{"id":"79f50468-6e8b-429b-94a7-bf373f0e9918","arxiv_id":"2504.15924","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"UDJ-FL is a single federated learning objective whose hyperparameters select among four distributive-justice fairness notions, with client weights set by aleatoric uncertainty.","lead":"The paper proposes UDJ-FL, a federated learning method that switches between four fairness goals (egalitarian, utilitarian, Rawls' difference principle, desert) by changing a few hyperparameters, using each client's data uncertainty as the measure of who needs help or who contributes most. It reports experiments on two image datasets showing the method matches or beats existing fair federated learning baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Desert and Rawls entries in Table 2 do not follow from Eq. 8, so the central 'one objective, four fairness notions' claim is not yet established.","rationale":"The reader's stated weakest assumption is the validity of softmax entropy as a measure of aleatoric uncertainty and hence of client advantage/contribution. That is a real concern, and the authors themselves flag it in Section 5.3. However, the more load-bearing issue is internal: even granting the entropy proxy, the mapping from Eq. 8 to the four objectives in Table 2 is not algebraically correct as printed. The desert sign error is particularly serious because it reverses the intended relationship between data quality and optimized loss. The reader's rationale does mention the unproved transformation from Eq. 5 to Eq. 8 and the desert/rawls equation problems, so there is partial agreement, but I would elevate the Eq. 8-to-Table 2 inconsistency above the uncertainty-proxy question as the primary barrier to accepting the central claim. Since the flaw is specific and potentially fixable by correcting the equations or clarifying the implemented objective, and since the empirical results and code may still support the method, the conditional verdict remains appropriate rather than a rejection.","tokens_in":23502,"tokens_out":5964,"duration_ms":60073,"concrete_test":"On a small two-client toy problem with known losses and uncertainty scores, compute the per-client gradients of (a) Eq. 8 with the desert hyperparameters r=1, beta=beta_i>0, gamma=0, and (b) Eq. 14. If the gradients differ in sign or in relative weighting, Eq. 14 is not a special case of Eq. 8, and the desert claim must be re-derived or the code inspected to identify the objective actually optimized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires each entry in Table 2 to be a genuine special case of Eq. 8 obtained by choosing r, beta, and gamma. For desert this is internally inconsistent. Eq. 8 with r=1, beta=beta_i (Eq. 13), gamma=0 gives sum_i H_i(theta)^{beta_i}, not the printed (1/N) sum_i H_i(theta)^{-beta_i} in Eq. 14. Since Eq. 13 defines beta_i > 0, the two objectives have opposite monotonicity in H_i, so the printed desert objective cannot be recovered from Eq. 8 and does not obviously implement the intended 'high-quality clients get higher accuracy' behavior. The Rawls entry has a second mismatch: Eq. 10 and Table 2 raise upsilon_i to the power 1+beta, while substituting gamma=1 into Eq. 8 yields upsilon_i^1. Thus the paper's derivation does not establish that one parametric objective realizes all four fairness notions; the empirical table may still be correct, but it is not supported by the stated algebra. This is a checkable internal inconsistency, not a disagreement with a fairness consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces UDJ-FL, a federated learning objective intended to realize four client-level fairness notions drawn from distributive justice—egalitarian, utilitarian, Rawls' difference principle, and desert—through different settings of three hyperparameters r, β, and γ. The client weights are aleatoric uncertainties estimated as the average softmax entropy of each client's training data (Eq. 9). The objective is written in Eq. 8 as sum_i υ_i^γ H_i(θ)^{rβ}, with derivations presented for each fairness notion in Section 4 and summarized in Table 2. The paper contributes Algorithm 1, a q-FedAvg-style solver, a Lipschitz-constant lemma, generalization bounds in Appendix D, and experiments on Dirty-MNIST and CURE-TSR with five clients whose data quality is controlled via clean/ambiguous shards.","tokens_in":23657,"tokens_out":12428,"duration_ms":111340,"significance":"If the algebraic route from Eq. 8 to the four objectives were valid, UDJ-FL would be a practically attractive contribution: one training routine, no architectural changes, hyperparameter selection of the fairness ethics, public code, and experiments