{"id":"d8b87e5d-1350-4515-ae20-d94d5f8caaad","arxiv_id":"2502.01532","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A federated weighted Naive Bayes method that shares only discriminative weights, not probability tables, matches or exceeds federated generative Naive Bayes accuracy on 12 discrete datasets.","lead":"This paper introduces a federated version of weighted Naive Bayes in which clients share only discriminative weights, while each client keeps its own probability tables local. In experiments on 12 discrete datasets, the method matches or beats federated generative Naive Bayes, and the authors argue that the shared weights are meaningless on their own.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central privacy claim rests on the unsupported assertion that shared discriminative weights are 'meaningless'; no threat model or formal bound is given, and the conclusion itself lists differential privacy as future work.","rationale":"The reader's weakest assumption is the same as the central concern: the privacy advantage is assumed, not demonstrated. The strongest technical claim, that NBw5fed is accurate, is supported by broad experiments, though missing error bars and an unspecified T; those are secondary. The unsupported privacy claim is load-bearing because the paper's motivation and contribution statements in the abstract, Section 2.3, and conclusion all advertise 'more reliable against possible attacks.' The paper itself flags the absence of a privacy guarantee by listing differential privacy as future work. Therefore I agree with the conditional verdict: acceptance should require either removing the privacy claims or substantiating them with a formal bound or attack evaluation. No change to the reader's verdict is needed.","tokens_in":18445,"tokens_out":5249,"duration_ms":49750,"concrete_test":"Set up a single synthetic client with known class counts and conditional feature counts, then run the released NBw5fed implementation for T rounds, recording the broadcast global weights and the client's local updates. Attempt to recover the client's counts from the sequence of weight updates, either by solving the CLL-gradient equations for counts (assuming the attacker knows the algorithm and, in the strongest attack, the local probability tables) or by a black-box model-inversion attack. If counts are recovered to within 10% relative error, the 'inherently meaningless' privacy claim is falsified; if recovery fails even with full knowledge of the algorithm and local tables, the privacy claim gains empirical support. Additionally, compare against recovering the same counts from directly shared probability tables to quantify any actual privacy gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's advertised advantage is that NBw5fed shares only 'meaningless' weights rather than probability tables (Sections 2.3 and 5, Abstract), making it 'more reliable against possible attacks.' This is asserted, not derived. No threat model, reconstruction attack, or privacy bound is provided, and Section 5 concedes that differential privacy would be needed 'to strengthen the privacy guarantees,' acknowledging that none currently exist. The shared artifacts are not arbitrary: each client sends w_c obtained by L-BFGS-M optimization of the conditional log-likelihood on local data. The CLL gradient with respect to w_{y,x_j} depends directly on local counts, with terms of the form #(x_j, y) log θ_{y,x_j} minus model expectations (see Section 2.2). An attacker with the public algorithm and observed weight updates could plausibly invert these to recover local class-conditional statistics. If that is possible, the privacy advantage over sharing perturbed probability tables is unsupported. This concern is independent of accuracy: NBw5fed may be accurate while its privacy claim fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Federated Weighted Naive Bayes (NBwfed), a federated learning algorithm for discrete discriminative Naive Bayes. In the proposed Algorithm 1, each client learns local conditional probability tables in a generative step and keeps them private, then participates in a federated discriminative step in which only weight vectors of a weighted NB (NBw) model are optimized locally, averaged on the server, and broadcast back. The authors compare NBwfed with non-federated NB and NBw, and with a federated generative NB baseline (NBfed) on 12 OpenML datasets across 5, 10, 20, 50, and 100 clients, using 5-fold cross-validation repeated over 5 seeds. They report that the variant with a 5-iteration L-BFGS-M limit and global weights (NBw5fed) generally achieves the best accuracy, except that NBfed performs better at large client counts. The central advertised advantage of the method is privacy: sharing 'meaningless' weight parameters rather than conditional probability tables is claimed to be more reliable against possible attacks.","tokens_in":18658,"tokens_out":7970,"duration_ms":71343,"significance":"If