{"id":"d171ab58-366f-4e72-a071-8399bf0653cf","arxiv_id":"2504.19103","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DRDFL splits each client model into a private Gaussian-mixture encoder and a shared adversarially trained encoder, reporting better personalization and generalization in ring-topology decentralized federated learning.","lead":"This paper introduces DRDFL, a training method for ring-shaped decentralized federated learning where models pass around a circle, not through a central server. It splits learning into a private personalization part and a shared generalization part, and reports higher accuracy with much smaller messages than eight comparison methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EMA class-statistic sharing with α=0.99 barely mixes on a 20-client directed ring, so the claimed 'global' priors are nearly local and may not drive the reported gains.","rationale":"The reader's CONDITIONAL verdict is reasonable: the empirical claim is supported by consistent tables but not by released code, error bars, or a rigorous proof. My stress-test pass identifies one more specific, load-bearing problem than 'missing code': the EMA update of class statistics in Algorithm 1 line 6 does not actually mix around the ring under the stated α=0.99, M=20, T=300. A direct eigenvalue analysis of the synchronous ring update shows the second eigenmode decays only to about 86% of its initial magnitude after 300 rounds, so the 'global' class means and covariances are essentially local. This is not a minor implementation detail: PersonaNet's loss (Eq. 6, Eq. 7) is driven by these statistics, and the convergence proof in Appendix D never models them. The theorem's Assumption 3 bounds only the Learngene difference and is itself assumed rather than derived; Lemma 2's step (a) approximates ΔL ≈ η∥Δφ∥² without a Lipschitz-gradient justification, and Theorem 2's learning-rate condition η < 2(ε−δ²)/(L1(ε+Eσ²)) implicitly requires ε > δ², so 'arbitrary ε' convergence is not established. I also note the adversarial objective in Eq. 9/10 is ambiguous: minimizing L_adv with respect to φ would make z_l more class-predictive, not invariant, unless the sign is reversed for φ, which Algorithm 1 line 13 does not specify. I focus on the EMA mixing issue because it directly attacks the paper's central contribution: if global priors are not shared effectively, the empirical gains may be attributable to the local personalized VAE, making the proposed ring-collaboration mechanism unverified. The concrete ablation test would settle this by removing only the EMA exchange while keeping all other communication; if performance is unchanged, the headline claim should be reframed or the method revised. Since this is a checkable empirical condition and the reader already requires conditions, I do not move the verdict: it remains conditional on this ablation and on the release of seeds/code.","tokens_in":23543,"tokens_out":11145,"duration_ms":118168,"concrete_test":"Run a full DRDFL experiment on CIFAR-10 with β=0.1 and β=0.4 (the settings behind the headline gains) with Algorithm 1 line 6 replaced by the client's own statistics only, i.e., µ(m)_k ← µ(m)_k and Σ(m)_k ← Σ(m)_k, while keeping all Learngene communication and all other hyperparameters identical. Compare Local-T and Global-T over at least 3 seeds. If removing the EMA exchange changes accuracy by less than the seed variance (or less than about 0.5 points), the shared Gaussian-prior mechanism is not the source of the reported gains. As an analytical companion, report the mixing factor ρ^T for M=20, α=0.99, T=300 and the maximum pairwise disagreement of class means after 300 rounds in the no-training ablation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 line 6 updates each client's class statistics as µ(m)_k ← 0.99 µ(m)_k + 0.01 µ̃_k (and similarly for Σ). On the directed 20-client ring this is the linear operator P = 0.99I + 0.01S, where S is the cyclic shift. Its second eigenvalue has magnitude ρ = |0.99 + 0.01e^{2πi/20}| ≈ 0.9995, so the non-uniform mode decays as ρ^300 ≈ 0.865 after the full 300 training rounds. The transmitted 'global' class means and covariances are therefore still about 86% of each client's local initialization or current state; consensus is far from reached. This matters because Eq. 6 and Eq. 7 make PersonaNet training depend directly on these drifting targets, and the convergence proof in Appendix D does not model the EMA statistics at all: Theorem 2 bounds only Learngene aggregation through Assumption 3 (∥φ̃−φ∥² ≤ δ²) and never establishes an analogous bound on ∥µ̃−µ∥. Consequently, the paper's claim that 'global implicit class representations undergo iterative optimization and sharing across clients' is not realized under the stated hyperparameters (M=20, α=0.99, T=300), and the reported Local-T/Global-T gains may come from the local VAE/personalization components rather than from the claimed cross-client information sharing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRDFL, a