against a broad set of baselines. The taxonomy of prior work along the four justice dimensions is also useful. However, the central derivation currently contains internal inconsistencies in the Rawls and desert cases, and the uncertainty weighting in Eq. 9 has a sign issue; these undermine the advertised 'one objective, four fairness notions' claim as written. The empirical results may still be meaningful, but they cannot be fully interpreted until the objectives being executed are identified unambiguously. The public code and the controlled uncertainty-partition experimental setup are concrete strengths.","major_comments":[{"comment":"The step from Lan et al.'s Eq. (5) to Eq. (8) is not derived. Eq. (5) has the exponent −rβ, while Eq. (8) uses +rβ, and the per-client weights υ_i^γ appear without any formal rule connecting q_i in Eq. (5) to υ_i^γ. The sentence that the sign change is due to federated learning minimizing costs rather than distributing utilities explains the sign convention but does not show that the resulting expression belongs to the same fairness family or is an equivalent transformation. Since every row of Table 2 is claimed to be a special case of Eq. (8), this missing equivalence is load-bearing.","section":"§4.1, Eq. (8)"},{"comment":"Setting r=1+1/β and γ=1 in Eq. (8) yields sum_i υ_i H_i(θ)^{1+β}. Eq. (10), as typeset, contains a prefactor (1+β) and a weight on υ_i that appears to be υ_i^{1+β}; if the intended weight is instead υ_i, the equation still is not a direct special case of Eq. (8). Thus the Rawls row of Table 2 is not obtained from Eq. (8) by the stated substitutions, and the claim that β→∞ recovers min-max fairness needs a corrected intermediate objective.","section":"§4.2, Eq. (10) and Table 2"},{"comment":"The desert objective is not a special case of Eq. (8). Eq. (13) defines β_i>0, yet Eq. (14) minimizes (1/N) sum_i H_i^{−β_i}, whose monotonicity in H_i is opposite to that of sum_i H_i^{β_i}, which is what Eq. (8) gives for r=1, scalar β=β_i, and γ=0. The text says 'we let β≤0', but Eq. (13) assigns positive values. Moreover, Eq. (8) has a single scalar β, so the per-client vector β cannot be inserted without changing the objective family. This invalidates the desert entry of Table 2 as written.","section":"§4.4, Eqs. (13)–(14)"},{"comment":"Eq. (9) defines υ as the average of p log p over the softmax distribution, which is non-positive, while the text and Table 3 treat υ_i as a positive uncertainty score. There is a missing minus sign, or equivalently an implicit log(1/p) convention, that must be stated. Because υ_i enters every objective through υ_i^γ and is used to identify the least advantaged client, the sign convention is load-bearing.","section":"§3.3/§4.1, Eq. (9)"},{"comment":"The generalization bound appears to be proved for an objective different from Eq. (8). Eq. (26) defines ZHat{L}^{rβ} as a sum of powered empirical risks, while the dual-norm representation in Eq. (27) yields the rβ-norm (sum_i F_i^{rβ})^{1/(rβ)}. Lemma 2 then bounds L_υ(h) in terms of ||υ||_p ]tilde{L}_{rβ}(h), i.e., the norm, not the sum used in Eq. (26) or in Eq. (8). The authors should state precisely which objective the bound covers and prove the bound for the actual UDJ-FL loss.","section":"Appendix D, Eqs. (26)–(28)"}],"minor_comments":[{"comment":"The text says the Rawls experiments use β→∞, while Table 4 reports β=5; please state explicitly that finite β=5 is an approximation and discuss sensitivity to β.","section":"§5.2, Rawls' Difference Principle"},{"comment":"Algorithm 1 treats β as a scalar input, but the desert objective uses a per-client vector β; clarify how the pseudocode and the released implementation handle the vector case.","section":"§4.4/Algorithm 1"},{"comment":"The uncertainty estimate in Eq. (9) depends on model parameters θ, but Algorithm 1 computes υ_i once after S rounds of solo training. The notation θ is reused for the global model; specify at which θ the entropy is evaluated and whether the values are held fixed during federated training.","section":"§3.3, Eq. (9)"},{"comment":"The notation υ_i is introduced as normalized weights υ_i/Σ_j υ_j in Section 4.1, while Eq. (9) defines υ as an average entropy; clarify whether the weights used in Eq. (8) are the normalized values or the raw uncertainties.","section":"§4.1, Eq. (8)"},{"comment":"The variance term in Eq. (28) uses v_i while the weights elsewhere are υ_i; align the notation.","section":"Appendix D, Eq. (28)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of FAccT and the public code is a concrete asset. My principal concern is the formal derivation of the unified objective: the Rawls and desert entries of Table 2 do not follow from Eq. (8) as written, and the uncertainty sign in Eq. (9) is ambiguous. These issues are load-bearing but appear fixable by rewriting the parameter mapping and the uncertainty definition, so I recommend major revision rather than rejection. I also suggest that the revision state explicitly that the evaluation metrics in Eqs. (11) and (15) are direct targets of the constructed objectives."