both the accuracy and privacy claims were substantiated, the paper would make a useful contribution to federated learning for explainable, lightweight classifiers: it demonstrates that a discriminatively trained NB can be federated by exchanging only weights, and it provides broad experimental evidence (12 datasets, 5 client counts, repeated runs) plus released code and OpenML dataset identifiers. The experimental design is commendable in scope and reproducibility, and the comparison with generative federated NB is a sensible baseline. However, the privacy claim is asserted rather than proven, and the statistical reporting is too thin to support the main comparative accuracy statement. The algorithmic description also leaves important implementation details (e.g., zero-count handling and the federated round count) unspecified. The useful, defensible core of the paper is the empirical exploration of a weight-sharing federated NB; the privacy advantage is currently a motivation, not a demonstrated property.","major_comments":[{"comment":"The paper's central privacy claim—that the shared discriminative weight vectors are 'inherently meaningless in revealing sensitive information'—is asserted without a threat model, an attack analysis, or any formal privacy bound. Each client transmits w_c^{(t)} obtained by L-BFGS-M optimization of the conditional log-likelihood on local data, and the CLL gradient with respect to w_{y,x_j} depends directly on local sufficient statistics such as #(x_j,y) (see §2.2), so the shared weights can encode information about local class-conditional counts. The conclusion (Section 5) lists differential privacy as future work, implicitly conceding that no quantifiable privacy guarantee is currently provided. The authors should either give a formal privacy analysis under an explicit adversary model, or substantially weaken the privacy claims and present the method as a communication-efficient alternative to sharing probability tables, with privacy as a motivating direction for further study.","section":"§2.3 and Abstract"},{"comment":"The experimental protocol (5-fold CV, 5 seeds, 25 runs) is described, but Table 2 and Figures 1–3 report only mean accuracy, with no standard deviations, confidence intervals, or significance tests. The statement that 'NBw5fed outperforms all other algorithms, except in scenarios with a large number of clients' rests on small mean differences (e.g., Table 2, 5 clients, House Votes: NBw5g 94.85 vs NBw5l 94.29), which may be within run-to-run noise. The authors should report variance and apply paired tests across datasets (e.g., Wilcoxon signed-rank or a sign test on per-dataset means) to support the cross-dataset conclusions.","section":"§4.2, Table 2, Figures 1–3"},{"comment":"Algorithm 1 depends on the federated round count T, but the experiments never state the value of T; Table 2 reports accuracy at 'the last iteration' without specifying how many iterations were performed. This is a load-bearing configuration choice: the behavior of the 5-iteration-limited variant cannot be fully interpreted without knowing the number of federated rounds, and the communication cost of the method is not quantified. Please report T (and any stopping criterion) for each experimental setting.","section":"§4.2, Algorithm 1"},{"comment":"The CLL objective for NBw contains terms of the form log θ_{y,x_j}, which are undefined when a local probability table contains zeros. The paper does not state whether Laplace smoothing or another zero-count correction is applied in the generative step or inside the L-BFGS-M optimization. This is particularly relevant because the authors explicitly discuss scenarios 'where most parameters are zero' (§4.3). Without a description of zero-count handling, the algorithm's behavior on small or sparse client partitions is not well-defined and the reported results are not fully reproducible.","section":"§2.2 and §4.3"}],"minor_comments":[{"comment":"There is a naming inconsistency: Algorithm 1 is titled 'FedNBw' while the text, tables, and figures use 'NBwfed'; please unify the notation.","section":"§3"},{"comment":"Line 7 contains a typo: 'Θ_i,y_c,x_c_j' should be 'Θ_c,y_c,x_c_j', and the subscript x_c_j is ambiguous; clarify that it denotes the j-th feature value of instance x_c.","section":"Algorithm 1, line 7"},{"comment":"The row for 'House Votes 84' is misaligned: the columns read 'House Votes 84 435 16 2 56', which suggests m=84 though the dataset has 435 instances; correct the table formatting.","section":"Table 1"},{"comment":"In the version of the manuscript provided, the figure text appears as garbled glyph sequences rather than readable labels; ensure that fonts are embedded and legends are legible in the final PDF.","section":"Figures 1–3"},{"comment":"The NBw probability formula writes θ_{y_k}^{w_y} and θ_{y_k,x_j}^{w_y,x_j} in the denominator without clarifying whether these are powers or subscripts; the exponent notation should be defined explicitly.","section":"§2.2"},{"comment":"Reference [12] contains a typo ('Proocedings'); please correct it and verify the conference proceedings venue.