ring-topology decentralized federated learning method that splits each client model into a personalized PersonaNet and a shared Learngene module. PersonaNet models class-conditional Gaussian latent distributions with EMA-updated class statistics, while Learngene is trained with a KL prior and an adversarial classifier to extract class-invariant representations. The two modules are combined in a VAE-style decoder, and only the Learngene plus class statistics are exchanged along the ring. Experiments on SVHN, CIFAR-10, and CIFAR-100 under Dirichlet and shard-based non-IID partitions compare DRDFL with eight baselines, reporting higher Local-T and Global-T accuracies with only 0.58 MB of communication per round. The paper also provides an ablation study, convergence curves, Grad-CAM visualizations, a new-client adaptation experiment, and a gradient-inversion robustness study, plus a convergence proof in Appendix D.","tokens_in":23882,"tokens_out":4967,"duration_ms":51309,"significance":"If the claims hold, DRDFL is a useful contribution to decentralized FL: it demonstrates a parameter-efficient way to combine personalization and generalization under ring-topology constraints, and the additional experiments on new-client adaptation and resistance to gradient inversion are valuable. The strongest assets are the low communication cost (0.58M parameters exchanged), the consistent gains across three datasets and multiple heterogeneity settings, and the explicit ablation showing both loss components matter. However, the significance is tempered by three correctness gaps: the convergence proof relies on an unproved approximation, the adversarial objective appears internally inconsistent as written, and the EMA class-statistic sharing is not covered by the theory and may not achieve consensus under the stated hyperparameters. These issues need to be resolved before the empirical and theoretical claims can be fully accepted.","major_comments":[{"comment":"The adversarial training objective is specified inconsistently. L_adv in Eq. (9) minimizes the negative log-likelihood of the true class, which should make zl class-discriminative, while L^u_adv in Eq. (10) pushes the classifier toward uniform outputs over all classes. The paper does not define a min-max game, a gradient reversal layer, or alternating updates for the classifier parameters, so the two terms directly conflict as written. The claim that Learngene learns class-invariant representations is therefore not supported by the stated objective. Please clarify the actual training procedure and report the exact loss used in the experiments.","section":"Section IV-B, Eqs. (9)-(10)"},{"comment":"The key step in Lemma 2 is the approximation Delta L ≈ η ||φ(t+1)E+0 − φ(t+1)E||², asserted without proof in step (a) of Eq. (27). This is not a consequence of Assumptions 1-3, and the loss change after aggregation is not generally proportional to the squared parameter displacement. Since Lemma 2 underpins Theorem 1 and Theorem 2, the claimed O(1/T) non-convex convergence rate is not established. Please either provide a correct proof under explicit assumptions on the Learngene loss module, or weaken the claim and present the analysis as a heuristic.","section":"Appendix D, Lemma 2, Eq. (27)"},{"comment":"The convergence analysis does not cover the EMA dynamics of class statistics. Algorithm 1 line 6 updates mu and Sigma with α=0.99, and lines 9-10 train PersonaNet using those statistics, but Assumption 3 only bounds the Learngene parameter variation; no analogue is provided for ||mu_tilde − mu|| or ||Sigma_tilde − Sigma||. Thus Theorem 2 does not apply to the full algorithm. Moreover, on a directed 20-client ring with α=0.99, the mixing rate of the EMA update is very slow: the second eigenvalue of 0.99I + 0.01S has magnitude about 0.9995, so after 300 rounds the non-uniform mode decays only to about 0.865 and the 'global' class statistics remain far from consensus. Please provide a mixing analysis, choose a hyperparameter that actually mixes, or supply direct experimental evidence that the shared statistics converge.","section":"Algorithm 1, lines 6-10, and Theorem 2"},{"comment":"No standard deviations, seeds, or significance tests are reported. Several of the claimed improvements over baselines are small (e.g., 0.11-0.22 points in some Global-T columns of Table I), and the abstract's 'up to 3.28%' claim rests on a single run. Without multiple seeds and error bars, the central empirical claim that DRDFL 'outperforms state-of-the-art methods' is not statistically verified. Please rerun with at least three random seeds and report mean±std, and perform a paired test for the main comparisons.","section":"Tables I-II and Figures 3-6"}],"minor_comments":[{"comment":"The sentence 'Following [57]-[59], we set the parameter α in EMA to 0.99' cites time-series references [57] and [58] and the authors' own [59], none of which clearly justify the EMA momentum choice for federated class