{"id":"854ad81a-3092-41a2-aa23-3fbf41e4d428","arxiv_id":"2504.20250","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Federated logistic regression with median or trimmed-mean aggregation reaches AUC close to centralized logistic regression on four public financial datasets, but not close to decision trees or k-nearest neighbors on all of them.","lead":"This paper applies federated logistic regression with median or trimmed-mean averaging to financial classification datasets, aiming for privacy, interpretability, and outlier resistance. A generalist might read it to see whether simple linear models can match centralized baselines in distributed finance settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own tables contradict the headline claim: on LoanPred, FLR AUC 0.53 vs DT 0.85 / KNN 0.87, so 'comparable to DT and KNN' is unsupported.","rationale":"The stress-test pass looked for the most load-bearing weakness in the central comparative claim. The claim is empirical and universal over the listed baselines, so its minimal support is FLR results within error of the baselines across all datasets. Table 4 versus Table 3 shows this holds against centralized LR everywhere and against DT/KNN on BankMarketing, CreditScore, and CreditRisk AUC, but it fails sharply on LoanPred for DT and KNN. One decisive counterexample in the authors' own results is sufficient to refute the unqualified claim, even before considering the Eq. (5) derivation error or the known-pout precondition for trimmed mean. The reader's REJECT verdict therefore stands without modification. Agreement is marked partial because the reader's stated weakest assumption concerned the trimmed-mean pout requirement, while the contradiction on LoanPred appears in the reader's rationale but not in the weakest-assumption field. The recommendation is UNCHANGED: the central claim is unsupported by the paper's own evidence.","tokens_in":17825,"tokens_out":4944,"duration_ms":47867,"concrete_test":"Obtain the repository/datasets promised at 'FLR' in Section 5 and rerun the no-outlier, IID, M=100 LoanPred FLR-mean experiment exactly as described in Section 5.2 on the same train/test split used for Table 3, then compare AUC and ACC with the centralized DT and KNN rows. If the reproduced FLR AUC is approximately 0.53 versus 0.85–0.87, the central claim is not supported; if the reproduced numbers differ materially from Table 4, the reported evidence is unreliable. If the repository is unavailable, a minimal independent implementation of FedAvg with per-client scikit-learn logistic regression under the paper's protocol would settle the point.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that federated logistic regression is comparable to centralized LR, DT, and KNN in both binary and multi-class tasks. The evidence for this is Table 4 (FLR without outliers) against Table 3 (centralized baselines). On LoanPred, FLR-mean AUC is 0.53 ± 0.01 under both IID and non-IID splits, while centralized DT and KNN achieve 0.85 ± 0.00 and 0.87 ± 0.01; ACC is 0.49 vs 0.88/0.89. Centralized LR also gives AUC 0.53 on LoanPred, so FLR matches LR on this dataset, but the abstract and conclusion explicitly list DT and KNN as algorithms to which the method is comparable. A ~0.34 AUC gap is far beyond the reported confidence intervals, so no reasonable reading of 'comparable' covers it. Since the claim covers both binary and multi-class tasks unqualified, failure on one of the two binary datasets is decisive. This is an internal inconsistency, not a disagreement with external consensus. The incorrect log-likelihood in Eq. (5) and the pout dependence of trimmed mean are real secondary issues, but the headline claim is already contradicted by the paper's own experimental results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a federated logistic regression (FLR) framework for financial data analysis, with three server-side aggregation rules: mean, coordinate-wise median, and trimmed mean. The method is evaluated on four public financial datasets (two binary, two multi-class) under IID and non-IID client partitions, both without outliers and with simulated outlier clients. The results are compared against centralized logistic regression (LR), decision trees (DT), and k-nearest neighbors (KNN). The central claim is that FLR achieves comparable classification performance to these centralized algorithms and remains robust to outliers when median or trimmed-mean aggregation is used. The paper also provides