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a paper with a genuinely useful idea—using aleatoric uncertainty as client weights to interpolate among four distributive-justice fairness notions—but the central derivation does not hold up. The specific claim that Table 2's objectives are all special cases of Eq. 8 is wrong as printed, and until that's fixed, the 'one objective, four fairness definitions' story is unsupported.\n\nWhat's genuinely good: the uncertainty proxy is low-cost and plausible; the philosophical mapping is thoughtful; Table 1's taxonomy is useful; the empirical study on Dirty-MNIST and CURE-TSR is reasonably thorough, and code is public. If the math were solid, this would be a nice contribution to fair FL.\n\nSoft spots, in order of severity. First, the algebra. Eq. 8 is written as sum_i upsilon_i^gamma H_i(theta)^{r beta}. For Rawls (Section 4.2), they set gamma=1, r=1+1/beta, and claim the objective becomes (1+beta) sum_i upsilon_i^{1+beta} H_i^{1+beta}. But substituting into Eq. 8 gives no (1+beta) factor and upsilon_i^1, not upsilon_i^{1+beta}. For desert (Eq. 14), they claim (1/N) sum H_i^{-beta_i} with beta_i defined by Eq. 13, which is positive. But Eq. 8 with r=1, gamma=0 gives sum H_i^{beta}; with beta=beta_i that has the opposite monotonicity. So two of the four entries do not follow. This is exactly the kind of thing a referee needs to see fixed before the headline claim is credible.\n\nSecond, the empirical validation is partially a fit. Appendix E says they chose hyperparameters that obtained the best fairness results, not a priori. That makes 'we can achieve all four' look like tuning to the benchmark rather than a verified prediction. The authors should report the full sweep and a proper validation split.\n\nThird, the generalization bound (Lemma 2) is standard AFL machinery, and the paper should be clear it doesn't certify any fairness guarantee.\n\nThe limitation section is honest about the aleatoric-uncertainty assumption, which is good. But the internal inconsistency in the main derivation is not just a presentation issue—it breaks the core contribution.\n\nWho should read this: practitioners and researchers in fair FL who want a menu of client-level fairness definitions and are willing to reconstruct the math. I'd send to peer review, but with the strong expectation of major revision: correct the derivations, report the hyperparameter selection honestly, and re-run or re-justify the experiments. If the authors can show that a single objective really does recover all four notions—even with a corrected equation—that would be a solid result.","headline":"Good idea, broken derivation: the claim that one objective realizes all four justice notions doesn't survive arithmetic.","tokens_in":24253,"tokens_out":3554,"would_cite":false,"duration_ms":31564,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that one federated learning objective, UDJ-FL, can deliver four distributive-justice fairness ethics—egalitarian, utilitarian, Rawls' difference principle, and desert—by choosing three hyperparameters.","keywords":["federated learning","client-level fairness","distributive justice","aleatoric uncertainty","uncertainty quantification","fair resource allocation","hyperparameter selection","softmax entropy"],"falsifier":"On a synthetic federation, flip labels for the client with the highest true aleatoric uncertainty so that the trained model becomes confidently wrong on that client's data; if UDJ-FL's computed $\\upsilon_i$ stays low and the desert correlation $r_{\\boldsymbol{\\upsilon},A}$ still approaches $-1$, the entropy measure misidentifies contribution and the claimed fairness behavior is an artefact of the metric rather than the method.","tokens_in":23182,"feed_emoji":"⚖️","tokens_out":10736,"duration_ms":89322,"temperature":0.7,"pith_summary":"This paper tries to establish that a single federated learning objective, UDJ-FL, can produce four different distributive-justice fairness behaviors—egalitarian, utilitarian, Rawls' difference principle, and desert—merely by selecting hyperparameters. The key move is to define each client's advantage and contribution by its average softmax entropy over its own local data, an aleatoric-uncertainty estimate, and then reweight or