{"id":"256d34a8-c498-43bc-bc6c-0a753bd99027","arxiv_id":"2505.11132","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Matching each group's normal data to a common target distribution via Sinkhorn distance yields an anomaly detector whose score distributions are approximately equal across groups, with better AUC-ADPD trade-offs than existing fair detectors.","lead":"This paper introduces FairAD, an unsupervised anomaly detection method that maps each demographic group's normal data to a shared, simple target distribution, making density-based anomaly scores fair across groups. It also proposes a threshold-free fairness metric and reports improved accuracy-fairness trade-offs on six benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 2 (Transferable fairness) is unformalized and vacuous as stated; without a concrete bound on μ, the fairness guarantee on unseen anomalies does not follow from training-time predictive equality.","rationale":"The reader's weakest-assumption analysis lands on the same point: Assumption 2 is the hinge from training-set predictive equality to fairness on anomalies. The paper honestly proves in Claim 1 that demographic parity and equal opportunity cannot be guaranteed without extra assumptions; Assumptions 2/3 are that extra. But because μ and τ are unbounded, the assumptions are not falsifiable and cannot support a formal guarantee. I considered the Eq. (11) mis-specification and the nonnegativity error in Proposition 2's proof; both are real but presentation-level and fixable by replacing (11) with the per-group sum and by using a Sinkhorn divergence instead of the raw entropy-regularized OT objective. The transfer assumption is the deeper issue: it is an intrinsic limitation of unsupervised fairness, not a typo. The paper does provide empirical evidence that ADPD on anomalies is low, which is valuable; but the formal claim that Assumptions 2/3 yield equal opportunity or demographic parity is unsupported as written. Verdict remains CONDITIONAL: the paper should be accepted only after formalizing the assumption (e.g., a Lipschitz continuity condition on the score under a perturbation model) and reporting μ/τ on held-out anomalies, and after fixing the objective notation and the Sinkhorn definition.","tokens_in":34358,"tokens_out":7796,"duration_ms":84004,"concrete_test":"Construct a controlled anomaly set with group-dependent perturbation magnitude: train FairAD on two groups with identical normal distributions, so E is near zero, then test on anomalies where group A's anomalies are pushed far from the target (large norm) and group B's are near it (small norm), both equally anomalous by construction. If Ẽ is large, Assumption 2 fails. Repeat across real datasets, reporting the empirical ratio μ = Ẽ/E with confidence intervals; if no small uniform μ emerges, the fairness-on-anomalies claim must be withdrawn or replaced with a Lipschitz-style bound on score behavior under the perturbation model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 2 (Sec. 3.2) asserts a small μ ≥ 1 with Ẽ ≤ μE, where E is the predictive-equality violation on training normal data and Ẽ is the violation on unseen anomalies. The paper never bounds μ; for any model with E > 0, some finite μ exists by taking μ = Ẽ/E, and the 'smallness' is undefined. If E = 0, the assumption either holds trivially or fails without warning. The informal support—anomalies as perturbed normals and empirical agreement on test anomalies—does not establish a uniform bound: a detector can be perfectly fair on normal data and arbitrarily unfair on a group-specific anomaly distribution while still satisfying the assumption for a suitably large μ. Since the training objective only controls normal-data projection, the advertised equal-opportunity/demographic-parity guarantee (via Assumptions 2/3) is not derived; the method's rigorous contribution reduces to predictive equality on the training normal set. This is the load-bearing gap between the theoretical framework and the claim that FairAD 'achieves group fairness' on anomalies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies group fairness in unsupervised anomaly detection (UAD). It argues that without additional assumptions, demographic parity and equal opportunity cannot be guaranteed in the unsupervised setting, and it introduces two assumptions under which fairness on training normal data transfers to unseen anomalies. The proposed method, FairAD, learns a projection that maps normal data of each demographic group to a common compact target distribution, so that anomaly scores (distances to the origin in the projected space) are comparable across groups; two variants are given: Im-FairAD, which uses a shared target distribution without an explicit fairness term, and Ex-FairAD, which adds a Sinkhorn-based fairness regularizer on the score distributions. The paper also proposes ADPD, a threshold-free fairness metric, and evaluates the methods on six datasets under balanced and skewed group splits, reporting