feature-importance analysis and makes code and datasets publicly available.","tokens_in":18074,"tokens_out":6010,"duration_ms":56801,"significance":"If the performance-parity claim were supported, the paper would provide a useful practical result: a simple, interpretable linear model can be trained in a federated fashion without a major accuracy loss relative to centralized non-linear models, while robust aggregation offers protection against malicious clients. The experimental design is straightforward and the reported metrics (ACC, F1, AUC with means and 1.96-sigma intervals) are appropriate. The authors also commit to releasing code and data, which is commendable for reproducibility. However, the paper's headline claim is contradicted by its own experimental results on one of the two binary datasets, and the mathematical formulation of the likelihood contains a serious error. These issues prevent the stated significance from being realized without substantial revision.","major_comments":[{"comment":"The abstract and conclusion claim 'comparable performance' to LR, DT, and KNN in both binary and multi-class tasks. This is contradicted by the paper's own Tables 3 and 4 on the LoanPred dataset. For LoanPred, all FLR variants in Table 4 achieve AUC 0.53 ± 0.01 under both IID and non-IID partitions, while centralized DT and KNN in Table 3 achieve 0.85 ± 0.00 and 0.87 ± 0.01, respectively. This is a 0.32–0.34 gap in AUC, far outside the reported confidence intervals. Since the claim is unqualified and covers both binary datasets, the failure on one of them is decisive. The text in Section 5.3 that explicitly states 'our proposed framework can achieve comparable performance in terms of AUC with the baseline performance obtained by LR, DT and KNN' is therefore not supported by the reported results.","section":"Abstract and Conclusion vs. Tables 3–4"},{"comment":"The log-likelihood expression in Eq. (5) is mathematically incorrect. For independent Bernoulli outcomes, the log-likelihood should be ln L = Σ_i [Y_i ln p_i + (1 − Y_i) ln(1 − p_i)]. The paper instead writes ln(L) = Σ_i [Y_i p + (1 − Y_i)(1 − p)], omitting the logarithms and using a single scalar p without an index. Moreover, Eq. (6) sets up an 'argmin' of the log-likelihood, whereas maximum likelihood estimation maximizes it. Although the implementation described later uses the cross-entropy loss, the stated derivation is invalid and should be corrected or removed.","section":"Section 3.2.1, Eqs. (5)–(6)"},{"comment":"The robustness demonstration for FLR-trim_mean depends on knowledge of the outlier fraction. Section 3.2.1 states that the trimming fraction α should satisfy α ≥ pout/2, where pout is the percentage of adversarial clients. In the outlier experiments of Section 5.3, pout is varied and the trimming fraction is evidently chosen using the known pout. The paper provides no procedure for estimating pout in practice, nor does it report the exact α values used. Consequently, the robustness result for the trimmed-mean variant is only established under an oracle assumption about the attack strength. The FLR-median results do not have this limitation, but the paper's general robustness claim covers both robust aggregation methods without qualification.","section":"Section 3.2.1 and Section 5.3 (Federated LR with Outliers)"}],"minor_comments":[{"comment":"The phrase 'Logistical Regression' should be 'Logistic Regression'.","section":"Abstract"},{"comment":"The equation label 'AUCmarco' contains a typo; it should be 'AUCmacro'.","section":"Section 4.2"},{"comment":"The text says the outlier percentage pout is varied 'from 0 to 0.2 with a step size of 0.5'; the step size should presumably be 0.05 to cover the stated range.","section":"Section 5.2"},{"comment":"The paper states that each experiment is conducted '10 times' using seeds 'spanning from 0 to 1000, incremented by 100', which yields 11 seeds. Please align the number of runs with the seed list.","section":"Section 5.2"},{"comment":"The pseudocode indentation is unclear: the 'if t % 10 == 0' block appears to enclose the client selection, initialization, and server update, implying updates occur only every 10 iterations. This contradicts the textual description in Section 3.2.1, where the server updates in every iteration. Please redraw the algorithm with explicit indentation.","section":"Algorithm 1"},{"comment":"There are several typographical and grammatical errors, e.g., 'dta' for 'data', 'we emphasis on' for 'we emphasize', and 