{"id":"cbb73865-77d1-48a6-8f13-69780ce46673","arxiv_id":"2412.18460","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Generative model-aided federated learning (GeFL) enables model-heterogeneous FL by sharing a federated generator, and its feature-level version GeFL-F improves scalability and privacy.","lead":"A new federated learning method lets clients with different model architectures share knowledge by training a shared generative model and using its synthetic samples as extra training data. A feature-level variant (GeFL-F) improves privacy and scales to more clients, but it requires clients to share a common feature extractor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: GeFL's gains may be pure synthetic-data augmentation; no experiment isolates federated generator aggregation, and the 10-client FedAvg baseline is local-only training.","rationale":"The reader's verdict already flags the weak 10-client baseline and the shared-feature-extractor limitation in GeFL-F, but the most load-bearing issue is more specific: the paper never controls for the possibility that the performance gain comes from synthetic-data augmentation alone rather than from federated aggregation of the generative model. The central claim in Section III-A depends on cross-client knowledge transfer, and that claim is untested without a local-generator baseline. The concrete test above would settle the question directly. Because the existing experiments are consistent with the paper's conclusion but do not yet rule out the augmentation-only explanation, the verdict remains CONDITIONAL: the paper should be accepted only after this control is added and the central claim is re-evaluated against it.","tokens_in":34334,"tokens_out":3786,"duration_ms":40443,"concrete_test":"Run the 10-client MNIST/FMNIST protocol with a LocalGen control: each client trains the same DCGAN/VAE/DDPM architecture only on its own D_k for the equivalent number of local epochs, with no parameter averaging; target networks then train with the same Ts/Tr/loss schedule on real plus LocalGen synthetic samples. Compare mean accuracy to GeFL in Table II. Also run the same control at 50 and 100 clients to test the scalability claim in Figure 4. If LocalGen matches GeFL, the federated aggregation is not load-bearing; if GeFL is significantly better, the central mechanism is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Algorithm 1 couples two mechanisms: federated aggregation of the generator (lines 4-10) and target-network training on synthetic samples (lines 14-22). The central claim in Section III-A — that the federated generator 'gathers global knowledge from every client' — requires showing that the aggregation mechanism causes the accuracy gain. No experiment isolates this. The 10-client FedAvg baseline in Table II is explicitly 'each client uses a unique model and trains independently without FL' (Section III-B-a), so the comparison already bundles a synthetic-augmentation effect into GeFL. Table IV compares against DA on real images, and Table VII compares Syn vs. Real+Syn using the same federated generator; neither includes a generator trained only on one client's data. If the client split is the evenly divided partition described in footnote 5, a locally trained generator could plausibly reproduce the augmentation benefit, making the federated aggregation unnecessary. GeFL-F inherits the same gap: Algorithm 3 stages (ii)-(iii) train the header on synthetic features, but there is no control where each client trains a feature-generator on its own features.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeFL and GeFL-F, two frameworks for model-heterogeneous federated learning. GeFL trains a conditional generative model in a federated manner on client data and uses its synthetic samples to augment the local training of heterogeneous target networks, without requiring shared model architectures or public data. GeFL-F extends this idea by training feature-level generative models on outputs of a shared feature extractor, claiming improved scalability, lower communication/computation cost, and reduced privacy risk. The empirical evaluation spans MNIST, Fashion-MNIST, CIFAR-10, SVHN, and CelebA with three generative model families (GAN, VAE, DDPM), reporting accuracy gains over several FL baselines, together with analyses of memorization, data scale, guidance score, and homogeneity level.","tokens_in":34529,"tokens_out":6386,"duration_ms":58250,"significance":"If the central claim holds, GeFL provides a practical method for knowledge sharing across clients with fully heterogeneous model architectures without public data, which is an important open problem in federated learning. The paper ships code, reports confidence intervals, and evaluates three generative families across four datasets, which is a solid empirical contribution. The paper is also honest in disclosing limitations of GeFL (memorization, scalability, resource cost) and motivates GeFL-F. However, the key attribution of the gains to federated aggregation of the generator is not experimentally isolated from the synthetic-augmentation effect, and the privacy claim for GeFL-F rests on a metric whose interpretation is questionable in the feature-reconstruction setting. These issues need to be addressed before the central claims can be accepted.","major_comments":[{"comment":"The experimental design does not isolate federated generator aggregation from synthetic-data augmentation. In Table II, the 10-client FedAvg baseline is explicitly described as 'each client uses a unique model and trains independently without FL' (Section III-B-a), so GeFL's gains over FedAvg bundle the effect of augmentation with synthetic samples and the effect of federated aggregation of the generator. Table IV