power-transform client losses with those entropy scores. On two image datasets with five clients of controlled data ambiguity, the paper reports that UDJ-FL attains the intended fairness behavior for all four settings and matches or beats existing fair federated learning methods on the corresponding fairness metrics. If this holds, a practitioner could change the ethical principle governing a federation without changing the training architecture.","feed_headline":"Four fairness ideals from one federated learning objective","feed_subtitle":"Softmax-entropy scores define who is least advantaged; tests match or beat prior fair-FL methods on two image datasets.","key_machinery":"The load-bearing object is the UDJ-FL objective of Eq. 8, $h(\\theta)=\\sum_i \\upsilon_i^{\\gamma}H_i(\\theta)^{r\\beta}$, coupled to the client weight of Eq. 9, $\\upsilon_i = \\frac{1}{|D_i^{tr}|}\\sum_{x\\in D_i^{tr}} H(Y|x,\\theta)$, the average softmax entropy over the client's local training data. The hyperparameter settings in Table 2 select the fairness theory: $\\gamma$ controls whether high-entropy or low-entropy clients get emphasized, $r$ sets the growth rate, and $\\beta$ controls the concave or power shape of the loss transformation, with $\\beta\\to\\infty$ for Rawls' principle, $\\beta\\to 0$ for utilitarian, $\\beta=1$ for egalitarian, and client-specific negative exponents for desert. The same machinery also supplies the optimization rule: following $q$-FedAvg, the server aggregates client updates using a local Lipschitz upper bound (Lemma 1) to handle the exponentiated losses, and Lemma 2 uses Hölder's inequality to bound generalization for the uncertainty-weighted objective.","core_discovery":"The paper's central claim is that one federated objective, $$h(\\$\\theta$) = \\sum_{i=1}^{N} \\$upsilon_i^{{\\gamma}}$ H_i(\\$\\theta$)^{r\\$\\beta$},$$ where $H_i(\\theta)$ is client $i$'s empirical loss and $\\upsilon_i$ is the average softmax entropy of client $i$'s local training data, recovers four distributive-justice fairness ethics purely through the hyperparameters $r$, $\\beta$, and $\\gamma$. Setting $r=1+1/\\beta$, $\\gamma=1$, and letting $\\beta\\to\\infty$ yields Rawls' difference principle; $r=1+1/\\beta$, $\\beta\\to 0$, $\\gamma=-1$ yields utilitarianism; $r=1$, $\\beta=1$, $\\gamma=1$ yields egalitarianism; and $r=1$, $\\gamma=0$ with client-specific exponents $\\beta_i = (1/\\upsilon_i)/\\sum_j (1/\\upsilon_j)$ yields desert-based fairness. The authors argue that a client's aleatoric uncertainty is irreducible and stable, so it is the right quantity for deciding who is least advantaged and who contributes most, unlike dataset size or participation counts in cross-silo settings. On Dirty-MNIST and CURE-TSR with five clients of controlled data ambiguity, UDJ-FL reports the desired fairness behavior for each setting and matches or improves on $q$-FFL, AFL, PropFair, TERM, FedMGDA+, and CFFL on the corresponding fairness measures. The paper also adapts the $q$-FedAvg Lipschitz step-size argument to handle the exponentiated losses and proves a generalization bound for the uncertainty-weighted objective by extending the AFL bound.","pith_inferences":["The same entropy-weighting recipe could be applied to other aggregation rules such as FedProx or SCAFFOLD, since nothing in the argument binds it to averaging-based updates; the paper does not test this.","If entropy scores are cheap to compute in one warm-up round, UDJ-FL could double as a diagnostic: comparing outcomes under the four ethics would reveal which clients are data-poor and which are data-rich, a use the authors mention only implicitly.","The empirical finding that $\\beta=0.1$ beats $\\beta=0$ for utilitarian fairness on Dirty-MNIST suggests the pure utilitarian objective of Eq. 12 may not be the practical optimum; a systematic sweep of $\\gamma$ and small $\\beta$ values is the implied but untested extension.","A direct comparison of aleatoric versus epistemic or predictive uncertainty weights would test the paper's irreducibility argument, since the authors do not run that experiment."],"forward_implications":["A practitioner can select a fairness ethics for a cross-silo federation by setting $r$, $\\beta$, and $\\gamma$; no change to the learning architecture is needed.","UDJ-FL contains $q$-FFL as a special case when $r=1+1/\\beta$ and $\\beta>0$, so results transfer to that popular baseline's settings.","Client contribution can be measured by data quality instead of dataset size or participation count, which matters when the largest client has the noisiest data.","The fairness metrics used for evaluation become decoupled from the training objective, so the same