{"id":"e500bfd4-0af9-4b81-b3f0-1675711d382f","arxiv_id":"2412.11463","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"FedCAR weights each client's generator by the inverse pairwise FID between its fake images and other clients' fake images, reporting modest FID improvements over FedAvg and centralized learning in non-i.i.d. chest X-ray generation.","lead":"This paper proposes FedCAR, a federated learning method that re-weights each hospital's generator by how similar its fake X-ray images are to other hospitals' fake images. The authors report lower FID scores than FedAvg and even centralized training on three public chest X-ray datasets, though the gains are small and statistical significance is not reported.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FedCAR's own severe non-i.i.d. results show α_n stays ≈0.33 for the data-deficient CheXpert client despite its individual FID of 53.05, so the fake-fake FID proxy does not track quality and the reported gains cannot be attributed to adaptive re-weighting.","rationale":"Reading in good faith, the paper's goal is to improve federated GAN training by adaptively re-weighting client contributions based on server-side pairwise FID between generated fake images. For that claim to hold, two conditions are necessary: α_n must track generation quality, and the resulting weights must materially differ from FedAvg in a way that explains the reported gains. The manuscript's own severe non-i.i.d. numbers undermine the first condition: CheXpert is known to be the weak client (individual FID 53.05), yet α_n is reported to stay near 0.33 with only a slight decrease. A quality-tracking weight would need to drop much more, as FedAvg's data-size weight of 0.048 illustrates. Thus the proposed fake-fake FID proxy is doing something other than measuring the quality gap that the paper claims to exploit. The mild scenario reinforces the problem: with α≈1/3 for all clients, FedCAR is effectively equal-weight aggregation, so the small FID improvements over FedAvg (7.12 vs 8.13) cannot be attributed to the adaptive mechanism. The absence of error bars or multiple seeds leaves even the size of the improvements uncertain. Algorithm 2's scalar F_bias is an additional internal inconsistency that makes the exact implemented algorithm unclear, though the release of code is a positive and makes the proposed test feasible. These concerns are the same load-bearing weakness identified by the reader, and they support keeping the REJECT verdict; no adjustment is needed.","tokens_in":7264,"tokens_out":11064,"duration_ms":97020,"concrete_test":"Run the released FedCAR code in the severe non-i.i.d. configuration for at least 5 seeds, logging per-round α_n, and compare three variants: (a) FedCAR as released, (b) FedCAR with α_n frozen to 1/3 (uniform weights), and (c) FedAvg with dataset-size weights. Report per-client and average FID as mean±std. If (a) does not significantly beat (b), the adaptive re-weighting contributes nothing beyond uniform weighting; if (b) does not beat (c), the severe-case gain is solely from ignoring dataset size. In the same logs, check whether α_CheXpert drops materially below 0.33 as the client's poor individual FID (53.05) would require; if not, the fake-fake FID proxy fails to track quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires α_n from Eq. 2–4 to measure each client's contribution quality. The severe non-i.i.d. experiment is the critical case: CheXpert has 1,000 images (10% of the other clients) and an individual FID of 53.05, versus 11.87 and 5.49 for NIH and VinDr. A quality-tracking weight should assign CheXpert far less than 1/3; FedAvg assigns it 0.048. The paper reports Average(α_n)≈0.33 in both scenarios and only a 'slight decrease' for CheXpert (Section 4, Fig. 2, Supplement Table 1). This shows the pairwise fake-fake FID proxy does not track the known quality gap. Consequently the reported severe-case improvement over FedAvg (12.76 vs 13.36 FID) is more plausibly explained by near-equal up-weighting of a weak client than by 'favoring clients with better performance,' and the mild-case α≈0.33 makes FedCAR nearly equal-weight FedAvg. Algorithm 2 line 10–11 also defines F_bias as a scalar 1/ΣF_i multiplied into all θ′, inconsistent with the per-client normalized α of Eq. 3–4, so the implemented re-weighting is ambiguous.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedCAR, a server-side aggregation algorithm for federated GAN training. At each round the server generates fake images from each client's generator, computes pairwise FID scores between these fake-image sets, converts the resulting per-client totals into normalized weights alpha_n (Eqs. 2-4), and re-weights the client generator updates during aggregation. The authors evaluate FedCAR on three public chest X-ray datasets (NIH, CheXpert, VinDr) using StyleGAN2 in two non-i.i.d. scenarios, and report average FID scores lower than centralized learning, FedAvg, and FedAdam. The code is released.","tokens_in":7542,"tokens_out":3924,"duration_ms":35260,"significance":"If the proposed mechanism worked, it would be a useful contribution to federated generative modeling, since server-side re-weighting based on generated outputs rather than raw data is an attractive idea for privacy-preserving medical imaging. The paper has several strengths: it uses three well-known public datasets, a strong StyleGAN2 baseline, comparisons to