{"id":"571f3e18-efde-4848-9a88-1bc899a9b63d","arxiv_id":"2501.02219","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In federated learning with very few labels, a collaboratively trained diffusion model can synthesize balanced examples for missing classes and lift accuracy by about 14 points on CIFAR-10.","lead":"This paper introduces a method for federated semi-supervised learning that trains a shared diffusion model to generate synthetic training images, filling gaps when some clients have few labeled examples and skewed class distributions. It reports large accuracy gains on CIFAR-10 and Fashion-MNIST, but only against simple baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claim of superiority over existing FSSL methods is unsupported: Table I compares only against FedAvg (non-SSL) and FedAvg-SL (upper bound), not against any prior FSSL algorithm, and the diffusion-synthesis component is not isolated.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper has substantial gaps. The reader's weakest_assumption focuses on the undefined local test set used for confusion-matrix estimation (Section III-B); that is a real issue because the method becomes unexecutable as written if no labeled validation set exists. However, I consider the missing comparison to existing FSSL methods and the lack of a no-synthesis control to be even more load-bearing for the paper's central claim, because the claim explicitly says 'compared to existing methods' and attributes gains to diffusion-based data synthesis. If the gains come from pseudo-labeling alone, the specific contribution is not validated. My concrete test would settle this by adding the necessary baselines. I do not recommend changing the verdict because the method might still be viable after such experiments; a conditional accept with required additions remains appropriate. This is why I choose UNCHANGED rather than REJECT or ACCEPT.","tokens_in":9012,"tokens_out":6881,"duration_ms":67862,"concrete_test":"Run the identical CIFAR-10 λ=0.1 (DIR,DIR) configuration with SemiFed, FedMatch, and FedDure as baselines under the same training schedule, and include a pseudo-label-only control (train classifier on labeled plus pseudo-labeled data without generating synthetic data). If DDSA-FSSL fails to exceed the best FSSL baseline and the control by a statistically significant margin (e.g., >1% accuracy), the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract, Section I) is that DDSA-FSSL 'significantly enhances classification accuracy compared to existing methods.' However, the experimental section (Table I) evaluates only two baselines: FedAvg, which uses no unlabeled data and is not an FSSL method, and FedAvg-SL, a fully supervised upper bound. No existing FSSL algorithm (SemiFed, FedMatch, FedDure, or SDA-FL) is compared, so the claimed superiority over 'existing methods' is not demonstrated. Equally important, the 'without/with data selection' ablation still includes the diffusion model in both arms; there is no control that uses the same pseudo-labeling but omits synthetic data generation. Therefore, the observed gains over FedAvg could arise entirely from standard pseudo-label-based semi-supervised learning, with the diffusion model contributing little. This concern is load-bearing because the paper's title and contribution are specifically about diffusion-model-based data synthesis; without an FSSL baseline or a no-synthesis control, the improvement cannot be attributed to the proposed mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DDSA-FSSL, a federated semi-supervised learning framework that trains a latent conditional diffusion model collaboratively across clients, using pseudo-labels assigned by a federated classifier. A precision-optimized data-selection step filters pseudo-labeled samples based on a global confusion matrix estimated from local test sets. Clients then generate synthetic data for classes that are underrepresented or absent locally, aiming to align local distributions with the global distribution. Experiments on CIFAR-10 and Fashion-MNIST under IID and Dirichlet-based non-IID settings report accuracy gains over FedAvg and a fully supervised upper bound (FedAvg-SL), with further gains when the precision-based data selection is enabled.","tokens_in":9204,"tokens_out":2722,"duration_ms":25634,"significance":"If the results are substantiated, the work addresses an important practical problem—combining label scarcity and data heterogeneity in federated learning—by showing that a collaboratively trained diffusion model can augment local data with class-conditional synthetic samples without relying on pre-trained generative models. The idea of using a global confusion matrix to optimize pseudo-label precision is a reasonable contribution, and the paper covers two datasets and several heterogeneity settings. However, the current empirical evidence is not yet sufficient to support the central claim of superiority over existing FSSL methods: the evaluation lacks comparisons with established FSSL baselines and lacks a control that isolates the contribution of diffusion-based synthesis from ordinary pseudo-labeling. The paper also does not report variance across runs, which is necessary given the small margins seen in several configurations.","major_comments":[{"comment":"The description of the confusion-matrix calculation is incomplete and load-bearing for the data-selection step. The text states that each client generates M_t_k by applying the global classifier to their local test set, but the system model in Section II defines only labeled and unlabeled training