compares GeFL only against FedAvg with standard DA methods, and Table VII compares Syn vs. Real+Syn using the same federated FedDCGAN; neither includes a control where each client trains a generator only on its own local data and uses it for the same augmentation scheme. Since Algorithm 1 couples generator aggregation (lines 4-10) with target-network training on synthetic samples (lines 14-22), the central claim in Section III-A that the federated generator 'gathers global knowledge from every client' is not directly supported by the current experiments. Please add a control with locally trained generators (one per client, trained on each client's partition) and compare the resulting accuracy against GeFL; if the local-generator control matches GeFL's performance, the central contribution would need to be reframed as synthetic-augmentation rather than federated knowledge sharing.","section":"Section III-B-b, Table II and Algorithm 1"},{"comment":"The abstract and introduction claim 'model-agnostic federated learning' for GeFL-F, but Algorithm 3 stage (i) requires a common feature extractor that is aggregated across all clients, and Section IV-A states that each target network consists of 'a common feature extractor and a unique heterogeneous header.' Thus GeFL-F supports heterogeneity only in the classification header, not in the full model architecture. The statement in the introduction that the framework enables 'FL with heterogeneous client models' is therefore overstated for GeFL-F. Please qualify the model-agnosticism claim for GeFL-F, and clarify in the discussion of Figure 7 that homogeneity level 0 (where no common feature extractor exists) falls outside GeFL-F's assumptions and corresponds to GeFL rather than GeFL-F.","section":"Section IV-A and Algorithm 3"},{"comment":"The privacy claim for GeFL-F relies on MND ratios computed between original images and images reconstructed from generated features via model inversion, as described in Section IV-B-b and Figure 5. Because model inversion of low-dimensional features typically yields blurry or unrecognizable images, the low MND values for GeFL-F (0.101-0.120) may reflect reconstruction error rather than absence of memorization or membership-inference risk. Moreover, Eq. (1) defines MND between a training sample and sets of synthetic and validation images; applying this ratio to reconstructed images changes its meaning, and the paper does not specify how the sets S and V are constructed in the feature case. Please validate the privacy claim with a direct membership-inference attack on the feature generator or with MND computed in feature space, and state explicitly what privacy property (if any) the feature-level approach provides.","section":"Section IV-B-b, Table VI"}],"minor_comments":[{"comment":"There is a typo in 'experminets' (should be 'experiments'), and the paper inconsistently uses both 'GeFL' and 'G EFL' throughout; please standardize the notation.","section":"Section V-A"},{"comment":"The FedALA row shows '80.03±0. 51.63±7.05', which appears to be a formatting or OCR error; the confidence interval entry is malformed and should be corrected.","section":"Table II"},{"comment":"The sentence 'the accuracy rapidly decreases thereafter due the significant heterogeneity' is missing the word 'to' after 'due'.","section":"Section III-B-a"},{"comment":"The notation 'min x ∈ S' should be typeset as 'min_{x \\in S}' with proper spacing, and the sets S and V should be defined immediately before the equation rather than only in the surrounding text.","section":"Equation (1)"},{"comment":"The caption 'GEFL-F shows the same trend across the homogeneity regardless of the number of clients' is grammatically unclear; consider rewording to 'GEFL-F exhibits the same trend across homogeneity levels for each number of clients'.","section":"Figure 7 caption"},{"comment":"Reference [1] is given only as a URL; since it serves as the code repository, please format it as a proper code/repository entry (e.g., with an accessed date and repository name) rather than leaving it as a bare arXiv-style hyperlink.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an empirical systems paper with a substantial amount of experimentation and a useful honest discussion of limitations. The main technical concern is the missing local-generator control, which is directly load-bearing for the central claim that federated aggregation of the generator enables knowledge sharing; this is fixable with an additional experiment. The privacy analysis for GeFL-F also needs stronger validation than the MND-on-reconstructed-images metric. If the authors add the control, temper the model-agnosticism claim for GeFL-F, and strengthen the privacy evaluation, the paper would be within the scope of the journal and could be acceptable. I did not find evidence of a fundamental flaw in the proposed mechanisms themselves."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead the GeFL paper. The core idea is clean: train a generative model federated across clients, then use it to synthesize samples that let heterogeneous target networks learn from each other's data without sharing models or raw data. The feature-level variant GeFL-F does the same on features from a shared extractor, which cuts communication and memorization. That combination is new as far as I know, and the experimental coverage is genuinely thorough: four datasets, GAN/VAE/DDPM families, client counts up to 100, plus resource tables and an honest limitations section. The code is out. The MND privacy numbers for GeFL-F are strikingly better than the raw-data variant.