FedAvg and FedAdam as well as centralized learning, and it releases source code. However, the reported results do not establish that the adaptive re-weighting mechanism is responsible for the observed performance, because the paper's own reported alpha values barely deviate from uniform weights and do not track the known quality gap between clients. The central claim therefore rests on an unvalidated proxy and an implementation ambiguity.","major_comments":[{"comment":"Table 1 reports a single FID value per method and scenario, with no variance, confidence intervals, or significance tests. The headline margins over centralized learning are 0.23 FID (severe) and 1.01 FID (mild), which are small relative to the typical stochasticity of StyleGAN2 training. Multiple random seeds and a statistical comparison are required before the claimed superiority can be assessed.","section":"Section 3.2, Table 1"},{"comment":"The authors state that Average(alpha_n) is approximately 0.33 in both the mild and severe scenarios, and the Discussion explicitly says that FedCAR 'does not re-weight by far compare to FedAvg in average.' In the severe scenario, the CheXpert client has an individual FID of 53.05 while the NIH and VinDr clients have FIDs of 11.87 and 5.49, yet the adaptive weights show only a slight decrease for CheXpert. A quality-tracking weight should assign this client far less than one-third of the weight. Consequently, the reported improvement over FedAvg in the severe scenario (12.76 vs. 13.36 FID) cannot be attributed to 'favoring clients with better performance' as claimed; it is more plausibly explained by a nearly uniform aggregation that happens to differ from data-proportional weighting. This directly undermines the central claim of the paper.","section":"Section 4, Fig. 2, Supplement Table 1"},{"comment":"There is an inconsistency between the mathematical formulation and the pseudocode. Equations (3)-(4) define a per-client normalized weight alpha_n = (1/FID_n) / sum_i (1/FID_i), which should produce a distinct weight for each client. Algorithm 2, however, computes F_bias as the scalar 1/(F1+F2+...+Fn) and then applies it uniformly to all theta' in the aggregation step. The paper does not clarify whether the actual implementation used per-client weights as in Eqs. (3)-(4) or a single scalar as in Algorithm 2. This ambiguity must be resolved because it changes the meaning of every reported result.","section":"Section 2.2, Eqs. (3)-(4), Algorithm 2 lines 10-11"},{"comment":"The method assumes that pairwise FID between fake images generated by different clients is a valid proxy for each client's contribution quality, but this proxy is never validated against the evaluation metric (FID to real data). The authors' own severe-scenario results contradict the proxy's expected behavior: the CheXpert client's individual FID is an order of magnitude worse, yet its learned alpha is close to 1/3. In addition, the number of fake images per client per round (x = 1,000) and kimg per round (k = 100) are free parameters that are not varied or justified, so the sensitivity of the results to these choices is unknown.","section":"Section 2.2 and Section 3.1"}],"minor_comments":[{"comment":"There are numerous typos and grammatical issues, for example 'obtacles' instead of 'obstacles', 'instablity' instead of 'instability', and 'previous researches' instead of 'previous research'. The manuscript would benefit from careful proofreading.","section":"Throughout"},{"comment":"The notation in Eq. (2) is confusing because 'n' is used both as the client index and in the expression N-n, and the summation limits are not clear. Using distinct indices such as i and j with explicit sums over all pairs would improve readability.","section":"Section 2.2, Eq. (2)"},{"comment":"The slice 'theta'[a : a + b - 1]' appears off by one; if the parameter list has length a + b, the discriminator part should be 'theta'[a : a + b]'. Please verify.","section":"Algorithm 1, line 17"},{"comment":"The sentence 'For experiment, we randomly sampled 10,000 images of \"No finding\" from each dataset, in total of 30,000 images' is repeated verbatim two sentences later. One copy should be removed.","section":"Section 3.1"}],"recommendation":"reject","confidential_remarks":"The core problem is not a missing ablation but the fact that the paper's own alpha dynamics contradict the proposed mechanism: the adaptive weights are effectively uniform and do not respond to a known, large quality gap. This cannot be repaired by additional experiments within the current method as described, and the inconsistency between Eqs. (3)-(4) and Algorithm 2 raises further doubt about what was actually evaluated. I would not encourage resubmission unless the aggregation algorithm or its validation is substantially redesigned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is genuinely new: re-weighting client contributions in a federated GAN by pairwise FID between fake images generated by each client. That specific mechanism is not in the FedGAN or FedOpt baselines they compare against, and they released code and used public datasets. Credit where due: the formulation is clean, and the paper honestly reports that alpha stays near 0.33 in both scenarios.