{"id":"82c6c6dd-94df-48ae-9c99-d5d2dbcd50d5","arxiv_id":"2504.21634","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Fairness metrics computed on differentially private synthetic data differ from real-data values by up to 0.32 for some measures, despite staying below 0.1 on average, across Adult, COMPAS, and Diabetes.","lead":"This paper tests whether AI fairness audits can safely use differentially private synthetic data instead of the real sensitive data. It compares fairness scores on real versus synthetic versions of three public datasets and finds that averages stay close but some important metrics drift far.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'average difference below 0.1' criterion hides per-metric errors that can flip audit conclusions; the synthesis model does not preserve P(S,Yhat), so even demographic parity is unreliable.","rationale":"The reader's weakest assumption—that the MST-selected marginals preserve the joint distribution needed for fairness metrics—is correct and is the core of the problem. However, the reader's emphasis on conditional-use-accuracy measures and the missing (S,Y) edge is too narrow. The paper's own tables show that even demographic parity, an independence metric with no ground-truth conditioning, has errors around 0.17–0.18 on Adult data, large enough to change an audit from 'disparity present' to 'no disparity detected.' This means the failure is not specific to sufficiency metrics or to the absent (S,Y) edge; it is a general failure to preserve P(S,Yhat), because Yhat is a deterministic function of features whose joint distribution a tree-structured marginal model cannot faithfully reproduce. The paper's central claim—that synthetic data supports meaningful fairness auditing after the real data are discarded—is therefore not supported by the reported experiments. The 'average difference below 0.1' statistic is misleading because averaging over six measures, several with near-zero baseline values, hides material per-metric errors. No error bars, no per-metric significance tests, and no code release with a commit hash are provided, so the empirical basis is also not independently checkable. The negative results in Section 7.2 acknowledge the limitation for conditional use accuracy but the conclusion and abstract still overclaim. Because the core empirical demonstration fails for at least one common fairness metric (demographic parity) on the largest dataset, and because no theoretical bound is offered, the paper cannot be accepted as establishing the framework's validity. A revision that narrows the claim, reports per-metric errors with intervals, and identifies precisely which metrics and data types are safe could be reconsidered, but as written the central claim should be rejected.","tokens_in":14796,"tokens_out":4852,"duration_ms":56088,"concrete_test":"Re-run the Adult random-forest experiment of Table 5 and compute the distribution of synthetic demographic parity over the 100 generated datasets, reporting the mean and a 95% confidence interval. If the original value 0.1933 falls outside the interval, or if the absolute error remains above 0.1, then the 'average below 0.1' claim does not justify an auditor discarding the real data. Ideally, also repeat with the (S,Y) edge added as described in Section 7.1 to separate the missing-edge explanation from the more general failure of the tree model to preserve P(S,Yhat).","verdict_should_be":"REJECT","load_bearing_attack":"The central claim rests on Section 7.1's statement that 'the average of the differences between the fairness measure values of the original and synthetic datasets is below 0.1 for all datasets across all models.' This average over six measures masks individual errors large enough to invalidate an audit. In Table 5 (Adult, random forest), demographic parity is 0.1933 on the original data but 0.0114 on synthetic data, an error of 0.1818; a disparity that would be flagged is almost completely erased. Table 6 (Adult, logistic regression) shows conditional use accuracy equality (true positive) error of 0.3154, and Table 7 (Adult, KNN) shows demographic parity error of 0.1675 and equalized odds (true positive) error of 0.1557. The paper's Section 7 analysis predicts conditional-use-accuracy measures should be hardest because they have the boxed variable Yhat in both numerator and denominator, but the data contradict this: demographic parity and equalized-odds also fail badly. The structural reason is that Yhat is not part of the original data at synthesis time, so preserving P(Yhat|S) requires preserving P(X|S) well enough that the audited model's predictions on synthetic rows have the correct conditional distribution. The default maximum-spanning-tree marginal selection (Section 5.1) keeps all 1-way marginals plus pairwise cliques on MST edges; a tree cannot encode the high-order feature interactions used by a random forest, and nothing in the method specifically targets the conditional distribution of features given S. Thus the load-bearing assumption is not just the missing (S,Y) edge: it is that the selected marginals induce an accurate joint distribution over (S, Yhat) for arbitrary downstream models. The paper provides only a handful of averaged empirical numbers, no confidence intervals, and no argument or bound for why this should hold.