'the number of iterationsTc' in Algorithm 1 is not used in the pseudocode. A careful proofread is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper appears to be an early draft with multiple presentation issues. The central claim is contradicted by the authors' own results, and the likelihood derivation in Section 3.2.1 is mathematically wrong. Although the robustness experiments are a useful empirical contribution, the trimmed-mean result depends on knowing the attack rate. These are substantial problems, not local typos. I would not rule out the possibility of a substantially revised version that honestly reframes the comparison and corrects the derivation, but as submitted the manuscript is not suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward application of known robust aggregation rules (coordinate-wise median, trimmed mean) to federated logistic regression, evaluated on four public financial datasets. The experiments are clean enough, but the abstract's claim that FLR is \"comparable\" to Decision Tree and K-Nearest Neighbors is contradicted by the paper's own Table 4 on LoanPred (AUC 0.53 for FLR vs 0.85 and 0.87 for DT and KNN), and the likelihood derivation in Eq (5) has a real mathematical error.\n\nWhat's actually useful: a careful, transparent benchmark of federated logistic regression with three aggregation strategies on IID and non-IID data, with and without outliers, plus a feature-importance inspection. The no-outlier results sensibly match centralized logistic regression, which is the correct sanity check. The robustness behavior of median vs trimmed mean is reported honestly, with standard deviations across repeated runs.\n\nThe main problems, in order of severity. First, the headline claim overreaches. FLR matches centralized LR on all four datasets, and beats DT/KNN on two of them, but it badly lags DT/KNN on LoanPred. Since the abstract promises \"comparable\" performance \"in both binary and multi-class classification tasks,\" one unambiguous failure among the two binary datasets is enough to sink the claim as stated. Narrowing the claim to \"comparable to centralized logistic regression\" would be accurate. Second, Eq (5) writes the log-likelihood as sum_i [Y_i p + (1-Y_i)(1-p)], which is not the log-likelihood; the correct form is Y_i log p + (1-Y_i) log(1-p). This is a derivation error and should be fixed even if the implementation uses the right cross-entropy loss. Third, the trimmed-mean robustness result depends on knowing the outlier fraction pout to set the trimming fraction alpha; the paper states this in Section 3.2.1, but it limits the practical claim. Fourth, the paper says the code and data are public but gives no working link in the arXiv version, and the related work gives little credit to the Byzantine-robust federated learning literature where these aggregation rules were established.\n\nThese are all fixable in revision. The novelty is modest, but the benchmark could be a useful reference for practitioners who want a simple, interpretable federated baseline in finance. I would send it to peer review with the expectation of heavy revision; if an editor wants to desk-reject on novelty grounds, that is defensible, but the experiments are solid enough to justify referee time.","headline":"Known robust federated aggregation rules applied to logistic regression for finance; useful benchmark, but the claim of comparability to DT/KNN is contradicted by the paper's own LoanPred results and Eq (5) has a real mathematical error.","tokens_in":18612,"tokens_out":3463,"would_cite":false,"duration_ms":33028,"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":"Federated logistic regression with median or trimmed-mean aggregation matches centralized logistic regression, decision trees, and k-nearest neighbors in AUC on four financial datasets, retaining that performance when a fifth of clients…","keywords":["Federated Learning","Logistic Regression","Robust Aggregation","Financial Data Analysis","Outlier Robustness","Non-IID Data","AUC","Interpretability"],"falsifier":"On the BankMarketing dataset with $M=100$ clients, sample size $s=100$, and 10% of clients corrupted, run FLR-trim_mean with a fixed trimming fraction $\\alpha=0.05$ instead of one chosen from the true corruption rate. If AUC drops by more than about 0.02 below the no-outlier value of 0.83, the paper's rule $\\alpha \\ge p_{\\mathrm{out}}/2$ is load-bearing and the practical robustness claim is only