data, with no held-out test set. If the confusion matrix is computed on the labeled training set instead, the precision estimates will be optimistically biased because the labels are known. If no test set exists, the method cannot be executed as described. Please clarify where the test set comes from and, if it is the labeled training set, analyze the bias and its effect on the reported gains.","section":"Section III-B"},{"comment":"The central claim that DDSA-FSSL 'significantly enhances classification accuracy compared to existing methods' is not supported by the experiments, because the only baselines are FedAvg, which does not use unlabeled data and is not an FSSL method, and FedAvg-SL, which is a fully supervised upper bound rather than an existing FSSL competitor. The paper does not compare against established FSSL algorithms such as SemiFed, FedMatch, FedDure, SDA-FL, or FedDISC, even though several are cited in the introduction. Please add comparisons with at least two or three of these methods under the same experimental settings, or revise the claim to state that DDSA-FSSL improves over FedAvg and approaches the supervised upper bound.","section":"Section IV-B, Table I"},{"comment":"The 'without/with data selection' ablation does not isolate the contribution of the diffusion model. Both arms include pseudo-labeling and synthetic data generation; the ablation only removes the precision-based filtering. There is no control that uses the same pseudo-labeling and classifier training but omits the diffusion-generated synthetic data (e.g., training only on labeled plus pseudo-labeled real data, or using a simpler augmentation). Without such a control, the observed gains over FedAvg could be entirely due to standard pseudo-label-based semi-supervised learning, and the paper cannot support the claim that diffusion-based data synthesis is the key mechanism. Please add a no-synthesis ablation.","section":"Section IV-B, Table I"},{"comment":"The experimental results are reported without standard deviations, confidence intervals, or multiple random seeds. This is a concern because several improvements are small—for example, Fashion-MNIST with α=0.2 gives 87.60% vs. 87.69% (without/with selection) under (IID, DIR), and the difference between 87.23% and 87.30% under (DIR, DIR) is under one percentage point. Without variance estimates, it is not possible to judge whether these differences are significant. Please report the mean and standard deviation over at least three independent runs, or otherwise justify the single-run results.","section":"Section IV-A and Table I"}],"minor_comments":[{"comment":"The abstract states that DDSA-FSSL improves CIFAR-10 accuracy from 38.46% to 52.14%, while Section I states 47.72% with 10% synthetic data and 53.01% with 90% synthetic data, and Table I reports 53.01% at α=1.0. These numbers should be reconciled.","section":"Section I and Abstract"},{"comment":"A typo appears in the phrase 'only on labled data'; it should be 'labeled data'.","section":"Section IV-B"},{"comment":"Reference [12] is a duplicate of reference [9]; both cite the same AAAI 2024 paper by Yang et al. on one-shot semi-supervised federated learning with pre-trained diffusion models. Remove the duplicate.","section":"References"},{"comment":"The hyperparameters w_L1, w_p, and τ in Eq. (7) are introduced but their values are not reported in the experimental setup. Please provide the values used and, ideally, a sensitivity study, as these weights directly control the trade-off between pseudo-label quantity and precision.","section":"Section III-B"},{"comment":"The description of Figure 3 is vague: it says precision and recall variations across classes are examined, but it does not explain how the curves are computed, which classes correspond to which x-axis positions, or what the error bars (if any) represent. Please clarify the figure description.","section":"Section IV-B, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the proposed framework is plausible, but the experimental evaluation is currently too weak to support the headline claim. The missing FSSL baselines and the absence of a no-synthesis control are the main blockers. I also note that the confusion-matrix test-set assumption is a methodological gap that could affect the validity of the data-selection step. If the authors add the missing comparisons and ablations, and clarify the test-set issue, the paper could become a solid contribution. I do not see an inherent flaw that would justify rejection, but the current manuscript is not ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible new combination — a federated-trained latent diffusion model feeding class-balanced synthetic data back into FSSL, with a precision-filtering step on pseudo-labels. The framework is clearly presented and the ablation on augmentation strength behaves as expected. But the evidence section doesn't currently support the abstract's claim of beating existing methods, and there's a load-bearing gap in how the confusion matrix is obtained.\n\nWhat's actually new: prior work either uses pre-trained DMs (FedDISC) or GANs (SDA-FL); here the DM is collaboratively trained from scratch under label scarcity. The precision-optimized pseudo-label selection is a reasonable heuristic, and the idea of using the global confusion matrix to filter pseudo-labels is neat, even if it's a cleaning step rather than a proven optimality argument. The paper also gives a complete algorithmic specification and reports results over multiple heterogeneity settings, with the expected monotone improvement as α grows.