{"id":"b518bceb-d5ec-4a0d-9567-ef140ff389ea","arxiv_id":"2505.16190","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A reputation-aware federated survival analysis framework using DP-protected peer feedback and clustering reports stable C-index gains, but its convergence argument assumes the true reliability it claims to prove.","lead":"Federated survival analysis gets a peer-driven reputation mechanism that privatizes client updates before peers score them, then clusters hospitals by data completeness. The reported C-index gains over standard federated averaging are small and lack error bars, and the convergence proof contains a circular assumption.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The peer-feedback signal m_{j,k} is never shown to track true contribution quality; Section 4.5 simply assumes this, so the reputation mechanism's central benefit is unproven and can invert under data heterogeneity.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the feedback signal is treated as an unbiased measure of contribution quality without validation. I agree, and the issue is more fundamental than the missing variance in Tables 2 and 3. Even with perfect convergence of the reputation recurrence, the reputation scores would only reflect local C-index deltas, not necessarily global value. Section 4.5's convergence lemma explicitly assumes reliable feedback m = RS* + eta, which is the very property that needs to be established; moreover the algebra inside that lemma is incorrect as written. The proposed DP decoupling is a reasonable design choice, and clustering on feature completeness is plausible, but neither rescues the central claim if the feedback signal is unfaithful. The concrete test above would settle whether the mechanism actually rewards clients that improve the global survival model, or merely clients that locally correlate with a peer's data. Since the reader already rejected on related grounds, my read leaves the verdict unchanged.","tokens_in":14547,"tokens_out":5202,"duration_ms":45435,"concrete_test":"On the Section 5.1 synthetic setup with heterogeneous centers and a held-out global evaluation set, compute per round: (a) each client k's average peer feedback mbar_k = mean_j [Omega(M(theta_j, theta_k)) - Omega(M(theta_j))] as in Section 4.4; and (b) the leave-one-out global contribution Delta_k = C_global(aggregate with all updates) - C_global(aggregate without theta_k). Report the Spearman correlation between mbar_k and Delta_k across clients and rounds, and compare reputation-weighted aggregation against equal-weight FedAvg over at least 10 seeds. If the correlation is approximately zero or negative, a client that helps one peer's local concordance can be rewarded while hurting the global model; if reputation weighting does not beat equal weights on held-out C-index, the central outperformance claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is the feedback signal of Sections 3.4 and 4.4: m_{j,k} = Omega(M(theta_j, theta_k)) - Omega(M(theta_j)), the change in peer j's local C-index when client k's update is included. For reputation-weighted aggregation to down-weight harmful clients, this local delta must be positively related to the effect of theta_k on the global model. With heterogeneous local data this can fail: a Cox coefficient update that improves concordance on j's data can move the aggregated model away from the population-level optimum or simply overfit j's censoring pattern. The paper provides no experiment linking m_{j,k} to held-out global C-index. Section 4.5's Lemma does not close the gap because it assumes m_{wv}(t) = RS*_v + eta, i.e., that feedback already equals true reliability plus zero-mean noise; that is exactly the property at issue. The proof is also algebraically wrong: the update gives E[e(t+1)] = E[e(t)] - alpha E[RS_uw] RS*_v, not the claimed contraction (1 - alpha E[RS_uw] RS*_v) E[e(t)]; the error is driven negative rather than to zero. Since selection probabilities P(i) are proportional to RS_i, if reputation scores do not track global contribution the method degenerates to weighting by a local correlation artifact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a federated survival analysis framework using Cox proportional hazards models with a peer-driven reputation mechanism. Reputation scores are updated from peer feedback defined as the change in a peer's local concordance index when incorporating another client's update. Clients are clustered by feature-completeness vectors, updates are DP-noised for reputation sharing, and global aggregation weights updates by reputation-based