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for fairness auditing of AI systems without retaining real data: an auditor generates differentially private synthetic data from the real data using the Private-PGM tool (winner of the 2018 NIST challenge), discards the real data, and evaluates an AI model with a fairness checker using standard metrics (demographic parity, equalized odds, conditional use accuracy equality, overall accuracy equality). The empirical evaluation compares fairness metrics computed on original versus synthetic data for Adult, COMPAS, and Diabetes datasets across several classifiers. The paper reports that the average absolute difference across six metrics is below 0.1 for every dataset/model combination and concludes that synthetic data generally preserves fairness properties well enough for meaningful auditing, while acknowledging in Section 7.2 that Conditional Use Accuracy Equality is harder to approximate.","tokens_in":15094,"tokens_out":6472,"duration_ms":62556,"significance":"If the central claim were established, this would be a practically valuable contribution: it would let third-party auditors avoid custody of sensitive data while retaining quantitative fairness evaluation, with a formal DP guarantee and an open-source implementation. The paper has real strengths: the fairness measures are standard definitions, the comparison against real-data metrics is an external benchmark rather than a circular evaluation, the synthesis pipeline is the well-known NIST-winning Private-PGM, and the code is publicly available. However, the reported results do not support the headline claim as stated; the average-difference criterion hides large per-metric errors that can reverse an audit verdict, so the significance hinges on a substantial reframing or additional analysis.","major_comments":[{"comment":"The criterion 'the average of the differences ... is below 0.1' masks errors that are decisive for audit outcomes. In Table 5 (Adult, random forest), Demographic Parity is 0.1933 on the original data and 0.0114 on synthetic data, a difference of 0.1818; under a common epsilon=0.1 threshold the original data would be flagged as unfair while the synthetic data would not. Table 6 (Adult, logistic regression) shows a Conditional Use Accuracy Equality (TP) difference of 0.3154, and Table 7 (Adult, KNN) shows Demographic Parity and Equalized Odds (TP) differences of 0.1675 and 0.1557. These are not small perturbations; they are large enough to flip the conclusion of a fairness audit, so 'generally preserves fairness with a reasonable degree of accuracy' (Section 7.1) is not supported by the paper's own tables.","section":"Section 7.1, Tables 5-7"},{"comment":"The proposed explanation that Conditional Use Accuracy Equality should be hardest because the predicted outcome appears in both numerator and denominator is contradicted by the data: Demographic Parity and Equalized Odds also show large errors (e.g., Table 7 differences of 0.1675 and 0.1557; Table 5 difference of 0.1818 for Demographic Parity). The more fundamental issue is structural: the model predictions Yhat are not part of the data at synthesis time, so the synthetic data does not directly preserve P(Yhat | S). The default maximum-spanning-tree selection in Section 5.1 retains pairwise marginals only, and Section 7.1 admits that adding the (S,Y) edge would improve accuracy; the paper should analyze this dependence rather than attributing the failure to the fairness measure's algebraic form.","section":"Section 7.2, Tables 5-9"},{"comment":"The paper averages fairness scores over 100 synthetic datasets but reports no standard deviation, range, or confidence interval. For a quantitative audit, an average error below 0.1 is not sufficient: the auditor needs to know how much a single synthetic dataset can deviate from the real-data value, because the proposed workflow uses one generated dataset after discarding the real data. Without this information, the framework's reliability for a single audit is not established.","section":"Section 6, paragraph on 100 runs"},{"comment":"The statement that adding the (S,Y) edge 'would have been more accurate' is not quantified anywhere; no table, figure, or magnitude accompanies it. Since the framework discards the real data immediately after synthesis, the auditor cannot generally know which higher-order edges are needed, so the default configuration is what matters. The paper should either validate the default workflow with decision-focused error bounds or explicitly restrict the framework's applicability to fairness measures that are robust to the default marginals.","section":"Section 5.1 and Section 7.1, added (S,Y) edge"}],"minor_comments":[{"comment":"The phrase 'predicated result' should be 'predicted result'.","section":"Section 3.1"},{"comment":"The RDP guarantee statement 'this procedure satisfies (α, α/2σ^2)-RDP' appears to omit the sensitivity Δ_f that appears in Theorem 3.3; please reconcile the notation.","section":"Section 5.1"},{"comment":"The dataset descriptions are imprecise (e.g., Adult 'about 30000 individuals' and COMPAS 'about 7000'); specify exact versions, preprocessing, and train/test splits to ensure reproducibility.","section":"Section 6"},{"comment":"Consider adding standard deviations or ranges across the 100 synthetic datasets, and provide a threshold context (e.g., epsilon=0.1) when interpreting the size of the differences.","section":"Tables 5-9"},{"comment":"The phrase 'by common sense or in an alternative assumption' is informal for a research paper; specify the exact procedure for adding the (S,Y) edge.","section":"Section 7.1"},{"comment":"The statement that 'biases present in the original data are captured and reflected in the synthetic counterpart' is stronger than the experimental results support; qualify it with the exceptions shown in Tables 5-9.","section":"Section 8"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claim is undercut by its own experimental tables. The paper may be salvageable as a negative result or as a narrowly scoped positive result for specific metrics (e.g., Overall Accuracy Equality on Diabetes), but the current framing and the average-difference criterion need major revision. I would also ask the authors to report per-run variability and to make the Section 7.1 edge-addition experiment concrete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a concrete empirical study of an operational workflow—generate DP synthetic data, discard the raw data, and compute fairness metrics on the synthetic data—but the paper's central claim that synthetic data \"generally preserves\" fairness is overstated. The average-difference criterion hides per-metric errors that would flip an audit conclusion.