trained model can be audited under egalitarian, utilitarian, Rawlsian, or desert criteria.","The generalization bound extends the AFL-style analysis to uncertainty-weighted federated objectives, giving a theoretical guarantee for the reweighted loss family."],"supporting_citations":[{"why":"Supplies the axiomatic resource-allocation fairness family $f_\\beta$ that UDJ-FL adapts into its objective.","marker":"[19]"},{"why":"Provides $q$-FFL, the $q$-FedAvg optimization rule with Lipschitz step-size estimates, and a baseline for comparison.","marker":"[22]"},{"why":"Defines FedAvg, the baseline objective and data-partitioning strategy UDJ-FL is measured against.","marker":"[24]"},{"why":"Provides Dirty-MNIST, the ambiguous-digit construction, and the softmax-entropy uncertainty estimate used in Eq. 9.","marker":"[28]"},{"why":"Justifies treating softmax entropy as an aleatoric uncertainty estimate via the conditional-entropy and mutual-information decomposition.","marker":"[38]"},{"why":"Supplies the definitions of egalitarian, utilitarian, Rawlsian, and desert distributive justice that structure the four fairness metrics.","marker":"[18]"},{"why":"Is the source of the difference principle that the Rawls' DP fairness definition and its min-max reading rely on.","marker":"[30]"},{"why":"Contributes the AFL objective and the generalization-bound technique that Lemma 2 extends to uncertainty-weighted losses.","marker":"[26]"},{"why":"PropFair is a comparison baseline for utilitarian and Rawlsian fairness experiments.","marker":"[40]"},{"why":"CFFL is the comparison baseline for desert-based fairness experiments.","marker":"[23]"}],"fun_headline_variants":["One FL loss, four distributive-justice ethics","Four fairness metrics from one uncertainty-weighted loss","Egalitarian to Rawls: one federated objective","Rawls, utilitarian, egalitarian, desert: one loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a client's average softmax entropy over its own local training data is a valid, stable measure of that client's data quality, and therefore of its advantage and contribution; if noisy or adversarial local data produces misleadingly confident predictions, the fairness guarantees built on those entropy weights collapse.","fun_headline_variants_meta":{"raw":{"variants":["One FL loss, four distributive-justice ethics","Four fairness metrics from one uncertainty-weighted loss","Egalitarian to Rawls: one federated objective","Rawls, utilitarian, egalitarian, desert: one loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000457,"raw_usage":{"total_tokens":2422,"prompt_tokens":1200,"completion_tokens":1222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":816,"completion_tokens_details":{"reasoning_tokens":1159}},"tokens_in":816,"tokens_out":1222,"duration_ms":11551,"temperature":1.0,"reasoning_tokens":1159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:15:15.829544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic federation, flip labels for the client with the highest true aleatoric uncertainty so that the trained model becomes confidently wrong on that client's data; if UDJ-FL's computed $\\upsilon_i$ stays low and the desert correlation $r_{\\boldsymbol{\\upsilon},A}$ still approaches $-1$, the entropy measure misidentifies contribution and the claimed fairness behavior is an artefact of the metric rather than the method.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the axiomatic resource-allocation fairness family $f_\\beta$ that UDJ-FL adapts into its objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Dirty-MNIST, the ambiguous-digit construction, and the softmax-entropy uncertainty estimate used in Eq. 9."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies treating softmax entropy as an aleatoric uncertainty estimate via the conditional-entropy and mutual-information decomposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the definitions of egalitarian, utilitarian, Rawlsian, and desert distributive justice that structure the four fairness metrics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the source of the difference principle that the Rawls' DP fairness definition and its min-max reading rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the AFL objective and the generalization-bound technique that Lemma 2 extends to uncertainty-weighted losses."},{"cited_title":"Proportional Fairness in Federated Learning","cited_arxiv_id":"2202.01666","evidence_quote":"PropFair is a comparison baseline for utilitarian and Rawlsian fairness experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CFFL is the comparison baseline for desert-based fairness experiments."}],"review_version":1}