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The disclaimer at the end indicates this contribution has already appeared as an LNCS proceedings paper (DOI 10.1007/978-3-031-77738-7_27). If this submission is intended as a journal extension, the editor may wish to evaluate the added value over the conference version, particularly since the privacy claim—the main selling point—remains unsupported in the present text. The strongest path to revision would be to add a concrete adversarial analysis of the shared weights, or to reframe the contribution as an empirical accuracy study with explicitly stated privacy limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: it takes weighted Naive Bayes (NBw), keeps the conditional probability tables local, and federates only the discriminative weights. That is a new combination, even if all the building blocks are pre-existing. The algorithm is clearly described, and the authors ship code and OpenML identifiers, so the work is reproducible.\n\nThe experimental side is broader than usual for this niche: 12 datasets, client counts from 5 to 100, and 25 runs per configuration. The accuracy trend is consistent and believable: NBw5fed is competitive or better than generative federated NB when clients are few, and degrades as partitioning fragments the data. The decision to limit L-BFGS-M to 5 iterations helps in the federated setting, and the paper says so without over-hyping it. I have no quarrel with the central accuracy claim.\n\nThe soft spots are where the reader's report points. The privacy argument is the weakest load-bearing element. Calling the shared weights \"meaningless\" is an assertion, not a result. Section 2.3 says it, the abstract says it, and the conclusion repeats it, but there is no threat model, no reconstruction attack, and no formal bound. The final section even lists differential privacy as future work, which admits the current guarantees are informal. The stress-test note is right that this is concerning: the weights are produced by optimizing conditional log-likelihood on local data, so the gradients and the resulting weights depend on local counts. An attacker with the public algorithm could plausibly invert them to recover class-conditional statistics. If that holds, the privacy advantage over perturbed probability tables dissolves. This is independent of accuracy, but it is the paper's main advertised benefit.\n\nSecond, the experimental reporting lacks variance. The protocol says 25 runs, but Table 2 and the figures report means only, with no significance tests. The choice of L=5 looks post-hoc, and the number of federated rounds T is never stated. That is sloppy but fixable. Also, the claim that federated versions consistently beat non-federated versions is strange and under-explained; since partitioning normally hurts, a sentence or two on why regularization from limited iterations flips it would help.\n\nWho gets value from this: researchers working on federated learning with simple, explainable, discrete-variable models. It is a modest addition, not a breakthrough. But it is honest, reproducible, and the central accuracy claim is likely correct. It deserves a serious referee, with the expectation of major revision: either demonstrate the privacy claim under a defined attacker model, or drop the emphasis to \"sharing fewer parameters\" and leave privacy out.\n\nI would send it to review.","headline":"A plausible, clearly-written federated extension of weighted Naive Bayes with solid experimental breadth, whose privacy selling point is asserted rather than demonstrated.","tokens_in":19185,"tokens_out":1784,"would_cite":true,"duration_ms":18249,"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":"This paper proposes a federated Naive Bayes variant that shares only discriminative weight parameters, not probability tables, and reports that it outperforms generative federated baselines on 12 datasets except when the number of clients…","keywords":["federated learning","Naive Bayes","discriminative learning","weighted naive Bayes","privacy-preserving machine learning","Bayesian network classifiers"],"falsifier":"Run FedNBw on synthetic datasets where the true class priors and feature-conditional counts are known, give an attacker the sequence of per-round weight vectors (or even the final global weights), and test whether the attacker can recover those counts or identify whether a specific record was in a client's data with better-than-chance accuracy; if any such reconstruction succeeds, the advertised privacy advantage is false. A simpler check is to compare the weight trajectories produced by two datasets that differ only in class priors: if the trajectories diverge