statistics. Please cite a more relevant source or provide a brief justification.","section":"Section VII-B"},{"comment":"The label 'DDRFL' in Figure 6 should be 'DRDFL'.","section":"Figure 6"},{"comment":"The word 'pamrameter' is a typo for 'parameter'.","section":"Algorithm 1, line 2"},{"comment":"The text 'CIFAR-100 setting with α = 0.1 and s = 30' should read 'β = 0.1' instead of 'α = 0.1'.","section":"Section VII-C"},{"comment":"The distinction between the Learngene paradigm and the Learngene module is confusing after the footnote; consider using a different name for the module to avoid ambiguity throughout the paper.","section":"Section I and abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a promising architecture and a broad experimental study, but the theoretical appendix is currently not reliable: Lemma 2's approximation is unproved, and Theorem 2 ignores the EMA class statistics entirely. The adversarial loss as written would not produce class-invariant features, which suggests the implementation may differ from the text. These are fixable in a revision, but they are load-bearing for the paper's claims, so I recommend major revision rather than rejection. If the authors can correct the proof, clarify the adversarial training, add mixing analysis or change the EMA parameter, and report error bars, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: DRDFL is a genuinely parameter-efficient personalized DFL method, and the reported gains are consistent across three datasets and two heterogeneity types. If you work on decentralized FL, it's worth a read even though the paper overstates its own machinery.\n\nThe new thing is the packaging: each client has a VAE-style PersonaNet that models class latents as a Gaussian mixture, plus a shared adversarial Learngene module that is averaged around the ring. That combination—personalized GMM prior plus adversarially invariant shared encoder—is new in a ring topology, and the communication cost of 0.58M params per round is an honest plus. The ablations show both modules contribute, and the Grad-CAM figure suggests the Learngene does latch onto class-common regions. Give credit where due: the empirical story is coherent and the numbers, while lacking error bars, are not cherry-picked in an obvious way.\n\nThe soft spots are real. The stress-test math is right: with α=0.99 on a 20-client directed ring, the class-statistic EMA barely mixes—the non-uniform mode decays to about 86% of its initial value after 300 rounds. So the 'global implicit class representations' are mostly local for the statistics. The Learngene itself does mix (its averaging operator has a much faster decay), so the cross-client sharing story isn't empty, but the paper's explanation of what is shared is overstated. Relatedly, the convergence proof in Appendix D doesn't model the EMA class statistics at all; Lemma 2's linear approximation is a hand-wave, and Assumption 3 only bounds Learngene variation. I'd call the proof illustrative, not rigorous.\n\nThe privacy claim in the appendix ('fundamentally prevents such leakage') is too strong. Sharing distributional statistics is safer than raw gradients, sure, but it is not a guarantee. And I'd note the absence of FedST as a baseline, since it is directly a disentangled shared/personalized FL method, even if centralized. The heavy self-citation around Learngene is noticeable but not circular, since the empirical results don't depend on that prior work.\n\nBottom line: this deserves a serious referee. It is a useful, plausible contribution to the FL subfield, but the current version oversells the global-statistics story and the theory needs reworking. I'd recommend a major revision with error bars, code release, and a toned-down narrative.","headline":"A communication-efficient personalized DFL method with consistent empirical gains, but the 'global' class statistics are mostly local under the stated hyperparameters and the convergence proof is hand-wavy; worth a serious but critical referee.","tokens_in":24359,"tokens_out":4222,"would_cite":false,"duration_ms":40180,"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 ring-topology federated learning method that splits each client model into a personalized PersonaNet and a shared Learngene improves both local and global accuracy under data heterogeneity while exchanging only 0.58 MB of parameters per…","keywords":["decentralized federated learning","ring topology","personalization","generalization","data heterogeneity","Gaussian mixture","adversarial learning","Learngene"],"falsifier":"Track the per-hop change in the shared knowledge encoder, $\\|\\tilde{\\phi}^t - \\phi_m^t\\|_2^2$, and the Local-T/Global-T accuracies while increasing label skew, for example moving from Dir(0.4) to Dir(0.05) or to shards with s=2; if the drift grows sharply and accuracy collapses when two clients with disjoint label sets pass the model, the paper's convergence and balance claims would