\n\nThe problems are empirical and one definitional. The big one: Table I compares only against FedAvg (which ignores unlabeled data) and FedAvg-SL (a fully supervised upper bound). No established FSSL method — SemiFed, FedMatch, FedDure, SDA-FL, FedDISC — appears as a baseline. And the 'without/with data selection' ablation still uses the DM in both arms, so it never isolates the contribution of synthetic data generation. With only these controls, the gains over FedAvg could come from vanilla pseudo-labeling. That's load-bearing for the title's claim.\n\nSecond, the confusion-matrix estimation in Section III-B requires each client to apply the global classifier to a 'local test set' with true labels, but the system model defines only labeled and unlabeled training sets. If that test set is actually the labeled training set, precision is optimistically biased; if it doesn't exist, the method can't be run as described. Either way, the filtering step is under-specified and the hyperparameters (wL1, wp, τ) aren't reported.\n\nThird, the numbers don't line up: the abstract says 52.14% for CIFAR-10 λ=0.1, while the intro and Table I say 53.01% at α=10.1 (and 47.72% at α=0.2). There are also no standard deviations or multiple runs, so I can't tell which of these differences are meaningful. Releasing code would help a lot.\n\nWho this is for: researchers working on FSSL or generative data augmentation in FL who want a concrete instantiation of the DM-based idea. It's a worthwhile method paper, but the referee should push for real baselines and uncertainty numbers. I'd accept it for peer review and require major revision, not desk reject.","headline":"The method is a real novelty, but the experiments don't yet show it beats FSSL baselines; the confusion-matrix step needs a defined test set.","tokens_in":9791,"tokens_out":2594,"would_cite":false,"duration_ms":23716,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that federated semi-supervised learning can overcome label scarcity and non-IID data by collaboratively training a latent diffusion model on precision-filtered pseudo-labels, letting each client synthesize samples for its…","keywords":["federated learning","semi-supervised learning","diffusion models","data synthesis","non-IID data","pseudo-labeling","latent diffusion","precision-optimized selection"],"falsifier":"Run DDSA-FSSL on a benchmark where the true labels of the 'unlabeled' partition are known, and compare the confusion-matrix-estimated precision of the selected pseudo-labeled samples with their actual precision; also replace the diffusion-generated synthetic samples with the same number of selected real pseudo-labeled samples. If the estimated precision is systematically optimistic, or if the real-sample ablation matches DDSA-FSSL's accuracy, the claimed gain is not caused by diffusion synthesis.","tokens_in":8778,"feed_emoji":"📈","tokens_out":11008,"duration_ms":95660,"temperature":0.7,"pith_summary":"Federated semi-supervised learning (FSSL) faces two problems at once: clients have very few labels, and their data distributions are uneven and non-IID. This paper proposes to attack both with a collaboratively trained conditional latent diffusion model that generates synthetic training data. In DDSA-FSSL, a federated classifier pseudo-labels unlabeled data, a precision-optimization step filters those pseudo-labels using a global confusion matrix, and a shared latent diffusion model is trained on the filtered data; each client then synthesizes samples for the classes its local dataset lacks, rebalancing local data toward the global distribution. The authors report that on CIFAR-10 with 10% labeled data under dual heterogeneity, accuracy rises from 38.46% to 53.01% as synthetic data is added, approaching the fully supervised upper bound. If true, this points to a way to exploit abundant unlabeled data and generative synthesis without sharing raw client data.","feed_headline":"Diffusion synthesis lifts federated semi-supervised accuracy to 53%","feed_subtitle":"A shared diffusion model trained on filtered pseudo-labels lets clients synthesize missing classes and beat fully supervised baselines.","key_machinery":"The load-bearing object is a class-conditional latent diffusion model (c-LDM): a VAE whose latent space is shared across clients, together with a U-Net denoising network that uses cross-attention on class labels so a client can request samples for a specific class. It is trained in two federated stages: first the VAE on all local data, then the denoiser on encoded labeled and precision-filtered pseudo-labeled data. The companion mechanism is precision-optimized data selection: each client estimates the global confusion matrix $M^t_g$ by applying the classifier to its local test set, scales its columns by the counts of pseudo-labeled samples per class, and solves the constrained problem P1, which maximizes the average diagonal precision under L1-sparsity and proportion penalties, using SLSQP. The optimal per-class retention rates $\rho^\\star_k$ decide how many pseudo-labels are kept, and the augmentation strength $\\alpha$ then fixes how many synthetic samples per class each client generates so the augmented local dataset matches the global class distribution.","core_discovery":"The central claim is that a globally shared conditional latent diffusion model (c-LDM), trained federatedly on labeled data plus precision-filtered pseudo-labeled data, can supply each client with synthetic samples for the classes its local dataset lacks. Because the diffusion model is trained across clients in a shared latent space rather than downloaded pre-trained, it can