_test":"Using the released code, run SemiDFL on CIFAR-10 (r=5%, α=0.1) and every 50 rounds compute pairwise MMD or FID between D_i and D_j for connected clients, sampling with fixed seeds. If the average pairwise distance does not shrink toward the within-client resampling noise level by round 500, the consensus data space claim fails. Then, for one sub-graph at round 500, recompute the Eq. (12) weights using a common held-out test set instead of each client's own validation subset; if the client ranking or the top-1 aggregated model changes, AdaGen (Table 4) is not equivalent to AdaTest and the adaptive aggregation results require re-interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the 'consensus data space' being real. Section 'Consensus MixUp' asserts that because diffusion models ψ_i are aggregated via Eq. (3), 'all local ψ_i converge to a unified model, thus all generated datasets D_i follow a similar data distribution.' This is the load-bearing assumption, and it is not established. Diffusion models are trained on locally pseudo-labeled data (Algorithm 1, line 11) and then parameter-averaged; no convergence proof is given, and generation is stochastic. The adaptive aggregation in Eq. (12) compounds the problem. Each client i computes its accuracy a_i on a 100-sample subset of its own generated data D_i (line 17-18). To set weights for neighbor j, client i must use a_j computed on D_j, not on a shared set. The weights therefore compare accuracies across different, possibly divergent generated distributions. These same weights are used in Algorithm 1, lines 22-23, to aggregate both classifiers and diffusion models. This creates a feedback loop: diffusion consensus determines D_i; D_i determines a_i; a_i determines the weights that aggregate the diffusion models. If the D_i do not actually converge to a common distribution, the evaluation is not a common benchmark, the weights are biased, and both pillars of the method—consensus data space and adaptive aggregation—lose their justification. The reported gains over baselines could then be an artifact of self-referential weighting rather than of a true consensus, since the baselines do not have this feedback.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SemiDFL, a semi-supervised decentralized federated learning method for networks in which clients have only labeled, only unlabeled, or mixed data. The method combines neighborhood pseudo-labeling (NPL), a consensus-based diffusion model that generates synthetic data for MixUp (C-MixUp), and an adaptive aggregation rule in which consensus weights are derived from classifier accuracy on the generated data. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 under different non-IID degrees, labeled ratios, and communication topologies report consistent improvements over adapted SSL baselines and over the paper's own ablation variants.","tokens_in":17145,"tokens_out":8527,"duration_ms":76838,"significance":"If the consensus data space is actually achieved, SemiDFL addresses a real gap: existing SSL methods for federated learning generally assume a central server, while DFL has mostly been studied under full supervision. The paper supplies a fairly extensive empirical study, including three datasets, three communication topologies, multiple non-IID degrees and labeled ratios, and three ablation studies. The reported gains over the adapted baselines are substantial in most settings, and the neighborhood pseudo-labeling ablation is particularly clean. The main risk is that the two novel mechanisms, C-MixUp and adaptive aggregation, rest on the unproven assumption that independently trained local diffusion models, after parameter averaging, generate a common data distribution. The public code URL is a useful reproducibility step, although the manuscript does not include a full code appendix. The contribution is therefore promising but needs additional evidence before the central claim can be accepted as established.","major_comments":[{"comment":"The consensus data space is load-bearing and is asserted rather than demonstrated. The text states that because the diffusion models psi_i are globally updated via Eq. (3), \"all local psi_i converge to a unified model, thus all generated datasets D_i follow a similar data distribution.