systematically, the weights carry information about the very quantities the paper says stay private.","tokens_in":18253,"feed_emoji":"🛡️","tokens_out":6232,"duration_ms":52371,"temperature":0.7,"pith_summary":"This paper tries to establish that a discriminatively trained Naive Bayes classifier can be federated by sharing only the learned weight vector, with each client keeping its own probability tables local. The proposed scheme, NBw5fed, combines generative estimation of local tables with federated discriminative optimization of a weight vector, and on 12 discrete datasets it achieves higher test accuracy than the non-federated discriminative model and the generative federated baseline in most client-count settings. The intended payoff is privacy: because only opaque weights cross the network, the method is claimed to be more reliable against attacks than sharing conditional probability tables. The main caveat is that the generative federated baseline performs better when the number of clients is very large.","feed_headline":"Federated Naive Bayes shares only weights, keeps accuracy","feed_subtitle":"A discriminative weight-sharing variant outperforms the table-sharing baseline on 12 datasets, except with many clients.","key_machinery":"The central mechanism is the Weighted Naive Bayes (NBw) hybrid, which first estimates the generative probability tables $\\theta$ and then learns a weight vector $w$, one weight per parameter, by maximizing conditional log-likelihood. The federated algorithm FedNBw wraps this in a federated averaging loop: clients keep $\\theta$ local, synchronize on the global $w$, take a bounded number of L-BFGS-M steps against local data, and the server averages the returned weight vectors. The five-iteration inner limit is the load-bearing tuning choice: it prevents the local discriminative step from overfitting sparse client data, and the paper shows it is what lets the federated weight-sharing variant beat both the unconstrained local model and the generative baseline in most regimes.","core_discovery":"The central claim is that the hybrid Weighted Naive Bayes (NBw) model, which attaches a real-valued weight to each generative parameter, can be turned into a practical federated classifier without ever transmitting counts or conditional probabilities. In the FedNBw algorithm each client first computes local probability tables, then repeatedly copies the current global weight vector, runs a limited number of L-BFGS-M optimization steps on its own data, sends the local weight updates to the server, and receives the averaged global weights back. The paper reports that limiting local optimization to five iterations and using the global weights directly, the NBw5fed configuration, gives the best accuracy among the variants tested and generally outperforms both the generative federated baseline and the non-federated discriminative model, with the exception of the 50- and 100-client regimes where the table-sharing baseline wins because it is unaffected by data partitioning. The paper claims this is more privacy-preserving because the shared weights are 'inherently meaningless in revealing sensitive information.'","pith_inferences":["The privacy claim is untested: because the paper gives no threat model or reconstruction attack, an attacker who observes the sequence of per-round weight updates might extract more information than a single averaged weight vector would suggest.","Even if final weights are opaque, intermediate per-client updates sent to the server each round may carry information about local class distributions; measuring this leakage empirically would settle whether the privacy advantage is real.","The same federated weight-sharing idea could extend naturally to semi-naive Bayes or tree-augmented classifiers, since the weights are generic per-parameter scalars rather than NB-specific quantities.","The large-client regime result hints at a scaling rule: when each client has very little data, discriminative fine-tuning of shared weights degrades, and directly federating the tables becomes the more robust choice."],"forward_implications":["Federated discriminative Naive Bayes can be deployed in settings where sharing conditional probability tables is considered too sensitive, since only the weight vector crosses the network.","The finding that a five-iteration local optimization limit improves federated accuracy provides a practical default for similar federated discriminative algorithms.","In regimes with many clients, the generative table-sharing baseline remains superior, so a deployer may need to choose between the two approaches based on the expected number of clients.","The authors identify differential privacy as a natural next step to give formal privacy guarantees on top of the weight-sharing scheme."],"supporting_citations":[{"why":"Establishes the federated averaging protocol that the proposed algorithm follows for distributed weight