selection probabilities. Theoretical claims include convergence of reputation estimates to true reliability and DP accuracy guarantees; experiments on synthetic data and SEER compare against TFFL and no-reputation FedAvg.","tokens_in":14827,"tokens_out":5777,"duration_ms":43771,"significance":"If the mechanism were shown to make reputation track true contribution quality, the paper would address a real problem in federated healthcare. The decoupling of DP-noised peer feedback from clean server aggregation is an interesting architectural choice, and the SEER evaluation is a strength. However, the central convergence proof assumes the very property at issue, contains an algebraic error, and the experiments do not establish that local C-index improvements correlate with global model quality. The theoretical contribution is therefore not load-bearing, and the empirical support is insufficient to support the abstract's claims.","major_comments":[{"comment":"The proof of the Lemma assumes m_wv(t) = RS*_v + eta, i.e., that the feedback signal is already the true reliability plus zero-mean noise. This is exactly the property the lemma is supposed to establish, so the convergence argument is circular. The paper itself notes this at the end of the proof ('provided the feedback mechanism is ideal'), which is a limitation statement but it is placed in a proof of a theorem, not as a caveat. Additionally, the algebra is wrong: from E[e_v(t+1)] = E[e_v(t)] - alpha E[RS_uw] RS*_v, the manuscript concludes E[e_v(t+1)] = (1 - alpha E[RS_uw] RS*_v) E[e(t)]. The second term is a constant, not proportional to E[e(t)], so the recurrence does not contract; it would drive the expected error to negative values, not to zero. Thus the main theoretical claim is unsupported.","section":"Section 4.5, Lemma"},{"comment":"The reputation feedback m_{j,k} = Omega(M(theta_j, theta_k)) - Omega(M(theta_j)) measures the change in peer j's local concordance when incorporating client k's update. With heterogeneous local data and censoring, an update can improve local concordance on peer j's data while degrading the global model's held-out concordance. No experiment in Section 5 measures the correlation between m_{j,k} and held-out global C-index; the reported reputation evolution plots (Figures 3-5) only show that scores change, not that they track contribution quality. Without such evidence, the central claim that the mechanism 'down-weighs noisy client updates' is not established.","section":"Sections 3.4 and 4.4"},{"comment":"The DP accuracy guarantee assumes Omega is L_Omega-Lipschitz and that the global loss is L_k-smooth. For the concordance index, which is a step-function-based ranking statistic, Lipschitz continuity is nontrivial and is not argued or referenced; the Cox partial likelihood's smoothness in the reputation-weighted objective is also not verified. The bounds in Section 4.6.2 are therefore conditional on unverified assumptions. In addition, the 'Privacy and Accuracy Guarantee' in Section 4.4 refers to Section 4.6 for proof, but the proof only provides error bounds in terms of these assumptions and does not connect to an end-to-end privacy-utility trade-off.","section":"Section 4.6"},{"comment":"The claim in the Abstract and Section 5.2.1 that the method 'consistently achieves high and stable C-index values... outperforming FL methods that lack a reputation system' is not supported by Table 2: the no-reputation baseline achieves 0.660216 in round 1, close to the proposed method's 0.661218, and TFFL reaches 0.6701 in round 7, higher than any value reported for the proposed method. The reported differences are not accompanied by confidence intervals or multiple-seed statistics, so it is unclear whether they are significant. The claim should be restated as 'on average' or supported with variance estimates.","section":"Section 5.2.1, Table 2"}],"minor_comments":[{"comment":"The threat model defines noise added to feature vectors, but the experiments state noise is injected into model parameters or reputation messages; clarify which attack surface is used in each experiment.","section":"Sections 3.3 and 5.1.2"},{"comment":"The formula for selection probability is written with an ambiguous juxtaposition; it should be P(i) = RSi(r) divided by the sum of RS_j(r) over j in C_k.","section":"Algorithm 1, line 8"},{"comment":"The caption says 'three plots' but lists four epsilon_max values; correct the count.","section":"Figure 3 caption"},{"comment":"The sentence 'ensures that clients do not gain knowledge of their peers' data or local model parameters' is too strong: clients receive DP-noised