\n\nThe main soft spot is exactly what the stress-test note flags: nothing isolates the federated aggregation. GeFL consistently beats the 10-client \"FedAvg\" baseline, but that baseline is just local training with no aggregation. The comparison bundles two things: having any synthetic-data augmentation and having a federated generator. Table IV compares against DA, and Table VII compares Syn vs Real+Syn, but no experiment trains a generator on a single client's data and feeds the same augmentation to the target networks. With the even data split described in footnote 5, a local generator could plausibly provide most of the augmentation benefit. That gap is a real hole in the central claim that the generative model 'gathers global knowledge from every client.'\n\nThere are two lesser issues. GeFL-F is called model-agnostic but actually requires every client to share the same feature extractor architecture, which limits the claim. And the privacy evaluation uses an MND proxy rather than an actual membership-inference attack; fine as a first pass, but not definitive.\n\nNone of this kills the paper. The framework is well-specified and the ablation can be added. I'd send it to review and ask for one control experiment: a locally trained generator (per client) with the same synthetic-augmentation budget, plus a membership-inference attack on the two variants.","headline":"Solid framework with thorough experiments, but the central claim about federated aggregation is not isolated from plain synthetic augmentation.","tokens_in":35081,"tokens_out":2673,"would_cite":true,"duration_ms":25437,"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":"A federated generative model lets clients with different architectures train together without sharing data.","keywords":["federated learning","model heterogeneity","generative model","feature-generative model","data augmentation","privacy","scalability","non-IID data"],"falsifier":"Train GeFL-F on a dataset with a deliberately weak or poorly warmed-up feature extractor and compare the mean-nearest-neighbor distance of reconstructed images with the image-level GeFL version; if the MND ratio rises above 1 or accuracy no longer beats grouped FedAvg, the privacy and scalability claims fail. For GeFL itself, run 100 clients with pathological label skew and check whether the federated generator collapses into low-diversity samples; collapse would falsify the claim that naive averaging of generator weights is sufficient.","tokens_in":34107,"feed_emoji":"🤖","tokens_out":7351,"duration_ms":63585,"temperature":0.7,"pith_summary":"The paper proposes GeFL, a federated learning framework in which clients collaboratively train a generative model on their local data and then use its synthetic samples to train their own locally chosen target networks. The aim is to make federated learning work when clients have different model architectures and no public dataset is available, two conditions that standard FL and distillation-based alternatives do not handle together. The paper argues that a federated generative model gathers the global data semantics and can transfer them to any architecture through generated, class-conditioned samples. A second variant, GeFL-F, shifts generation from raw images to features produced by a warmed-up common feature extractor, which the paper reports improves scalability to larger client populations, reduces communication and computation cost, and lowers memorization-based privacy risk. Across MNIST, FMNIST, CIFAR10, SVHN, and CelebA, both variants report accuracy gains over the comparison methods in heterogeneous settings.","feed_headline":"One shared generator lets clients with different models learn together","feed_subtitle":"GeFL shares knowledge through synthetic samples without public data; GeFL-F instead generates features, cutting cost and privacy risk.","key_machinery":"The load-bearing object is the federated generative model: a conditional GAN, VAE, or DDPM whose weights are trained by federated averaging of client updates and whose class-conditioned outputs serve as training data for every client's target network. In GeFL-F the object is the feature-generative model, a generative model trained on representations from a warmed-up common feature extractor, so the synthesized objects are low-dimensional features instead of images. This single shared generative object is what transfers knowledge across heterogeneous architectures without any client exposing raw data or another client's model parameters.","core_discovery":"On the paper's own terms, the central discovery is that cross-client knowledge in model-heterogeneous federated learning can be carried by a generative model trained federatively. In GeFL, each client trains the same generator backbone on private local data; the server simply averages the generator parameters and returns the aggregate, and the resulting conditional generator produces class-labeled synthetic samples. Each client trains its target network on those synthetic samples and then refines on real local data, while target-network parameters are averaged only within groups sharing the same architecture. The paper states that 'the trained generative model effectively gathers global knowledge from every client, enabling the training of target networks despite the different model architectures among clients.' GeFL-F repeats the idea at the feature level: clients warm up a shared feature extractor, train a feature-generative model on its outputs, and then train heterogeneous classifier headers on the generated features, which the paper reports as more scalable, cheaper, and less prone to memorization. The model-agnosticism claim is thereby qualified, because GeFL-F clients must share the feature-extractor architecture.","pith_inferences":["Beyond the paper: GeFL-F's privacy claim is measured by memorization of reconstructed images, so a stronger test would be a white-box gradient-based inversion attack on