capture the global data distribution and avoid the domain mismatch of pretrained generators. The precision-optimized data selection step estimates per-class label reliability from an aggregated confusion matrix and solves a constrained optimization problem to remove the pseudo-labeled samples most likely to be mislabeled before diffusion training, so the generator learns from cleaner supervision. The paper reports that with 10% labeled CIFAR-10 data under dual heterogeneity, DDSA-FSSL raises classification accuracy from 38.46% to 53.01% when enough synthetic data is added, approaching the fully supervised FedAvg-SL upper bound.","pith_inferences":["The paper's system model never defines the local test set used to compute the confusion matrix; if no such labeled held-out set exists, the precision estimates are likely optimistic and the filtering step needs a privacy-preserving calibration or server-side validation set.","Because the baselines are FedAvg and FedAvg-SL rather than an FSSL method with pseudo-labeling but no diffusion synthesis, the reported gains do not isolate the contribution of the generative model; an ablation that reuses the selected pseudo-labels directly in classifier training would separate these effects.","The observed class-level pattern, where recall improves most for classes with high initial precision and can decline for low-precision classes, suggests a class-conditional precision threshold or targeted retraining for low-precision classes could extend the method.","The precision-optimized pseudo-label filter is stated generally and could be transferred to other generative models or non-federated semi-supervised pipelines, though the paper only tests it inside DDSA-FSSL."],"forward_implications":["If the central claim holds, FSSL clients can compensate for missing classes without uploading raw data or relying on pre-trained generative models, because the diffusion model itself is trained collaboratively.","On CIFAR-10 with 10% labels under dual heterogeneity, accuracy climbs from 47.72% at augmentation strength $\\alpha=0.2$ to 57.48% at $\\alpha=10.1$, showing that more synthetic data monotonically closes the gap to the fully supervised upper bound.","The precision-optimized selection step adds consistent gains, and its largest effect appears under dual heterogeneity (3.41 percentage points at $\\alpha=0.2$ on CIFAR-10).","With sufficient synthetic data, DDSA-FSSL surpasses fully supervised FedAvg-SL at $\\lambda=0.7$ and $\\lambda=0.9$ under dual heterogeneity, indicating that rebalancing can outweigh the imperfection of generated samples.","The framework does not depend on a specific aggregator; FedAvg can be replaced by other federated aggregation algorithms tailored to FSSL."],"supporting_citations":[{"why":"Supplies FedAvg, the aggregation algorithm used to train the classifier, VAE, and denoiser across clients.","marker":"[13]"},{"why":"Supplies the latent-diffusion architecture, including the shared VAE latent space and class-conditioning cross-attention, that the c-LDM uses.","marker":"[15]"},{"why":"Provides the forward-noise and reverse-denoising formulation with the noise schedule used by the conditional diffusion model.","marker":"[18]"},{"why":"Supplies the VAE training objective with GAN loss used to learn the shared latent space.","marker":"[17]"},{"why":"The GAN-based synthetic-data FL baseline whose limitations motivate training a conditional diffusion model instead.","marker":"[7]"},{"why":"The pre-trained diffusion model baseline whose domain-mismatch limitations motivate federated training of the generator.","marker":"[9]"}],"fun_headline_variants":["Diffusion model synthesizes missing classes, lifting federated accuracy to 53%","Diffusion-aided FSSL jumps accuracy from 38% to 53% on CIFAR-10","Shared diffusion model creates missing-class samples for clients","Diffusion synthesis lifts semi-supervised federated accuracy by 14%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes every client has a held-out local test set with true labels for computing the confusion matrix $M^t_g$, although the system model in Section II defines only labeled and unlabeled training data.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model synthesizes missing classes, lifting federated accuracy to 53%","Diffusion-aided FSSL jumps accuracy from 38% to 53% on CIFAR-10","Shared diffusion model creates missing-class samples for clients","Diffusion synthesis lifts semi-supervised federated accuracy by 14%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2844,"prompt_tokens":931,"completion_tokens":1913,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1829}},"tokens_in":547,"tokens_out":1913,"duration_ms":14243,"temperature":1.0,"reasoning_tokens":1829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:14:00.729178+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DDSA-FSSL on a benchmark where the true labels of the 'unlabeled' partition are known, and compare the confusion-matrix-estimated precision of the selected pseudo-labeled samples with their actual precision; also replace the diffusion-generated synthetic samples with the same number of selected real pseudo-labeled samples. If the estimated precision is systematically optimistic, or if the real-sample ablation matches DDSA-FSSL's accuracy, the claimed gain is not caused by diffusion synthesis.","supporting_citations":[{"cited_title":"Autoencoding beyond pixels using a learned similarity metric,","cited_arxiv_id":null,"evidence_quote":"Supplies the VAE training objective with GAN loss used to learn the shared latent space."}],"review_version":1}