parameter vectors from peers, so they do gain information about noised parameters; suggest rephrasing to 'do not learn exact local parameters.'","section":"Section 4.4"},{"comment":"The table header 'Thonest' should be 'T_honest', and 'reputation Stability' should be 'reputation stability.'","section":"Table 1"},{"comment":"The symbol alpha is used both for the reputation learning rate in Eq. (4) and for the noise scaling in Section 3.3 (alpha_i); renaming one of the two would avoid confusion.","section":"Notation throughout"}],"recommendation":"reject","confidential_remarks":"The manuscript is not ready for publication. The main theoretical result is invalid as stated, and the core mechanism's validity is unverified. The paper has interesting components (DP decoupling, SEER experiments) but they do not rescue the central claim. I would not encourage resubmission in its current form; a major revision would require a different proof strategy and an empirical validation of the feedback signal against global held-out performance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines known building blocks — reputation-based aggregation, clustering for heterogeneity, DP for privacy — into a new artifact for federated Cox survival analysis, with a P2P reputation channel and a sensible decoupling of privatized updates for peer evaluation from unaltered updates for aggregation. That packaging is genuinely new and worth a look. The empirical section is honest about being largely illustrative: simulations on synthetic data plus one real SEER split, with TFFL and no-reputation baselines. I agree the DP accuracy bounds in Section 4.6 are standard and fine if their assumptions hold.\n\nSoft spots, in proportion. The load-bearing problem is that the entire mechanism hinges on m_{j,k} = local C-index delta tracking global contribution quality, and nowhere is that shown. The convergence lemma in Section 4.5 simply assumes feedback equals true reliability plus zero-mean noise — that is the property at issue, not a proof of it. The algebra also does not work: the recurrence gives E[e(t+1)] = E[e(t)] - alpha*RS_uw*RS*_v, not the claimed contraction; the error is pushed negative, not to zero. That is a genuine load-bearing flaw, and the stress-test note is right about it. The threat model in Section 3.3 injects feature noise, but the experiments in Section 5.1.2 inject noise into parameters/reputation/messages without connecting them. Tables 2 and 3 report single runs with no variance, and the abstract's \"outperforming\" claim is stronger than the modest, noisy differences in the tables. The clustering-concordance objective in Section 3.2 is also not clearly operationalized.\n\nIs the paper worth engaging? Yes, with the right expectations. The artifact is new and relevant to federated healthcare analytics, and the empirical setup is reproducible in principle (SEER is public, the data-generating process is described). But the central theoretical claim is unproven, and the empirical evidence is too weak to rescue it. A serious referee should ask for: (1) a corrected convergence proof or a properly stated assumption, (2) an experiment linking m_{j,k} to held-out global C-index under heterogeneity, and (3) variance/standard errors across runs. The paper deserves peer review rather than a desk reject because the question — can peer-assessed local C-index deltas safely weight federated Cox updates — is interesting and the framework is a plausible candidate. I would not cite it until the feedback-signal question is answered.","headline":"A plausible reputation-based FL/Cox framework whose central mechanism is unproven: the feedback signal is assumed to track true contribution quality, and the convergence proof in Section 4.5 is algebraically off.","tokens_in":15375,"tokens_out":633,"would_cite":false,"duration_ms":6991,"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 survival-analysis framework uses differentially privatized peer feedback to build client reputation, down-weighting noisy hospitals and preserving C-index accuracy.","keywords":["Federated Learning","Survival Analysis","Cox Proportional Hazards","Decentralized Aggregation","Healthcare Informatics","Data Heterogeneity","Privacy-Preserving Machine Learning","Concordance Index"],"falsifier":"Run the scheme under heterogeneous local datasets and compare each client's reputation trajectory with the effect of its update on a held-out global evaluation set across rounds; if clients with rising reputations consistently contribute nothing or negatively to global concordance while low-reputation clients contribute