the shared feature extractor; if reconstructed images remain recognizable there, the privacy margin would shrink.","Beyond the paper: the homogeneity-level sweep shows a peak at intermediate sharing (HL 1), which suggests the feature-extractor size is a tunable knob; one could test whether the optimal sharing level shifts with dataset difficulty or client count and use that as deployment guidance.","Beyond the paper: the observation that five times more synthetic samples did not improve accuracy suggests a saturation effect, implying that synthetic diversity rather than synthetic volume drives the gain; varying generator guidance or truncation would test this directly.","Beyond the paper: GeFL aggregates the generator by naive parameter averaging, so on highly skewed label distributions the 100-client regime may benefit from clustered generator aggregation or server-side model interpolation, neither of which the paper explores."],"forward_implications":["Clients with entirely different CNN architectures can take part in one federated learning round and each improve its own model, removing the need for a shared global model or public data.","GeFL-F maintains or improves accuracy as the client count grows from 10 to 100, while image-level GeFL degrades, making the feature-level variant the scalable option for large federations.","Switching from raw-image generation to feature generation reduces total parameters, training and sampling cost, and the mean-nearest-neighbor distance ratio used to measure memorization (roughly 0.1 versus above 0.6 on MNIST), meaning the privacy risk from generative memorization is reduced.","Synthetic samples from the federated generator can be mixed with local real data and with standard data-augmentation methods such as MixUp, CutMix, AugMix, and AutoAugment, with gains that appear additive.","The accuracy advantage over grouped FedAvg widens as the amount of real data per client shrinks, so the method is most valuable in data-limited federated deployments."],"supporting_citations":[{"why":"Supplies the federated averaging baseline and the aggregation rule used to combine client updates.","marker":"[2]"},{"why":"The main public-data-dependent baseline that GeFL contrasts against by requiring no public dataset.","marker":"[13]"},{"why":"A model-agnostic baseline that averages logits across clients, with communication and divergence issues GeFL avoids.","marker":"[4]"},{"why":"Prior feature-level conditional GAN work in FL that shares a classifier; GeFL-F extends the idea to heterogeneous headers with a shared feature extractor.","marker":"[31]"},{"why":"Provides the GAN backbone used as one instantiation of the federated generative model.","marker":"[17]"},{"why":"Provides the VAE backbone used as another instantiation of the federated generative model.","marker":"[18]"},{"why":"Provides the DDPM backbone used as a third instantiation of the federated generative model.","marker":"[44]"},{"why":"LPIPS perceptual distance metric used to compute the mean-nearest-neighbor distance memorization ratio.","marker":"[61]"}],"fun_headline_variants":["One generator lets heterogeneous federated clients share knowledge","Generative model unifies federated learning across model types","Federated learning with generative models: no shared architecture needed","Feature-level generation boosts federated scalability and privacy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework depends on a shared generator, trained by simple averaging of client updates, producing synthetic samples that are at the same time high-quality, diverse, and private enough to train every client's differently built network; the feature-based variant adds the requirement that all clients use the same feature-extractor architecture.","fun_headline_variants_meta":{"raw":{"variants":["One generator lets heterogeneous federated clients share knowledge","Generative model unifies federated learning across model types","Federated learning with generative models: no shared architecture needed","Feature-level generation boosts federated scalability and privacy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1875,"prompt_tokens":979,"completion_tokens":896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":832}},"tokens_in":595,"tokens_out":896,"duration_ms":8555,"temperature":1.0,"reasoning_tokens":832,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:41:18.498365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GeFL-F on a dataset with a deliberately weak or poorly warmed-up feature extractor and compare the mean-nearest-neighbor distance of reconstructed images with the image-level GeFL version; if the MND ratio rises above 1 or accuracy no longer beats grouped FedAvg, the privacy and scalability claims fail. For GeFL itself, run 100 clients with pathological label skew and check whether the federated generator collapses into low-diversity samples; collapse would falsify the claim that naive averaging of generator weights is sufficient.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The main public-data-dependent baseline that GeFL contrasts against by requiring no public dataset."},{"cited_title":"Afonin and S","cited_arxiv_id":null,"evidence_quote":"A model-agnostic baseline that averages logits across clients, with communication and divergence issues GeFL avoids."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GAN backbone used as one instantiation of the federated generative model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DDPM backbone used as a third instantiation of the federated generative model."},{"cited_title":"Zhang, P","cited_arxiv_id":null,"evidence_quote":"LPIPS perceptual distance metric used to compute the mean-nearest-neighbor distance memorization ratio."}],"review_version":1}