aggregation.","marker":"[6]"},{"why":"Introduces the weighted Naive Bayes model whose parameters the federated method optimizes.","marker":"[16]"},{"why":"Provides the NBw algorithm and the experimental evidence that discriminative weighting improves over generative NB; source of the baseline implementation.","marker":"[17]"},{"why":"Supplies the discriminative learning theory linking NB parameters to logistic-regression-style optimization, motivating the conditional log-likelihood objective.","marker":"[8]"},{"why":"Prior federated/differential-privacy Naive Bayes that shares probability tables; the privacy contrast the paper argues against.","marker":"[2]"},{"why":"The L-BFGS-B optimizer used for the local discriminative weight updates.","marker":"[19]"},{"why":"Source of the 12 public discrete datasets used in the evaluation.","marker":"[13]"},{"why":"Another federated Naive Bayes baseline that relies on generative information, used as a related comparison.","marker":"[7]"}],"fun_headline_variants":["Federated NB shares weights, not counts, for privacy","Weight-shared Naive Bayes outperforms table-sharing federated model","Federated Naive Bayes: weight-sharing wins, but not with 100 clients","Discriminative NB goes federated with weight updates only","FedNBw: sharing meaningless weights keeps data private and accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the shared weight vector is inherently 'meaningless' and therefore cannot reveal sensitive information about a client's data; the paper assumes this privacy property rather than deriving it from a threat model, an attack analysis, or a formal privacy bound.","fun_headline_variants_meta":{"raw":{"variants":["Federated NB shares weights, not counts, for privacy","Weight-shared Naive Bayes outperforms table-sharing federated model","Federated Naive Bayes: weight-sharing wins, but not with 100 clients","Discriminative NB goes federated with weight updates only","FedNBw: sharing meaningless weights keeps data private and accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1504,"prompt_tokens":854,"completion_tokens":650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":470,"tokens_out":650,"duration_ms":5653,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:01:34.774049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedNBw on synthetic datasets where the true class priors and feature-conditional counts are known, give an attacker the sequence of per-round weight vectors (or even the final global weights), and test whether the attacker can recover those counts or identify whether a specific record was in a client's data with better-than-chance accuracy; if any such reconstruction succeeds, the advertised privacy advantage is false. A simpler check is to compare the weight trajectories produced by two datasets that differ only in class priors: if the trajectories diverge systematically, the weights carry information about the very quantities the paper says stay private.","supporting_citations":[{"cited_title":"In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics","cited_arxiv_id":null,"evidence_quote":"Establishes the federated averaging protocol that the proposed algorithm follows for distributed weight aggregation."},{"cited_title":"Journal of Machine Learning Research 14, 1947–1988 (2013)","cited_arxiv_id":null,"evidence_quote":"Introduces the weighted Naive Bayes model whose parameters the federated method optimizes."},{"cited_title":"Machine Learning 106, 1289–1329 (2017) Federated Learning with Discriminative Naive Bayes Classifier 13","cited_arxiv_id":null,"evidence_quote":"Provides the NBw algorithm and the experimental evidence that discriminative weighting improves over generative NB; source of the baseline implementation."},{"cited_title":"Machine Learning59, 267– 296 (2005)","cited_arxiv_id":null,"evidence_quote":"Supplies the discriminative learning theory linking NB parameters to logistic-regression-style optimization, motivating the conditional log-likelihood objective."},{"cited_title":"In: Proceedings of the 20th International Conference on Smart Business Technologies","cited_arxiv_id":null,"evidence_quote":"Prior federated/differential-privacy Naive Bayes that shares probability tables; the privacy contrast the paper argues against."},{"cited_title":"ACM Transactions on Mathematical Software 23, 550–560 (1997)","cited_arxiv_id":null,"evidence_quote":"The L-BFGS-B optimizer used for the local discriminative weight updates."},{"cited_title":"ACM SIGKDD Explorations Newsletter15, 49–60 (2013)","cited_arxiv_id":null,"evidence_quote":"Source of the 12 public discrete datasets used in the evaluation."},{"cited_title":"In: Proceedings of the 14th International Conference on Computing Communication and Networking Technologies","cited_arxiv_id":null,"evidence_quote":"Another federated Naive Bayes baseline that relies on generative information, used as a related comparison."}],"review_version":1}