\n\nWhat is actually new: the tripartite framework (data provider, model maker, auditor) is clearly described, and the idea of discarding the real data immediately after synthesis is a defensible operational proposal. The paper reuses Private-PGM (the NIST 2018 winner) and the authors' own fairness checker, both appropriate choices. Running 100 synthetic datasets per model and averaging is good practice. The intuition about Conditional Use Accuracy Equality being hardest to approximate—because it conditions on Y_hat in both numerator and denominator—is reasonable and partially supported.\n\nThe soft spots are real. In Table 5, Demographic Parity on Adult-RF differs by 0.18; in Table 6, Conditional Use Accuracy Equality (TP) differs by 0.32. Those are not small. The paper's own Section 7.1 admits that adding the (S,Y) edge improves accuracy, but the results of that experiment are not reported. There are no confidence intervals or standard deviations across the 100 runs, and no code link is given despite the claim that the notebooks are public. The stress-test point is correct: the maximum-spanning-tree marginal selection does not guarantee preservation of P(X|S), so even demographic parity can fail badly. The boxed-variable analysis in Section 7 is too weak to explain why demographic parity fails.\n\nThe citation pattern is fine. The authors cite the existing work that already raised the fairness-of-DP-synthetic-data question [13, 23, 61] and position their contribution as an evaluation rather than a new method. Self-citation of their own fairness checker is acceptable.\n\nThe paper deserves a serious referee. It is an honest empirical validation of a niche but growing practice, and the flaws are correctable: report per-metric errors with uncertainty, include the (S,Y) experiment, release code with a hash, and soften the conclusion to name the metrics and settings where synthetic auditing is reliable. As is, it would need heavy revision; as a case study with those fixes, it would be a solid contribution to the DP-synthetic-data and fairness-auditing community.\n\nRecommendation: send to peer review, but require the revisions above before acceptance.","headline":"A useful empirical validation of a DP-synthetic-data auditing workflow, but the headline claim that fairness is 'generally preserved' is not supported by the paper's own tables.","tokens_in":15689,"tokens_out":1942,"would_cite":false,"duration_ms":21592,"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":"Differentially private synthetic data can stand in for real data in AI fairness audits, keeping average metric error below 0.1.","keywords":["fairness auditing","differential privacy","synthetic data","Markov random field","Rényi differential privacy","fairness metrics","conditional use accuracy equality","equalized odds"],"falsifier":"Run the same pipeline on a dataset in which the protected attribute influences the label only through a three-way interaction that no single tree edge captures; if the average difference between synthetic and real fairness scores then exceeds 0.1 for demographic parity or equalized odds, the claimed approximation would be shown not to generalize. A cheaper check is to rerun the authors' notebooks on a dataset with more than a dozen attributes and compare the error distribution.","tokens_in":14586,"feed_emoji":"⚖️","tokens_out":8995,"duration_ms":84235,"temperature":0.7,"pith_summary":"The paper tries to establish that an auditor can replace a sensitive real dataset with a differentially private synthetic copy, discard the original, and still reach trustworthy conclusions about an AI system's fairness. It proposes a workflow in which the auditor generates the synthetic data immediately upon receiving the real data and keeps only the synthetic version for all future audits. Across the Adult, COMPAS, and Diabetes datasets and several models, the average absolute difference between fairness measures computed on real and synthetic data stays below 0.1. This would let third-party fairness audits shed their main privacy liability: auditors would no longer need to be custodians of raw sensitive records.","feed_headline":"Private synthetic data reproduces fairness metrics within 0.1","feed_subtitle":"Auditors can discard the originals and still catch bias in Adult, COMPAS, and Diabetes.","key_machinery":"The engine is the select-measure-generate method for differentially private synthetic data, built around a Markov random field—a graphical model whose nodes are attributes and whose edges are dependencies preserved in the synthetic output. The selection step computes pairwise mutual information between all attributes, takes a maximum spanning tree over those weights, and protects every 1-way marginal plus the 2-way marginals on the tree's edges; by default the edge between the protected attribute and the ground truth is not selected. The measure step adds Gaussian noise calibrated to Rényi differential privacy, splitting the privacy budget equally between the 1-way marginals and the selected cliques. The generate step samples new rows from the fitted Markov random field. This construction lets the synthetic data inherit the joint statistics that demographic parity, equalized odds, and accuracy equality depend on, while severing the record-level link to real individuals.","core_discovery":"The central finding is that fairness-relevant statistics of a dataset survive differentially private synthesis well enough for auditing. Using a select-measure-generate pipeline, the