{"id":"545b1f4f-106c-4c1f-9241-bf5d5a937e27","arxiv_id":"2501.15486","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"FedAlign reports state-of-the-art average accuracy on PACS, OfficeHome, Office-Caltech-10, and miniDomainNet by combining MixStyle augmentation with supervised contrastive and prediction alignment losses.","lead":"FedAlign is a federated learning method that mixes style statistics across clients and aligns feature representations and predictions to help models generalize to unseen visual domains. The paper reports higher average accuracy than prior federated domain generalization methods on four benchmarks, but the algorithm as written omits the cross-client sharing that the method description relies on.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed cross-client feature extension is not present in the executable description: Algorithm 1 transmits only model parameters, no style statistics are exchanged, and the adversarial loss of Sec. 3.4 is absent from Eq. (16).","rationale":"I agree with the reader's weakest_assumption: the load-bearing gap is that the cross-client feature extension and dual-stage alignment are not actually realized in the algorithm and loss presented. Algorithm 1's server side only aggregates model parameters, so the 'sample statistics' mentioned in Fig. 2 are never communicated; the referenced Algorithm 2 is absent; and the adversarial component from Sec. 3.4 is omitted from Eq. (16). These are not stylistic or presentational issues: the abstract and conclusion explicitly credit these modules for the reported state-of-the-art results, and Table 1's 82.96% PACS average is the central quantitative evidence. Without a complete, executable method, the empirical comparison cannot support the claimed contribution. I also note that no code or formal verification is provided, so the inconsistency cannot be resolved by inspection. For these reasons the current manuscript does not meet the standard for a reliable report of the stated method, and the REJECT verdict is appropriate. If the authors later release code and show that a local-only MixStyle variant reproduces the numbers, the contribution would need to be restated; if it does not, the reported numbers would require an unspecified mechanism. Either way, the claim as written is not verifiable.","tokens_in":10852,"tokens_out":5425,"duration_ms":48091,"concrete_test":"Re-run the exact protocol of Sec. 4 using only Algorithm 1 and Eq. (16): the server aggregates model parameters only, each client applies local MixStyle to its own batch, no statistics are exchanged, and no discriminator is trained. Compare the PACS average to the reported 82.96% under the same MobileNetV3-Large backbone, 10 communication rounds, 3 local epochs, and upload ratio r=0.1. If the reproduced average is within one percentage point, the cross-client feature extension is not responsible for the headline gain; if it is not within one percentage point, the authors must supply the missing Algorithm 2, the style-statistics exchange protocol, and the adversarial loss before the claim can be evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"The abstract's central claim is that FedAlign's cross-client feature extension and dual-stage alignment together deliver superior generalization to unseen domains, with Sec. 5.1 reporting 82.96% PACS average versus 79.85% for CCST. This claim is load-bearing only if the executed method actually contains the cross-client operation. As written, it does not. In Algorithm 1, the server loop (lines 3-10) receives only updated client parameters and broadcasts the global parameters; no channel-wise style statistics (Eqs. 7-10) are collected, aggregated, or redistributed, even though Fig. 2's caption says clients share sample statistics with the server. The MixStyle module M() is deferred to 'Algorithm 2', but no Algorithm 2 appears in the manuscript. The total loss in Eq. (16) is L = L_CLS + lambda1(L_SC + L_RC) + lambda2 L_JS; the domain discriminator and adversarial objective introduced in Sec. 3.4 appear nowhere in this loss, no update rule for the discriminator is given, and no adversarial term is reported in the experiments. Thus the components that make the method cross-client and privacy-preserving are unspecified. The measured accuracy could be produced by local MixStyle plus supervised-contrastive, consistency, and JS losses, a substantially weaker and already familiar method. The empirical comparison in Table 1 therefore does not validate the paper's distinctive mechanism, and the central claim is not independently checkable from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedAlign, a federated domain generalization method that combines a cross-client feature extension module (based on MixStyle-style style statistics interpolation and sharing) with a dual-stage alignment module (supervised contrastive loss, representation consistency loss, and Jensen-Shannon divergence on predictions). The authors claim state-of-the-art accuracy on PACS, OfficeHome, miniDomainNet, and Caltech-101 benchmarks under a federated leave-one-domain-out protocol, with minimal communication and privacy overhead. The central empirical claim is that this combination of modules yields superior generalization to unseen domains while preserving privacy.","tokens_in":11212,"tokens_out":2141,"duration_ms":20528,"significance":"If fully validated, FedAlign would be a genuinely useful contribution to federated domain generalization, offering a light-weight alternative to adversarial or generative approaches. The paper targets a real problem—limited