\n\nBut the central claim falls apart under their own numbers. In the severe non-i.i.d. case, the CheXpert client has 1,000 images and an individual FID of 53.05, versus 11.87 and 5.49 for the other two. A quality-tracking weight should assign CheXpert a small fraction of the aggregate; FedAvg gives it 0.048. FedCAR reports only a slight decrease from 0.33. That means the fake-fake FID proxy does not track the known quality gap. So the reported severe-case gain over FedAvg (12.76 vs 13.36) is more plausibly explained by near-uniform up-weighting of a weak client than by \"favoring clients with better performance.\"\n\nIn the mild case, alpha ≈ 0.33 makes FedCAR nearly identical to FedAvg, yet it beats FedAvg by 1.01 FID. Without error bars, variance, or significance tests—Table 1 has single FID values—that margin could easily be noise. The gains over centralized learning are 0.23 and 1.01 FID, again without any uncertainty. One GAN architecture (StyleGAN2) and three chest X-ray datasets are a thin basis for the broad superiority claim.\n\nThere is also an internal inconsistency in Algorithm 2: line 10–11 defines F_bias as a scalar 1/ΣF_i multiplied into all θ', which contradicts the per-client normalized alpha in Eq. 3–4. Discriminator aggregation is left as a vague \"F(θd)\" in Eq. 5, so the implemented update is ambiguous.\n\nThis is a paper with a new but unvalidated mechanism. The reader's verdict is right: reject in current form. Still, if the authors can validate the proxy, add error bars, and clean up the algorithm description, the idea has enough novelty to warrant a serious look. I would not cite it as evidence for anything yet, but I would bring it to a reading group to discuss why the proxy fails so badly.","headline":"The core claim does not survive its own data: FedCAR's weights stay at ~0.33 even for a client with FID 53, so the mechanism is FedAvg in disguise and the small gains are not attributable to the re-weighting.","tokens_in":8083,"tokens_out":1909,"would_cite":false,"duration_ms":18337,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A federated GAN aggregation rule that re-weights clients by cross-client image similarity outperforms centralized training on chest X-ray generation.","keywords":["federated learning","generative adversarial networks","medical image generation","adaptive re-weighting","Fréchet Inception Distance","non-i.i.d. data","chest X-ray","StyleGAN2"],"falsifier":"Run FedCAR with alpha_n fixed at 1/N (uniform weights) under the same experimental setup; if the average FID is statistically indistinguishable from the adaptive version, the paper's central claim about adaptive re-weighting is falsified. Alternatively, compute the correlation between the adaptive alpha_n and each client's FID against the real global data distribution; a zero or negative correlation would invalidate the proxy.","tokens_in":7038,"feed_emoji":"🩻","tokens_out":4758,"duration_ms":37981,"temperature":0.7,"pith_summary":"This paper proposes FedCAR, a server-side aggregation rule for generative adversarial networks trained under federated learning. Instead of weighting each hospital's model by dataset size, FedCAR generates fake images from each client's generator each round and weights that client inversely to the Fréchet Inception Distance between its fake images and those of the other clients. On three public chest X-ray datasets, the authors report that FedCAR beats both conventional federated aggregation (FedAvg, FedAdam) and centralized learning on average FID, in both mild and severe non-i.i.d. splits. The key claim is that adaptive re-weighting based on cross-client distribution distance makes federated GAN training more efficient than training on pooled data, which would let hospitals collaborate on generative models without sharing raw images.","feed_headline":"Federated GAN beats centralized training on chest X-rays","feed_subtitle":"A new aggregation rule that re-weights hospitals by fake-image similarity reaches FID 7.12 versus 8.58 for pooled training.","key_machinery":"The machinery is a pairwise FID matrix between fake image sets. In each round the server samples fake images from each client's generator, computes Fréchet Inception Distance between every pair of clients, sums row-wise to get a total FID for each client, converts these to weights alpha_n = 1/FID_n, normalizes them, and aggregates the generators as a weighted sum while aggregating the discriminator separately. This gives a data-free, server-side evaluation that uses only generated images, preserving privacy.","core_discovery":"The central discovery claim is that a federated GAN can outperform a GAN trained on the pooled dataset, and that the aggregation rule matters: re-weighting client contributions by the inverse of their cross-client FID yields average FID of 7.12 in the mild non-i.i.d. scenario and 12.76 in the severe scenario, versus 8.58 and 12.99 for centralized learning. The authors attribute this to FedCAR 'preserving the characteristics of the client' each round while still sharing knowledge. They also observe that the learned weights stay near 0.33 for all three clients, so the mechanism is a slight per-round adjustment rather than a strong re-weighting.","pith_inferences":["The paper's own alpha values staying near 0.33 imply that FedCAR's practical advantage may come less from the specific adaptive weights and more from the fact that even a small, data-dependent perturbation of FedAvg changes GAN training dynamics; a direct comparison against fixed uniform weights would settle this.","A natural testable extension is to