strongly, the peer-feedback signal is not faithful and the central benefit collapses. Equivalently, inject a fixed constant bias into the feedback messages $m_{j,k}$ and check whether the Section 4.6 error bounds still hold.","tokens_in":14282,"feed_emoji":"🏥","tokens_out":11350,"duration_ms":87173,"temperature":0.7,"pith_summary":"Federated survival analysis in healthcare can be made trustworthy without a central reputation authority: each client rates a peer by how much that peer's model update improves the concordance index on the client's own local data, and these ratings accumulate over rounds into reputation scores that weight the aggregation of Cox proportional-hazards models. The reputation channel is deliberately separated from the learning channel, with updates differentially privatized before they are shown to peers and unaltered updates sent to the server, so privacy protection does not degrade the global model. A clustering step groups institutions by feature completeness and risk profile to handle data heterogeneity and to filter noisy feedback. If the claim holds, hospitals can collaboratively fit survival models whose quality degrades gracefully when some participants gradually begin poisoning their updates, and no node needs to see another node's data. Reported experiments on synthetic data and the SEER breast-cancer registry show stable, high concordance-index values and clear down-weighting of noisy clients relative to reputation-free federated averaging.","feed_headline":"Peer ratings shield federated survival models from noisy nodes","feed_subtitle":"Peer-reviewed updates stay private, and noisy centers get down-weighted as their reputations fall.","key_machinery":"The load-bearing object is the peer-feedback message $m_{j,k}(t)=\\Omega(M(\\theta_j(t),\\theta_k(t)))-\\Omega(M(\\theta_j(t)))$, in which client $j$ scores client $k$'s update by the change in concordance index $\\Omega$ of a Cox model evaluated on $j$'s local data with and without $k$'s update. This message, weighted by the sender's reputation $RS_{ij}(t)$, drives the update $RS_{ik}(t+1)=RS_{ik}(t)+\\alpha\\sum_j RS_{ij}(t)m_{j,k}(t)$, and the resulting scores set client-selection probabilities $P(i)=RS_i/\\sum_{j\\in C_k}RS_j$ inside each cluster. The privacy half of the mechanism clips updates to norm $Q$ and adds Gaussian noise $\\sigma=Q\\sqrt{2\\ln(1.25/\\delta)}/\\zeta$ before peer evaluation, while unperturbed updates flow to the server. Clustering by feature-completeness vectors and predicted risk, with a concordance term weighted by $\\lambda$, turns peer evaluations into cluster-specific, noise-resistant aggregation.","core_discovery":"Working with the Cox proportional-hazards model, the central claim is that a reputation score built from peer feedback, specifically the change in a peer's local concordance index when a client's update is added, can serve as a reliable aggregation weight in heterogeneous federated learning, as long as the feedback itself is reliable. The paper further claims that differential privacy can be applied to client updates before peer evaluation without materially changing the outcome: the expected errors induced by DP noise in reputation feedback, reputation scores, client-selection probabilities, and the aggregated global model are each bounded by explicit constants involving the privacy parameters. A formal lemma states that, under zero-mean feedback noise, estimated reputations converge to the true reliability of each peer over rounds. Empirically, the method keeps the global C-index near 0.66 across ten synthetic rounds, beats the no-reputation baseline consistently, and matches or exceeds the reputation-based TFFL baseline on most SEER states.","pith_inferences":["Looking beyond the paper, the same DP decoupling should transfer to other federated objectives: any secondary client score whose value changes smoothly with the update could be computed on a privatized channel while the clean channel serves the global objective.","The convergence lemma assumes zero-mean feedback noise; the paper's static-bias experiment suggests resilience but does not prove it, so a natural extension is to derive reputation-error bounds under fixed non-zero mean bias.","Whether the mechanism helps the global model ultimately hinges on local concordance gains tracking global contribution; a direct test is to correlate each client's reputation trajectory with its effect on a held-out global evaluation set after every round.","One could also apply the reputation-weighted, cluster-selection scheme to other survival losses, such as discrete-time or deep survival