fail.","tokens_in":23361,"feed_emoji":"🔁","tokens_out":9048,"duration_ms":77802,"temperature":0.7,"pith_summary":"The paper argues that the conflict between generalization and personalization in serverless ring-topology federated learning can be resolved by splitting each client's model into two modules: a private PersonaNet that models class-specific features as a Gaussian mixture, and a shared Learngene that learns class-invariant representations through adversarial training. Only the Learngene and per-class mean/covariance statistics travel along the ring, so each round exchanges about 0.58 MB of parameters rather than full models. The claim is that this division lets clients keep their own data distributions while still converging to a common, transferable representation, improving both local test accuracy and global test accuracy across Dirichlet and shard-based non-IID settings. If true, it would give decentralized federated learning a way to get the benefits of shared knowledge and communication efficiency without a central server.","feed_headline":"One shared module lifts both local and global accuracy","feed_subtitle":"A 0.58 MB Learngene circulates around the ring, giving decentralized clients personalization plus transferable knowledge.","key_machinery":"The load-bearing mechanism is the divide-and-conquer latent decomposition with ring-based averaging. Each client's model is written as $w_m = [\\psi_m, \\phi, \\theta_m, \\omega_m]$, where $\\psi_m$ is PersonaNet, $\\phi$ is the shared Learngene, $\\theta_m$ the decoder, and $\\omega_m$ the classifier. On receiving its neighbor's Learngene and class statistics, the client sets $\\phi \\leftarrow (\\tilde{\\phi}+\\phi)/2$ and updates $\\mu_k, \\Sigma_k$ by exponential moving average with $\\alpha=0.99$, then trains locally with the ELBO objective: reconstruction, KL divergence to the class-conditional Gaussian prior for PersonaNet, and KL divergence plus an adversarial uniform-class loss for Learngene. The adversarial classifier, with parameters $\\vartheta$, and the uniform-prior term enforce that $z_l$ carries no class identity, which is what makes the module transferable around the ring.","core_discovery":"DRDFL's central discovery is that personalization and generalization need not be traded off if they are assigned to separate latent channels. PersonaNet maps each input to a class-conditioned latent $z_p$ regularized toward a per-class Gaussian prior with mean $\\mu_k$ and covariance $\\Sigma_k$ updated by exponential moving average with the inherited global statistics, so the representation stays discriminative for the client's own label distribution. Learngene maps the same input to a class-independent latent $z_l$, regularized toward $\\mathcal{N}(0,I)$ and pushed by an adversarial classifier toward a uniform class distribution, so the representation captures knowledge that transfers across clients. The two latents are concatenated and fed to a decoder; reconstruction with injected noise and a classifier on both original and perturbed data completes the objective. The claim is that this two-channel disentanglement lets the ring propagate a stable consensus representation while each client retains its personalized head, and the experiments report gains up to 3.28% in local test accuracy over eight comparison methods while sharing only the 0.58 MB Learngene and class statistics per iteration.","pith_inferences":["The same two-channel split could be lifted out of the ring: the Learngene could serve as a task-agnostic feature extractor for newly arriving domains or tasks, since its adversarial uniform-class training makes the shared latent class-independent; the paper only demonstrates fast convergence for new clients with unseen distributions, not transfer to new class sets.","A natural stress test would vary the EMA smoothing factor α rather than fixing it at 0.99; under fast-moving class distributions, a fixed slow EMA may lag, and adapting α per class could extend DRDFL to non-stationary federated settings without changing the architecture.","The privacy claim rests on the assumption that per-class means and covariances leak nothing about individual instances; that is plausible but unproven, so a membership-inference test on the shared Gaussian statistics would tell whether the privacy benefit is as strong as the gradient-inversion reconstruction experiment suggests.","If the learned shared representation is truly class-invariant, it could double as a calibration or out-of-distribution detector; a reader could test by checking whether the shared latent separates known from novel classes, which the paper does not do."],"forward_implications":["Clients can keep the personalized part of their model private while sharing only the Learngene and class-level Gaussian statistics, cutting per-round communication from hundreds of millions of parameters to 0.58 MB.","The same architecture produces high Local-T and high Global-T under both Dirichlet (β=0.1, 0.4) and shard-based (s=4,5,20,30) non-IID splits, whereas