\" No convergence proof is given, and the generation process is stochastic. If the generated datasets diverge, the C-MixUp training data are not drawn from a common augmented distribution, and the method's first pillar loses its justification. Please provide either a convergence argument for the diffusion-model consensus or a direct empirical measure of generated-distribution divergence across clients over training rounds, for example FID or MMD between D_i and D_j.","section":"Consensus MixUp"},{"comment":"The adaptive weights in Eq. (12) are computed from accuracies a_j measured on each client's own validation set Dhat_j, sampled from that client's generated data. This is not evaluation on a shared dataset unless the consensus data space assumed in the previous comment has already been established. If the generated distributions differ across clients, a high a_j can simply mean that client j's diffusion model produces easy samples, so the softmax weighting rewards generator easiness rather than classifier quality. The same weights are then used to aggregate the diffusion models themselves (Algorithm 1, line 23), creating a feedback loop: generated distributions determine accuracies, accuracies determine aggregation weights, and those weights determine the next generation of generated distributions. A cross-evaluation ablation, in which each client evaluates all neighbor classifiers on a fixed generated reference set, is needed to show that the reported gains are not an artifact of this self-referential design.","section":"Adaptive Aggregation, Eq. (12); Algorithm 1, lines 17-23"},{"comment":"The paper states that MixMatch, FlexMatch, and CBAFed were adapted to the DFL setting, but the adaptation is not described. These methods were designed for centralized or server-based training, so the validity of the comparison depends on how their aggregation, pseudo-labeling, and augmentation procedures were modified. Please provide the adaptation protocol or the code used for each baseline. In addition, the Reproducibility Checklist claims that statistical tests were used to judge significance, but the paper reports only means and standard deviations and no statistical test anywhere; this discrepancy should be corrected.","section":"Baseline methods / Experiments"}],"minor_comments":[{"comment":"The MixUp formula for x' has x_m on both sides, so the mixed input has no component from x_n; it should presumably read x' = lambda*x_m + (1-lambda)*x_n.","section":"Eq. (10)"},{"comment":"The denominator max_i(max_c(sigma_i^{t,c})) can be zero when no pseudo-label exceeds the threshold tau; the update rule should specify a fallback, such as keeping the threshold unchanged.","section":"Eq. (9)"},{"comment":"The set D_t^i is used at line 15 before it is generated for rounds t < R; the pseudocode should state that D_t^i is empty during the diffusion warm-up rounds.","section":"Algorithm 1"},{"comment":"The last setting is listed as alpha = 0.1, r = 1%, which breaks the pattern of the other tables; if this is a typo for r = 0.1%, it should be corrected.","section":"Table 4"},{"comment":"The abstract refers to the method as \"DFL-Semi\" while the rest of the paper uses \"SemiDFL\"; please unify the terminology.","section":"Abstract"},{"comment":"The checklist says that source code is included in a code appendix, but no code appendix appears in the manuscript; the GitHub URL in the abstract and the \"upon acceptance\" statement in the supplementary material should be reconciled.","section":"Reproducibility Checklist"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical contribution to an understudied problem, and I do not see grounds for rejection. However, the central mechanism, consensus in both data and model spaces, is currently supported only by an assertion about diffusion-model convergence, and the adaptive aggregation rule is self-referential in a way that could explain part of the reported gain. The requested convergence evidence or divergence measurements and a cross-evaluation ablation are, in my view, necessary before the paper is ready for publication. The checklist/test discrepancy and the missing baseline adaptation details are secondary but should also be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on SemiDFL (arXiv:2412.13589). It is a reasonable, well-packaged first crack at semi-supervised decentralized federated learning, and the experiments are more thorough than most. The core idea—neighborhood pseudo-labeling, a consensus diffusion model for generating data, then adaptive aggregation weighted by accuracy on that generated data—hangs together and outperforms adapted SSL baselines on MNIST, Fashion-MNIST, and CIFAR-10 across label ratios, non-IID degrees, and topologies. The ablations are informative, especially the comparison of AdaGen (weights on generated data) against AdaTest (weights on a real test set), which shows only a small gap. That is the right way to test the adaptive aggregation.\n\nWhere the paper is softest is the 'consensus data space' claim. The paper asserts that because the diffusion models are parameter-averaged, all local generated datasets share a similar distribution, and uses that as justification for evaluating classifier accuracy on each client's own 100 generated samples. No convergence analysis is given, and the aggregation weights in Eq. (12) depend on those very evaluations, so if the diffusion models do not, in practice, converge to a common distribution, the benchmark is not shared and the weighting is self-referential. This is not fatal on the evidence: the AdaGen vs AdaTest ablation suggests the mechanism works, and standard consensus arguments probably close the gap. But the paper should either prove convergence under the adaptive weights or add a diagnostic showing the generated distributions are actually similar across clients.