conditional.","tokens_in":17571,"feed_emoji":"📊","tokens_out":13570,"duration_ms":130134,"temperature":0.7,"pith_summary":"The paper aims to show that financial classification can be done well without ever pooling raw customer data. It proposes a federated logistic regression framework in which each client trains a local linear model and sends only coefficient vectors to a central server. Across four public datasets covering loan default, deposit subscription, credit scoring, and credit risk, the authors report AUC comparable to centralized logistic regression, decision tree, and k-nearest-neighbor baselines for both binary and multi-class tasks. They further claim that replacing the usual average of client updates with a coordinate-wise median or a trimmed mean keeps that AUC stable when up to 20% of clients are corrupted, while plain averaging fails. If these results hold, a simple and interpretable linear model can deliver privacy, outlier robustness, and centralized-level accuracy on the same financial problems.","feed_headline":"Federated logistic regression matches centralized baselines","feed_subtitle":"It matches centralized AUC while keeping raw data local and surviving corrupted clients.","key_machinery":"The load-bearing mechanism is server-side aggregation of client coefficient vectors. After local logistic-regression updates, the server either averages the per-coordinate values, takes their per-coordinate median, or removes the top and bottom $\\alpha$ fraction of each coordinate before averaging (trimmed mean). In the robust variants, one corrupted local model cannot shift the global model the way it shifts a mean, and the paper sets $\\alpha \\ge p_{\\mathrm{out}}/2$ so that trimming removes at least the corrupted updates. For multi-class problems the same machinery is repeated for each one-vs-rest binary classifier, and the client only transmits $C$ coefficient vectors per round.","core_discovery":"On the paper's own terms, the discovery is that a straightforward federated linear model is enough to reproduce centralized performance on these financial benchmarks. With no outliers, all three aggregation strategies (mean, median, trimmed mean) give nearly identical results, matching centralized logistic regression's AUC on BankMarketing, LoanPred, and CreditScore within 0.01 and on CreditRisk within 0.02, under both IID and non-IID data splits. When outliers are injected, mean aggregation degrades as the corruption rate grows, while median and trimmed-mean aggregation keep AUC largely unchanged up to $p_{\\mathrm{out}} = 0.2$, with median slightly ahead. The multi-class version trains one-vs-rest binary models in parallel, so communication grows linearly with the number of classes. The coefficients themselves serve as interpretable feature-importance reports.","pith_inferences":["Beyond the reported experiments, a natural stress test is to run the same protocol on raw, unfiltered financial features; the paper first removes highly correlated features, and that preprocessing may be hiding part of the advantage non-linear centralized models would have.","Because the trimmed-mean rule requires knowing the corruption fraction, an untested practical variant is to set the trimming fraction adaptively from the spread of client updates; the reported median results suggest per-coordinate median may already give this robustness without tuning.","The protocol's small payload (one coefficient vector per class per round) suggests it could scale to much larger client populations than the 200-client maximum tested here, provided communication is the bottleneck.","Rerunning the released code with a temporal split—training on earlier months and testing on later ones—would indicate whether the centralized-level AUC is a property of the datasets or of the method itself."],"forward_implications":["Financial institutions can train a classification model with centralized-level AUC while raw customer data never leaves local devices; only coefficient vectors cross the network.","Median aggregation is a simple defense against outlier clients: it preserves AUC up to 20% corruption without requiring the server to know the exact corruption rate.","Multi-class credit-scoring tasks fit the same framework by training one-vs-rest logistic models in parallel, at a communication cost that grows linearly with class count.","Coefficient-based feature importance is available directly from the federated model, so decisions can be explained even though training is distributed.","Non-IID client data degrades