improved accuracy-fairness trade-offs relative to several baselines.","tokens_in":34604,"tokens_out":6299,"duration_ms":65349,"significance":"If the central mechanism is correct, this is a valuable contribution: fairness in unsupervised anomaly detection is understudied, and the idea of achieving group fairness by mapping all groups to a common target distribution is simple, avoids adversarial training, and is potentially task-compatible. The empirical evaluation is reasonably extensive (multiple tabular and image datasets, balanced/skewed splits, threshold-free and threshold-dependent metrics, ablations, contamination experiments), and the proposed ADPD metric is a useful addition to the evaluation toolbox. However, as detailed below, two load-bearing theoretical pieces—the printed Im-FairAD objective and the proof of Proposition 2—are not currently sound as written, and the transferability assumptions are vacuous as formalized. The empirical results may still support the method's practical utility, but the theoretical claims need substantial revision.","major_comments":[{"comment":"The stated objective for Im-FairAD is min_{φ,ψ} Σ_{s∈S} Sinkhorn(h_φ(X), Z) + (β/n) Σ_i ∥x_i − g_ψ(h_φ(x_i))∥². The first term does not depend on s, so the sum over groups equals |S|·Sinkhorn(h_φ(X), Z), which is exactly the fairness-unaware objective in Eq. (7). As printed, Eq. (11) therefore does not implement the group-wise distribution matching described in Eq. (10) and in the surrounding text, and the claim that Im-FairAD achieves fairness through a shared target distribution is not supported by the printed objective. The authors should correct Eq. (11) to use Sinkhorn(h_φ(X_{S=s}), Z) inside the sum and verify that the reported Im-FairAD results were obtained with that corrected objective.","section":"Section 3.3.2, Eq. (11)"},{"comment":"The proof of Proposition 2 relies on the assertion that M(·,·) is a distance metric with M(·,·)≥0 and that M(P,Q)=0 implies P=Q. However, the Sinkhorn objective defined in Eq. (6) includes the term α Σ_{i,j} P_{ij} log P_{ij}, which is non-positive and can make the objective negative; moreover, with entropy regularization the optimal value is not a metric and zero value does not imply equality of distributions. Therefore the chain of implications in Eq. (17) is not justified for the Sinkhorn distance as used in Eq. (11). The proposition should either be stated for a genuine distributional distance (e.g., Sinkhorn divergence, Wasserstein distance, or MMD with a characteristic kernel) with a proof that zero distance implies distributional equality, or the entropy regularization should be defined so that the resulting quantity is nonnegative and has the identity-of-indiscernibles property.","section":"Appendix A.3, proof of Proposition 2"},{"comment":"Assumptions 2 and 3 are vacuous as formalized. For any detector with E>0, the inequality Ẽ ≤ μE holds for μ = Ẽ/E, and the condition that μ is 'small' is never defined; thus the assumption does not constrain the model. If E=0, the assumption requires Ẽ=0, which is an unverified strong condition that the paper does not establish. Consequently, the statement that solving (10) together with Assumption 2 or 3 yields equal opportunity or demographic parity on unseen anomalies is not a formal derivation; the rigorous content reduces to predictive equality on the training normal set. The authors should either (i) provide a concrete quantitative bound on μ and τ derived from the generative model or from measurable quantities, or (ii) explicitly reframe the claim as an empirical regularity rather than a theoretical guarantee. The current supporting arguments—that existing methods 'implicitly verified' the assumptions and that test-set ADPD values are small—do not establish a uniform bound on μ.","section":"Section 3.2, Assumptions 2 and 3"},{"comment":"The ADPD metric is presented as a threshold-free measure of demographic parity. With finite samples, ADPD=0 holds if and only if the empirical score distributions are equal across groups, which is stronger than demographic parity at any single threshold; the paper should state this explicitly so that the metric is not overclaimed as an exact proxy for demographic parity. This is a conceptual clarification rather than a technical error, but it affects how the experimental comparisons are interpreted.","section":"Section 3.4 and Appendix G"}],"minor_comments":[{"comment":"There are corrupted placeholder strings in the manuscript text after Figure 2 and Figure 3 (e.g., '/uni00000013/...'), which appear to be artifacts of the source file and should be removed before publication.","section":"Section 4.1"},{"comment":"The quantity in Eq. (6) is called a 'Sinkhorn distance,' but with the added entropy term it is not a distance in the mathematical sense; consider renaming it to 'entropic optimal transport cost' or 'Sinkhorn loss' throughout the paper.","section":"Eq. (6)"},{"comment":"In the Ex-FairAD objective, the notation