\n\nOther, smaller issues: statistical significance tests promised in the reproducibility checklist are not reported in the text. The GitHub link in the abstract and the 'code upon acceptance' in the supplementary are inconsistent. A few hyperparameters (K, warm-up round R) are not specified. The adapted baselines are described only briefly, so it is hard to tell if they got fair tuning.\n\nWho is this for? Researchers working on federated learning in edge settings where labels are scarce. The problem setting is real, the method is novel in combination, and the experiments are systematic. It deserves a serious referee; I would send it to review, with the expectation that the authors add a convergence or divergence analysis for the diffusion consensus and clean up the reproducibility details.","headline":"SemiDFL is a credible first semi-supervised DFL method with solid experiments; the consensus data-space assumption needs formal support, but the paper deserves review.","tokens_in":17567,"tokens_out":2887,"would_cite":false,"duration_ms":26176,"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":"SemiDFL claims that decentralized federated learning can train accurate classifiers from mostly unlabeled, non-IID client data by coupling neighborhood pseudo-labeling with a consensus-trained diffusion model.","keywords":["decentralized federated learning","semi-supervised learning","pseudo-labeling","diffusion models","consensus aggregation","non-IID data","MixUp","adaptive aggregation"],"falsifier":"Take a trained SemiDFL system on non-IID CIFAR-10 and measure a distribution distance (e.g., maximum mean discrepancy or class-conditional FID) between the synthetic datasets $D_i$ of neighboring clients; the consensus data-space claim predicts small distances, and if large distances appear, the MixUp and adaptive-aggregation components are not actually sharing a common data distribution.","tokens_in":16514,"feed_emoji":"🔗","tokens_out":10896,"duration_ms":92477,"temperature":0.7,"pith_summary":"The paper tackles a practical gap in decentralized federated learning: standard DFL assumes every client has labels, but real deployments often have clients with a few labels, only unlabeled data, or a mix. It proposes SemiDFL, a semi-supervised DFL method whose core claim is that a network of clients can train accurate classifiers from very few labels and strongly non-identical data distributions by building consensus in two spaces at once: a data space of synthetic samples and a model space of classifier and generator weights. The method couples neighborhood-aware pseudo-labeling with a consensus-trained diffusion model that generates synthetic data for MixUp, then adaptively aggregates models based on accuracy on that synthetic data. If the claim holds, decentralized training becomes usable in settings where labeling is expensive and no central server can orchestrate the learning.","feed_headline":"SemiDFL beats SSL baselines with almost no labels","feed_subtitle":"Neighborhood pseudo-labels plus consensus diffusion close the gap to fully supervised DFL in non-IID settings.","key_machinery":"The load-bearing object is the consensus-trained local diffusion model $\\psi_i$ together with the adaptive aggregation rule in Eq. (12). A diffusion model is a generative model that learns to remove noise and can sample new images conditioned on a class; here each client trains one on its labeled and pseudo-labeled data, then all $\\psi_i$ are updated through the consensus rule in Algorithm 1 step 23. If those generators converge to a shared distribution, the synthetic sets $D_i$ form a common data space for MixUp and for evaluating classifiers, and the softmax weights $w_{ij}$ computed from accuracies on 100-sample subsets of $D_i$ become a privacy-preserving stand-in for a shared test set.","core_discovery":"On the paper's own terms, SemiDFL's discovery is that the semi-supervised DFL problem is solvable by enforcing consensus in both data and model spaces. Each client uses neighborhood classifiers and an adaptive class-wise threshold to produce high-quality pseudo-labels; trains a local diffusion model on labeled plus pseudo-labeled data; and the local diffusion models are aggregated by the same consensus rule as classifiers, so that every client can generate synthetic data with a similar distribution. MixUp over labeled, pseudo-labeled, and synthetic data creates a consensus data space for classifier