validate the cross-client FID proxy against each client's FID to the real pooled data distribution; if they disagree, the re-weighting signal is not actually tracking generator quality.","The authors' finding that federated GAN beats centralized learning in both scenarios could be an artifact of the specific StyleGAN2 training budget or initialization; replicating with different seeds and training lengths would check robustness."],"forward_implications":["If the reported gains hold, hospitals can jointly train high-quality generative models without transmitting patient data, with results that beat a single centralized model on heterogeneous data.","The method's principle—server-side evaluation of client generators using only generated samples—extends beyond GANs to any generative model in federated settings.","The observation that federated GAN training can beat centralized training suggests that federated learning may act as a regularizer for GANs on non-i.i.d. data.","In severe non-i.i.d. settings with one data-poor client, FedCAR's re-weighting lets the data-poor client benefit more than data-proportional averaging would allow."],"supporting_citations":[{"why":"Supplies the FedAvg baseline, the standard data-proportional averaging that FedCAR is compared against.","marker":"[12]"},{"why":"Supplies the FedAdam adaptive federated optimization baseline used as a comparison.","marker":"[17]"},{"why":"Provides the StyleGAN2 architecture used as the backbone in all experiments.","marker":"[9]"},{"why":"Is the federated GAN method whose synchronous generator-discriminator training FedCAR extends.","marker":"[16]"},{"why":"Is the NIH ChestX-Ray8 dataset, one of the three client datasets in the experiments.","marker":"[19]"},{"why":"Is the CheXpert dataset, which serves as the data-poor client in the severe non-i.i.d. scenario.","marker":"[7]"},{"why":"Is the VinDr-CXR dataset, one of the three client datasets in the experiments.","marker":"[13]"},{"why":"Supplies the synchronized federated GAN training approach that stabilizes convergence in FedCAR.","marker":"[3]"}],"fun_headline_variants":["Federated GAN with adaptive client weighting beats pooled training","Cross-client re-weighting gives federated GANs an edge","Adaptive re-weighting improves federated GANs on chest X-rays","FedCAR: Re-weighting clients by fake-image similarity beats central"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The re-weighting rests on treating pairwise FID between fake images as a measure of each client's contribution quality, but the paper never checks that this proxy matches FID to real data, and its own reported weights stay near uniform, so the claimed improvement may not come from the adaptive mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Federated GAN with adaptive client weighting beats pooled training","Cross-client re-weighting gives federated GANs an edge","Adaptive re-weighting improves federated GANs on chest X-rays","FedCAR: Re-weighting clients by fake-image similarity beats central"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3500,"prompt_tokens":886,"completion_tokens":2614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":2538}},"tokens_in":502,"tokens_out":2614,"duration_ms":15282,"temperature":1.0,"reasoning_tokens":2538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:53:36.114729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedCAR with alpha_n fixed at 1/N (uniform weights) under the same experimental setup; if the average FID is statistically indistinguishable from the adaptive version, the paper's central claim about adaptive re-weighting is falsified. Alternatively, compute the correlation between the adaptive alpha_n and each client's FID against the real global data distribution; a zero or negative correlation would invalidate the proxy.","supporting_citations":[{"cited_title":"In: International Conference on Artificial Intelligence and Statistics (2016), https: //api.semanticscholar.org/CorpusID:14955348","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg baseline, the standard data-proportional averaging that FedCAR is compared against."},{"cited_title":"In: International Confer- ence on Learning Representations (2021), https://openreview.net/forum?id= LkFG3lB13U5","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAdam adaptive federated optimization baseline used as a comparison."},{"cited_title":"2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp","cited_arxiv_id":null,"evidence_quote":"Provides the StyleGAN2 architecture used as the backbone in all experiments."},{"cited_title":"In: 2017 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Is the NIH ChestX-Ray8 dataset, one of the three client datasets in the experiments."},{"cited_title":"In: AAAI Conference on Artificial Intelligence (2019), https://api.semanticscholar.org/CorpusID:58981871","cited_arxiv_id":null,"evidence_quote":"Is the CheXpert dataset, which serves as the data-poor client in the severe non-i.i.d. scenario."},{"cited_title":"Scientific Data9 (2020), https://api","cited_arxiv_id":null,"evidence_quote":"Is the VinDr-CXR dataset, one of the three client datasets in the experiments."},{"cited_title":"In: Peng, Y., Liu, Q., Lu, H., Sun, Z., Liu, C., Chen, X., Zha, H., Yang, J","cited_arxiv_id":null,"evidence_quote":"Supplies the synchronized federated GAN training approach that stabilizes convergence in FedCAR."}],"review_version":1}