personalized baselines typically sacrifice one for the other.","New clients joining the ring after 200 rounds of training converge faster when initialized with the inherited Learngene and global priors than with a sparse-mask initializer.","Because raw gradients and instance-level representations are never exchanged, gradient-inversion attacks cannot reconstruct recognizable training images from the transmitted information.","Under the stated assumptions, each client's non-convex objective converges at rate $O(1/T)$, so the dual-module split does not by itself introduce a convergence barrier."],"supporting_citations":[{"why":"Introduces the Learngene paradigm of distilling shared knowledge into lightweight models, the conceptual basis for the shared module.","marker":"[23]"},{"why":"Extends Learngene to variable-sized models, supporting the claim that the shared component can initialize diverse downstream models.","marker":"[25]"},{"why":"Provides the variational autoencoder and ELBO that Equation (3) is built on.","marker":"[35]"},{"why":"Conditional VAE formulation that motivates conditioning the decoder and latent on class labels, also used in the Fedcvae baseline.","marker":"[36]"},{"why":"Large-margin Gaussian mixture loss used to model class-conditional feature distributions in PersonaNet.","marker":"[49]"},{"why":"DisPFL, a decentralized personalized FL method with sparse masks, used as a main DFL baseline and comparison for new-client adaptation.","marker":"[16]"},{"why":"DFedPGP, a decentralized personalization baseline that DRDFL compares against and outperforms.","marker":"[18]"},{"why":"FedRep, the personalized FL baseline that learns shared representations with local heads, used for comparison.","marker":"[12]"}],"fun_headline_variants":["Dual latents: personal and shared knowledge without trade-off","Separate channels: one for personalization, one for generalization","Ring FL: disentangling personal and shared knowledge","No compromise: coordinated personal and transferable latents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence proof assumes that the shared knowledge encoder barely changes when one client's copy is averaged with the next client's copy along the ring; whether that drift stays small under extreme label skew is what determines whether the balance holds.","fun_headline_variants_meta":{"raw":{"variants":["Dual latents: personal and shared knowledge without trade-off","Separate channels: one for personalization, one for generalization","Ring FL: disentangling personal and shared knowledge","No compromise: coordinated personal and transferable latents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001185,"raw_usage":{"total_tokens":4912,"prompt_tokens":985,"completion_tokens":3927,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":3861}},"tokens_in":601,"tokens_out":3927,"duration_ms":26902,"temperature":1.0,"reasoning_tokens":3861,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T06:01:34.734520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track the per-hop change in the shared knowledge encoder, $\\|\\tilde{\\phi}^t - \\phi_m^t\\|_2^2$, and the Local-T/Global-T accuracies while increasing label skew, for example moving from Dir(0.4) to Dir(0.05) or to shards with s=2; if the drift grows sharply and accuracy collapses when two clients with disjoint label sets pass the model, the paper's convergence and balance claims would fail.","supporting_citations":[{"cited_title":"Learngene from open-world to your learning task,","cited_arxiv_id":null,"evidence_quote":"Introduces the Learngene paradigm of distilling shared knowledge into lightweight models, the conceptual basis for the shared module."},{"cited_title":"Initializing variable-sized vision transformers from learngene with learnable transformation,","cited_arxiv_id":null,"evidence_quote":"Extends Learngene to variable-sized models, supporting the claim that the shared component can initialize diverse downstream models."},{"cited_title":"Rethinking feature distribution for loss functions in image classification,","cited_arxiv_id":null,"evidence_quote":"Large-margin Gaussian mixture loss used to model class-conditional feature distributions in PersonaNet."},{"cited_title":"Dispfl: Towards communication-efficient personalized federated learning via decentralized sparse training,","cited_arxiv_id":null,"evidence_quote":"DisPFL, a decentralized personalized FL method with sparse masks, used as a main DFL baseline and comparison for new-client adaptation."},{"cited_title":"Decentralized directed collaboration for personalized federated learning,","cited_arxiv_id":null,"evidence_quote":"DFedPGP, a decentralized personalization baseline that DRDFL compares against and outperforms."},{"cited_title":"Exploiting shared representations for personalized federated learning,","cited_arxiv_id":null,"evidence_quote":"FedRep, the personalized FL baseline that learns shared representations with local heads, used for comparison."}],"review_version":1}