training, and adaptive aggregation weights—computed from each classifier's accuracy on a small sample of that synthetic data—create a consensus model space. The experiments report that SemiDFL consistently outperforms MixMatch, FlexMatch, and CBAFed, as well as the DFL lower bound, across MNIST, Fashion-MNIST, and CIFAR-10 under IID and non-IID settings and multiple labeled-data ratios, with only the fully supervised DFL upper bound scoring higher.","pith_inferences":["Beyond the paper, the consensus synthetic data could be reused as a common validation benchmark for other decentralized tasks—client selection, drift detection, or Byzantine-robust aggregation—since SemiDFL currently uses it only for MixUp and for aggregation weights.","The ablation compares a consensus GAN with the consensus diffusion generator; if newer generative models improve the quality and homogeneity of the synthetic sets, the same SemiDFL scaffolding should widen its lead on harder image datasets.","A direct test of the method's internal logic would be to see whether per-client generated distributions actually become similar after consensus; the paper reports accuracy gains but does not measure the distance between $D_i$ sets."],"forward_implications":["A DFL network can continue learning when some clients have no labeled data at all, rather than being limited to clients that happen to have labels.","No shared or public test set is needed to decide which neighbors should have more influence in aggregation; the generated data doubles as the evaluation set.","The method's reported gains persist across three communication topologies, so the mechanism is not tied to one graph structure.","In the low-label regimes the paper tests, several SSL baselines fail to converge, while SemiDFL keeps improving, suggesting the consensus mechanisms also stabilize training."],"supporting_citations":[{"why":"Provides the pseudo-labeling baseline that SemiDFL extends with neighborhood classifiers and adaptive thresholds.","marker":"Lee et al. 2013"},{"why":"Supplies the label-sharpening step and the MixUp operation used for pseudo-labels and synthetic data.","marker":"Berthelot et al. 2019"},{"why":"Provides the class-balanced adaptive threshold idea and the CBAFed baseline that SemiDFL compares against.","marker":"Li, Li, and Wang 2023"},{"why":"Defines the denoising diffusion objective used to train each client's generator.","marker":"Ho, Jain, and Abbeel 2020"},{"why":"Supplies the deterministic sampling procedure used to create synthetic datasets.","marker":"Song, Meng, and Ermon 2020"},{"why":"Gives the classifier-free guidance used for class-conditioned generation.","marker":"Ho and Salimans 2022"},{"why":"Provides the consensus aggregation rule applied to classifiers and diffusion models.","marker":"Liu and Ding 2021b"},{"why":"Defines the non-IID data partition used to construct the experimental settings.","marker":"Hsu, Qi, and Brown 2019"}],"fun_headline_variants":["SemiDFL: Semi-supervised DFL without a central server","SemiDFL: Beats semi-supervised baselines in decentralized FL","No labels? SemiDFL makes decentralized FL work","SemiDFL: Consensus data and model spaces improve DFL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the belief that consensus aggregation of the local diffusion models makes every client's generated data statistically similar, even though the aggregation weights used in that consensus are themselves computed from each classifier's accuracy on those very same generated sets; if the generators do not converge to a shared distribution, the consensus data space and the adaptive weights built on it have no common reference.","fun_headline_variants_meta":{"raw":{"variants":["SemiDFL: Semi-supervised DFL without a central server","SemiDFL: Beats semi-supervised baselines in decentralized FL","No labels? SemiDFL makes decentralized FL work","SemiDFL: Consensus data and model spaces improve DFL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3719,"prompt_tokens":1006,"completion_tokens":2713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2638}},"tokens_in":622,"tokens_out":2713,"duration_ms":19715,"temperature":1.0,"reasoning_tokens":2638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:58:09.311370+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained SemiDFL system on non-IID CIFAR-10 and measure a distribution distance (e.g., maximum mean discrepancy or class-conditional FID) between the synthetic datasets $D_i$ of neighboring clients; the consensus data-space claim predicts small distances, and if large distances appear, the MixUp and adaptive-aggregation components are not actually sharing a common data distribution.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pseudo-labeling baseline that SemiDFL extends with neighborhood classifiers and adaptive thresholds."}],"review_version":1}