domain diversity on clients under privacy constraints—and the proposed loss combination is plausible as a mechanism for learning domain-invariant features. However, the manuscript as written does not substantiate the central claim because the cross-client feature-sharing machinery described in the text and figure is absent from the executable algorithm, and the adversarial component is never integrated into the optimization. The strengths are limited to a plausible idea and an apparently competitive table of numbers; reproducibility is currently impossible given the missing implementation details and hyperparameter values. The significance is therefore conditional on a substantial revision that reconciles the description with the actual executed method.","major_comments":[{"comment":"The central described mechanism, cross-client feature extension via sharing of sample statistics (Eqs. 7–10), is not implemented in Algorithm 1. The server loop (lines 3–10) only receives updated model parameters and broadcasts global parameters; no channel-wise mean/variance statistics are collected, aggregated, or redistributed, contrary to Fig. 2's caption that 'clients share local model parameters and sample statistics with the server.' Consequently, the executed method appears to be local MixStyle augmentation plus alignment losses, not the cross-client feature transfer claimed in the abstract. The reported accuracy gains cannot be attributed to the paper's distinctive contribution.","section":"Algorithm 1 and Fig. 2"},{"comment":"Section 3.4 introduces an adversarial domain discriminator, but no adversarial loss appears in the total loss (Eq. 16), no update rule for the discriminator is given, and Algorithm 1 contains no discriminator training step. The paper therefore describes a component that has no identifiable effect on the optimized objective or the reported results. Either the adversarial module is unused, in which case it should be removed, or it is used and the manuscript fails to specify how—either way, the current text is inconsistent.","section":"Sec. 3.4, Eq. (16)"},{"comment":"The MixStyle module M() is said to be described in 'Algorithm 2', but no Algorithm 2 appears in the manuscript. The claimed 'clustering' and 'probabilistic sampling weights' enhancements are described only verbally; there are no equations, no definitions of the clustering procedure, and no statement of how the cluster weights enter the sampling of λ in Eqs. (9)–(10). These are not cosmetic omissions: they are necessary to reproduce the method and to assess whether the claimed cross-client extension is actually present.","section":"Section 3.3 and Algorithm 2"},{"comment":"The experimental section omits standard deviations, the number of clients, the exact non-IID partitioning procedure, and the values of hyperparameters λ1, λ2, α, batch size, mix_feat size, and clustering parameters. With only three runs reported and no variance, the claim in Section 5.1 that FedAlign 'consistently outperforms all baseline methods across the evaluated datasets' cannot be statistically assessed. The reported margins over CCST (e.g., 82.96 vs 79.85 on PACS) may be within noise, and without these details the result is not independently checkable.","section":"Section 4 (Experiments) and Table 1"},{"comment":"The supervised contrastive loss in Eq. (12) is written only for a single augmented view convention, while Algorithm 1 computes L_SC(Z^(1), Z) + L_SC(Z^(2), Z), but the definition in Eq. (12) does not specify how multiple views are handled (e.g., whether views are included in the positive set). The representation consistency loss in Eq. (13) uses |mix_feat| without defining it, and the notation h(X_aug) is ambiguous because Algorithm 1 calls the augmented batches X^(1) and X^(2). These ambiguities prevent a reader from implementing the loss exactly as intended.","section":"Eqs. (12)–(13), notation"}],"minor_comments":[{"comment":"The phrase 'superior generalization to unseen domains' is a strong claim that is only supported by a single table without variance; please phrase as 'competitive' until statistical significance is established.","section":"Abstract and Section 1"},{"comment":"The reference to 'CycleGAN' for style transfer is tangential to the federated setting; consider citing a more relevant federated style transfer baseline or removing it.","section":"Section 2.2"},{"comment":"The dataset name 'Caltech-10' in Table 1 is inconsistent with 'Caltech-101' in the text; also, the column header abbreviates domains inconsistently (e.g., 'C', 'P', 'R', 'S' without a legend).","section":"Table 1 and Section 4"},{"comment":"The sentence beginning 'This adversarial mechanism effectively mitigates domain shift' is asserted without any supporting ablation; consider removing or substantiating.","section":"Section 3.4"},{"comment":"The paper says 'We follow the partitioning strategy presented in Section 3.1', but Section 3.1 only defines the problem setup and does not describe a concrete data-partitioning procedure. Please provide the explicit split (e.g., number of clients, Dirichlet concentration, or per-client domain composition).","section":"Section 4"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is unsupported because the described algorithm does not implement the cross-client feature sharing or the adversarial training. This is not a matter of presentation; the executable algorithm (Algorithm 1) and the loss (Eq. 16) are at odds with the text. The empirical results therefore cannot be interpreted as validating the proposed method. I see no path to accept