authors fit a Markov random field to noise-perturbed marginals selected by a maximum-spanning-tree heuristic, generate 100 synthetic datasets per (dataset, model) pair, and compute six fairness measures on each. Averaged over those runs, the absolute difference between synthetic and original fairness scores is below 0.1 for every dataset and model tested. The paper also identifies a structural exception: sufficiency measures such as conditional use accuracy equality condition on the model's prediction, which is not present in the original data and therefore is not captured by the marginals; those measures show larger errors, up to roughly 0.32 in one Adult experiment. The authors conclude that the synthetic-data workflow remains valid for auditing and that a known edge between the protected attribute and the ground truth can be added to reduce the sufficiency-measure error.","pith_inferences":["Inference: because the 0.1-bound result is empirical on three tabular datasets, a natural next test is high-dimensional or text-derived data, where the maximum-spanning-tree marginal set will omit more of the joint structure that fairness metrics may depend on.","Inference: the paper's diagnosis of sufficiency-metric errors implies an inexpensive remedy an auditor could adopt without changing the pipeline—include the protected-attribute/ground-truth clique explicitly, at the cost of a small slice of the privacy budget.","Inference: the privacy guarantee protects individuals in the training data, but the published fairness scores themselves are aggregate statistics; inference attacks on those statistics remain outside what this framework claims to prevent."],"forward_implications":["Auditors can adopt a generate-then-discard protocol: synthesize on receipt of the data, erase the originals, and keep the synthetic copy for repeated audits of future models.","Independence and separation metrics—demographic parity, equalized odds, and overall accuracy equality—are reproduced closely enough on the tested datasets that a fairness violation visible in real data should remain visible in synthetic data.","Sufficiency metrics such as conditional use accuracy equality should be interpreted cautiously in this workflow, since their errors can exceed the 0.1 average by a wide margin.","The framework requires no dataset-specific tuning beyond choosing the protected attribute, so it transfers to a new dataset and model without re-engineering.","With the Gaussian mechanism under Rényi differential privacy at a fixed budget, the auditor can state a formal bound on what the released synthetic data reveals about any individual."],"supporting_citations":[{"why":"Supplies the select-measure-generate synthetic-data algorithm: mutual-information maximum spanning tree, noisy marginals, and Markov-random-field sampling.","marker":"[39]"},{"why":"Gives the graphical-model estimation and inference machinery used to fit the Markov random field from the measured marginals.","marker":"[40]"},{"why":"Provides the fairness checker that computes the six fairness measures on both real and synthetic datasets.","marker":"[64]"},{"why":"Defines the fairness measure taxonomy (independence, separation, sufficiency) and identifies the datasets' known fairness concerns.","marker":"[47]"},{"why":"Establishes Rényi differential privacy, the accounting used to state the privacy guarantee of the Gaussian mechanism.","marker":"[41]"},{"why":"Provides the Gaussian mechanism and sensitivity framework underlying the noise added to selected marginals.","marker":"[19]"}],"fun_headline_variants":["Synthetic data audits fairness, error under 0.1","Private data replicates fairness metrics within 0.1","Audit AI bias with synthetic data, 0.1 accuracy","Fairness checks persist in private synthetic data","DP synthetic data preserves fairness scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the assumption that the one-way counts and pairwise links chosen by the maximum-spanning-tree heuristic preserve how the protected attribute, the true label, and the model's predictions depend on one another; the paper's own results show this assumption breaks for sufficiency metrics, whose errors reach about 0.32 when the protected-attribute/ground-truth link is not encoded.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic data audits fairness, error under 0.1","Private data replicates fairness metrics within 0.1","Audit AI bias with synthetic data, 0.1 accuracy","Fairness checks persist in private synthetic data","DP synthetic data preserves fairness scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2543,"prompt_tokens":912,"completion_tokens":1631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1556}},"tokens_in":528,"tokens_out":1631,"duration_ms":13242,"temperature":1.0,"reasoning_tokens":1556,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:58:01.282142+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline on a dataset in which the protected attribute influences the label only through a three-way interaction that no single tree edge captures; if the average difference between synthetic and real fairness scores then exceeds 0.1 for demographic parity or equalized odds, the claimed approximation would be shown not to generalize. A cheaper check is to rerun the authors' notebooks on a dataset with more than a dozen attributes and compare the error distribution.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the graphical-model estimation and inference machinery used to fit the Markov random field from the measured marginals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fairness checker that computes the six fairness measures on both real and synthetic datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes Rényi differential privacy, the accounting used to state the privacy guarantee of the Gaussian mechanism."}],"review_version":1}