performance only slightly relative to IID data, so the method does not depend on a uniform partition of the population."],"supporting_citations":[{"why":"Supplies the BankMarketing deposit-subscription dataset used in the binary benchmarks.","marker":"[48]"},{"why":"Supplies the LoanPred consumer-loan dataset used in the binary default-prediction benchmarks.","marker":"[49]"},{"why":"Supplies the CreditScore dataset used in the three-class credit-scoring benchmarks.","marker":"[50]"},{"why":"Supplies the CreditRisk loan dataset used in the six-class credit-risk benchmarks.","marker":"[51]"},{"why":"Defines the federated-learning protocol in which clients send model updates rather than raw data.","marker":"[13]"},{"why":"Gives the logistic-regression logit formulation that the local and global models are built on.","marker":"[35]"},{"why":"Provides the one-vs-rest scheme used to extend binary federated logistic regression to multi-class tasks.","marker":"[40]"},{"why":"Defines AUC, the primary metric used to compare federated and centralized classifiers.","marker":"[47]"}],"fun_headline_variants":["Federated logistic regression robust to outliers","Median aggregation shields regression from outliers","Private regression, robust to corrupted clients","Federated LR: central accuracy, outlier resistance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trimmed-mean robustness result assumes the server knows roughly how many clients are corrupt, because the trimming fraction is set from that number; without that knowledge the reported outlier resilience is not established, although the median-aggregation results do not need that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Federated logistic regression robust to outliers","Median aggregation shields regression from outliers","Private regression, robust to corrupted clients","Federated LR: central accuracy, outlier resistance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000647,"raw_usage":{"total_tokens":2939,"prompt_tokens":877,"completion_tokens":2062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":2008}},"tokens_in":493,"tokens_out":2062,"duration_ms":20080,"temperature":1.0,"reasoning_tokens":2008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:34:04.325604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the BankMarketing dataset with $M=100$ clients, sample size $s=100$, and 10% of clients corrupted, run FLR-trim_mean with a fixed trimming fraction $\\alpha=0.05$ instead of one chosen from the true corruption rate. If AUC drops by more than about 0.02 below the no-outlier value of 0.83, the paper's rule $\\alpha \\ge p_{\\mathrm{out}}/2$ is load-bearing and the practical robustness claim is only conditional.","supporting_citations":[{"cited_title":"banking institution, Bank marketing, https://archive.ics.uci.edu/dataset/222/bank+marketing, accessed: 2023-08-01 (2014)","cited_arxiv_id":null,"evidence_quote":"Supplies the BankMarketing deposit-subscription dataset used in the binary benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LoanPred consumer-loan dataset used in the binary default-prediction benchmarks."},{"cited_title":"global financial company, Credit score classification,https://www.kaggle.com/datasets/parisrohan/ credit-score-classification, accessed: 2021-02-01 (202x)","cited_arxiv_id":null,"evidence_quote":"Supplies the CreditScore dataset used in the three-class credit-scoring benchmarks."},{"cited_title":"global financial company, Credit risk dataset, https://www.kaggle.com/datasets/ranadeep/ credit-risk-dataset, accessed: 2021-02-01 (202x)","cited_arxiv_id":null,"evidence_quote":"Supplies the CreditRisk loan dataset used in the six-class credit-risk benchmarks."},{"cited_title":"Konecny, H","cited_arxiv_id":null,"evidence_quote":"Defines the federated-learning protocol in which clients send model updates rather than raw data."},{"cited_title":"Maalouf, Logistic regression in data analysis: an overview, International Journal of Data Analysis Techniques and Strategies 3 (3) (2011) 281–299","cited_arxiv_id":null,"evidence_quote":"Gives the logistic-regression logit formulation that the local and global models are built on."},{"cited_title":"Galar, A","cited_arxiv_id":null,"evidence_quote":"Provides the one-vs-rest scheme used to extend binary federated logistic regression to multi-class tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines AUC, the primary metric used to compare federated and centralized classifiers."}],"review_version":1}