without major new experiments and a fully specified algorithm, which goes beyond the scope of a revision. The authors should be encouraged to resubmit a version that either removes the unsupported components or actually implements them with full details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before spending time on this one. First, the load-bearing component in the abstract and intro—\"cross-client feature extension\"—does not appear in the algorithm the paper gives. Algorithm 1 communicates only model parameters; no channel-wise style statistics are collected, aggregated, or redistributed. The MixStyle module is deferred to \"Algorithm 2,\" which does not exist. Second, Section 3.4 describes an adversarial domain discriminator, but its objective is absent from the total loss in Eq. (16), and no update rule is given. So the central mechanism that makes FedAlign \"cross-client\" and privacy-preserving is unspecified.\n\nThe paper is not without merit. The identity of the method as written is local MixStyle augmentation plus three alignment losses: supervised contrastive (LSC), MSE representation consistency (LRC), and Jensen-Shannon prediction consistency (LJS). That package is reasonable and known to help in DG. The experiments follow the standard leave-one-domain-out protocol and compare against the right baselines (FedAvg, FedProx, FedADG, FedSR, FedIIR, CCST, GA). The reported gains—e.g., 82.96% average on PACS vs. 79.85% for CCST—are plausible and, if reproducible, would be a small but useful improvement. The paper should get credit for combining these known modules in a federated setting and for testing on four benchmarks.\n\nThe soft spots, however, are load-bearing. The algorithm/equation mismatch means the \"cross-client\" claim is not independently checkable from the manuscript. A reader cannot tell whether the reported numbers come from the method advertised or from a weaker local-only variant. There are also no standard deviations, no hyperparameter values (lambda_1, lambda_2, alpha, number of clients, etc.), and no code release. The dataset description is sloppy—Caltech-101 is described with a reference to the Caltech-256 paper. These are fixable, but in the current form the paper's central claim is unverifiable.\n\nWho is this for? Researchers in federated domain generalization who want to see whether MixStyle-style augmentation with alignment losses transfers to FL. It could be a useful baseline. I would not cite it as is.\n\nFor peer review: I'd send it out rather than desk-reject, because the experimental question is worth referee time and the missing details could be supplied. The authors should be pushed hard to provide the actual algorithm, code, hyperparameters, error bars, and a corrected narrative that matches the implemented method. If they can do that, the paper has a chance. In the current state, it is not a reliable report of the stated method. Recommendation: send to review with major-revision expectations.","headline":"The paper's advertised cross-client mechanism is missing from its own algorithm and loss, so the reported SOTA gains are unverifiable; the local MixStyle+alignment package is plausible and deserves referee attention to get fixed.","tokens_in":11674,"tokens_out":2699,"would_cite":false,"duration_ms":24233,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FedAlign claims that augmenting each federated client's local batches with cross-client style statistics and aligning both representations and predictions yields top generalization to unseen domains under privacy constraints.","keywords":["federated learning","domain generalization","domain-invariant features","MixStyle","supervised contrastive loss","Jensen-Shannon divergence","feature alignment","privacy-preserving learning"],"falsifier":"Run FedAlign with the cross-client statistics sharing disabled, keeping only each client's local MixStyle and the dual-stage alignment losses; if the average PACS accuracy stays at 82.96% or the gap over CCST disappears, the cross-client extension is not the source of the gains.","tokens_in":1443,"feed_emoji":"🖼️","tokens_out":3046,"duration_ms":76196,"temperature":0.7,"pith_summary":"This paper proposes FedAlign, a way to train a single image classifier across several clients that each hold data from a different visual domain, so that the shared model works on a domain none of the clients have seen. The central claim is that two additions to ordinary federated training—a cross-client feature extension that diversifies each client's style statistics, and a dual-stage alignment that pulls representations and predictions closer across augmented views—let federated models beat existing federated domain generalization baselines. On the four benchmarks tested, FedAlign reports the highest average accuracy, for instance 82.96% on PACS versus 79.85% for the strongest baseline CCST, while transmitting only model parameters and sample statistics rather than raw data.","feed_headline":"FedAlign tops four federated domain generalization benchmarks","feed_subtitle":"Feature extension plus dual alignment hits 82.96% on PACS, beating prior federated DG methods","key_machinery":"The central object is the per-client augmented-training loop: for a batch $X$, generate $X^{(1)} = \\mathcal{M}(X)$ and $X^{(2)} = \\mathcal{M}(X)$ via a MixStyle module $\\mathcal{M}$ that interpolates channel-wise means and standard deviations, then compute representations and predictions for all three views and minimize the combined classification, representation-alignment, and prediction-alignment loss. The cross-client