models, where the same concordance-based feedback and DP decoupling would carry over."],"forward_implications":["A hospital that starts injecting noise after an honest warm-up sees its reputation decline over rounds, so its influence on the shared survival model shrinks.","Reputation can be computed without any central trusted dataset, because each evaluation uses only local concordance-index changes and privatized updates.","The shared Cox model can reach accuracy comparable to reputation-free federated averaging while keeping each client's update private from its peers.","Institutions with similar feature-completeness and risk profiles can be clustered to improve concordance under data heterogeneity.","Reputation updates can be throttled to cut message overhead, at the cost of some accuracy, so an intermediate update frequency balances communication and performance."],"supporting_citations":[{"why":"Supplies the reputation-free Federated Averaging baseline that the method must outperform.","marker":"[1]"},{"why":"Supplies the Trustworthy and Fair Federated Learning baseline with dynamic reputation-based consensus.","marker":"[9]"},{"why":"Defines the Cox proportional-hazards model that each client trains and shares.","marker":"[12]"},{"why":"Defines the concordance index used as the performance metric $\\Omega$ in the feedback messages.","marker":"[19]"},{"why":"Supplies the peer-feedback formula $m_{j,k}=\\Omega(M(\\theta_j,\\theta_k))-\\Omega(M(\\theta_j))$ that the reputation update builds on.","marker":"[20]"},{"why":"Supplies the differentially private SGD mechanism motivating the privatized client updates.","marker":"[10]"},{"why":"Supplies the Gaussian mechanism noise calibration $\\sigma=Q\\sqrt{2\\ln(1.25/\\delta)}/\\zeta$ used for the DP protections.","marker":"[11]"},{"why":"Supports the client-level differential privacy accuracy bounds used in Section 4.6.","marker":"[24]"}],"fun_headline_variants":["Peer reputation scores down-weight noisy nodes in federated survival","Federated survival models benefit from peer-driven reputation","Privacy-preserving peer feedback improves federated survival analysis","Peer-reviewed updates keep federated survival models robust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when a client's update improves a peer's local concordance index, that update is genuinely helpful to the global survival model; if local gains diverge from global contribution, reputation weighting could reward exactly the wrong clients.","fun_headline_variants_meta":{"raw":{"variants":["Peer reputation scores down-weight noisy nodes in federated survival","Federated survival models benefit from peer-driven reputation","Privacy-preserving peer feedback improves federated survival analysis","Peer-reviewed updates keep federated survival models robust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1432,"prompt_tokens":921,"completion_tokens":511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":537,"tokens_out":511,"duration_ms":4581,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:05:45.982649+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the scheme under heterogeneous local datasets and compare each client's reputation trajectory with the effect of its update on a held-out global evaluation set across rounds; if clients with rising reputations consistently contribute nothing or negatively to global concordance while low-reputation clients contribute strongly, the peer-feedback signal is not faithful and the central benefit collapses. Equivalently, inject a fixed constant bias into the feedback messages $m_{j,k}$ and check whether the Section 4.6 error bounds still hold.","supporting_citations":[{"cited_title":"Trustworthy and fair federated learning via reputation-based consensus and adaptive incentives,","cited_arxiv_id":null,"evidence_quote":"Supplies the Trustworthy and Fair Federated Learning baseline with dynamic reputation-based consensus."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the concordance index used as the performance metric $\\Omega$ in the feedback messages."},{"cited_title":"Federated learning using peer-to-peer network for decentralized orchestration of model weights,","cited_arxiv_id":null,"evidence_quote":"Supplies the peer-feedback formula $m_{j,k}=\\Omega(M(\\theta_j,\\theta_k))-\\Omega(M(\\theta_j))$ that the reputation update builds on."},{"cited_title":"Calibrating noise to sensitivity in private data analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian mechanism noise calibration $\\sigma=Q\\sqrt{2\\ln(1.25/\\delta)}/\\zeta$ used for the DP protections."}],"review_version":1}