component is the sharing of channel-wise style statistics between clients and server, which the paper says broadens each client's effective domain coverage without exposing raw data.","core_discovery":"FedAlign's central claim is that federated domain generalization improves when each client augments its local batches with MixStyle-style style interpolation and is trained with a total loss $L = L_{CLS} + \\lambda_1(L_{SC} + L_{RC}) + \\lambda_2 L_{JS}$, where $L_{SC}$ is a supervised contrastive loss on representations, $L_{RC}$ is an MSE consistency loss between original and augmented representations, and $L_{JS}$ is a Jensen–Shannon divergence between predictions on original and augmented views. The paper reports consistent superiority over the compared methods across PACS, OfficeHome, Caltech-10, and miniDomainNet, and presents t-SNE visualizations as evidence that the learned representations are more compact and class-separated than those of federated baselines.","pith_inferences":["Inference: if FedAlign works as claimed, the dual-stage alignment losses may be the main driver, and a simplified baseline using only local MixStyle plus consistency losses could match the reported accuracy.","Inference: the privacy guarantee is weaker than stated, since channel-wise style statistics can still leak information about local data distributions; a reconstruction or attribute-inference attack could quantify the residual risk.","Inference: the method should transfer to non-image modalities whenever style statistics can be defined, such as sensor or text embeddings, where domain shifts follow similar patterns."],"forward_implications":["On PACS, FedAlign reports an average accuracy of 82.96%, beating the best prior federated domain generalization method CCST at 79.85%.","The same pattern holds on OfficeHome (68.03%), Caltech-10 (93.72%), and miniDomainNet (63.76%), with FedAlign reaching the top accuracy in each target domain on PACS and miniDomainNet.","FedAlign's advantage persists as the number of participating clients changes, whereas the comparison methods' accuracy deteriorates more markedly with larger client populations.","Sharing only channel-wise statistics rather than raw data or high-dimensional feature maps keeps communication and computation overhead minimal, according to the paper.","The learned representations are qualitatively more compact and class-separated under t-SNE, which the paper interprets as evidence of domain-invariant features."],"supporting_citations":[{"why":"Supplies the FedAvg baseline that all methods, including FedAlign, build on and compare against.","marker":"McMahan et al., 2017"},{"why":"Provides the MixStyle augmentation that FedAlign's feature-perturbation mechanism is based on.","marker":"Zhou et al., 2021"},{"why":"Defines CCST, the strongest cross-client style-transfer baseline that FedAlign outperforms on PACS.","marker":"Chen et al., 2023"},{"why":"Defines FedSR, the representation-alignment baseline whose regularizations FedAlign extends.","marker":"Nguyen et al., 2022"},{"why":"Provides the PACS dataset, the primary benchmark where FedAlign reports its largest margin.","marker":"Li et al., 2017"},{"why":"Defines FedADG, an adversarial federated DG baseline whose training instability FedAlign aims to avoid.","marker":"Zhang et al., 2021"}],"fun_headline_variants":["FedAlign beats federated baselines on four domain generalization benchmarks","Cross-client alignment boosts federated DG on four datasets","Privacy-preserving feature alignment beats baselines on four DG benchmarks","Federated DG improves with cross-client feature alignment"],"cache_read_input_tokens":13824,"weakest_assumption_plain":"The load-bearing premise is that clients and the server actually exchange channel-wise style statistics, so that the feature-extension step genuinely spans cross-client styles rather than only local ones.","fun_headline_variants_meta":{"raw":{"variants":["FedAlign beats federated baselines on four domain generalization benchmarks","Cross-client alignment boosts federated DG on four datasets","Privacy-preserving feature alignment beats baselines on four DG benchmarks","Federated DG improves with cross-client feature alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001216,"raw_usage":{"total_tokens":4957,"prompt_tokens":855,"completion_tokens":4102,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":4035}},"tokens_in":471,"tokens_out":4102,"duration_ms":30127,"temperature":1.0,"reasoning_tokens":4035,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:14:22.367542+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedAlign with the cross-client statistics sharing disabled, keeping only each client's local MixStyle and the dual-stage alignment losses; if the average PACS accuracy stays at 82.96% or the gap over CCST disappears, the cross-client extension is not the source of the gains.","supporting_citations":[{"cited_title":"Domain adaptive ensemble learning","cited_arxiv_id":null,"evidence_quote":"Provides the MixStyle augmentation that FedAlign's feature-perturbation mechanism is based on."},{"cited_title":"Federated domain generalization for image recognition via cross-client style transfer","cited_arxiv_id":null,"evidence_quote":"Defines CCST, the strongest cross-client style-transfer baseline that FedAlign outperforms on PACS."},{"cited_title":"Fedsr: A simple and effective domain generalization method for federated learning","cited_arxiv_id":null,"evidence_quote":"Defines FedSR, the representation-alignment baseline whose regularizations FedAlign extends."}],"review_version":1}