Sinkhorn(ζ(h_φ(X_{S=si})), ζ(h_φ(X_{S=sj}))) should be clarified: ζ(h_φ(X)) denotes the set of anomaly scores for samples in that group, and the Sinkhorn cost is computed on these one-dimensional score sets. Using 1-D scores is a special case that could be replaced by a simpler closed-form distance; a remark would help reproducibility.","section":"Section 3.3.2, Eq. (12)"},{"comment":"The column heading 'APDP' in Table 14 should be 'ADPD' for consistency with the rest of the paper.","section":"Table 14"},{"comment":"The ablation study for Ex-FairAD reports results without the fairness term, but the hyperparameter λ is varied only on three datasets; a sensitivity analysis for λ on the remaining datasets (Titanic, SP, Credit) would strengthen the claim that the fairness term is necessary and that performance is not overly sensitive to λ.","section":"Appendix D.2"},{"comment":"The Sinkhorn entropy coefficient α is fixed to 0.1 in all experiments, but no sensitivity analysis is reported; given that Proposition 2's proof depends on the distance-like properties of M, a brief study of α's effect on both fairness and detection would be useful.","section":"Appendix B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a preprint (arXiv:2505.11132v1) and the text contains some corrupted placeholder strings, which suggests the submitted version is not final. The empirical contribution is promising, but the theoretical framework has two fixable yet load-bearing issues (the mis-specified Eq. (11) and the invalid proof of Proposition 2) and one more fundamental issue (the vacuity of Assumptions 2 and 3). The authors should be encouraged to revise carefully, because with the objective corrected and the assumptions either formalized or explicitly downgraded to empirical claims, the paper could make a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the central idea is good: learn a projection that pushes each demographic group's normal data onto the same compact target distribution, then use the norm of the projection as an anomaly score. That is a clean, end-to-end way to enforce group-level score parity without adversarial training, and it is not in the cited prior work. Second, as printed, the main objective in Eq. (11) does not implement that idea: it sums Sinkhorn(h_phi(X), Z) over groups, with the same full sample X each time, so the sum has no group-specific effect. That is almost certainly a typo (Eq. (10) has the correct group-wise term), but it has to be fixed.\n\nWhat the paper does well: the assumption analysis in Section 3.2 (Claim 1) is a useful statement of why unlabeled anomalies make demographic parity and equal opportunity hard in unsupervised AD. The experiments are unusually broad: balanced and skewed splits, tabular and image data, and a reasonable set of fair-AD baselines. On COMPAS, Adult, and Credit, the AUC/ADPD trade-off genuinely favors Im-FairAD and Ex-FairAD. The ADPD metric is a sensible threshold-free summary of flagging-rate disparity across thresholds, and it improves on the threshold-dependent fairness ratio for comparison purposes.\n\nThe soft spots, in order of seriousness. (1) Proposition 2's proof relies on M >= 0; the Sinkhorn objective as defined in Eq. (6) includes an unnormalized entropy term + alpha sum P log P, which can be negative, so the nonnegativity step is false as written. Using a proper Sinkhorn divergence or KL-regularized OT fixes it, but the proof needs to be redone. (2) Assumptions 2 and 3 are vacuous in the form stated: any nonzero violation on training normals gives some finite mu, and \"small\" is never defined. The rigorous claim should be limited to predictive equality on the training normal data; equal opportunity and demographic parity for unseen anomalies are empirical properties here, not consequences of the theory. The authors should say that plainly. (3) Some empirical claims outrun the tables: on CelebA both proposed methods have lower AUC than Deep SVDD, and on Titanic the locked-in F1 of 66.66 with 1.00 fairness ratio suggests the detector is flagging everything, which makes the fairness comparison trivial there.\n\nBottom line: this is a solid, fixable paper with a useful idea and mostly honest experiments. I would send it to review. The authors should be asked to correct Eq. (11), repair the Sinkhorn proof, either bound or explicitly de-emphasize Assumptions 2 and 3, and soften the CelebA and Titanic claims.","headline":"A genuinely useful fairness-through-target-distribution idea for unsupervised AD, with a sensible threshold-free metric, but the printed objective and the transfer assumptions need fixing before the fairness guarantees can be taken at face value.","tokens_in":35112,"tokens_out":5043,"would_cite":true,"duration_ms":56178,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mapping each demographic group to one shared compact distribution enforces group fairness in unsupervised anomaly detection.","keywords":["fairness-aware anomaly detection","unsupervised anomaly detection","group fairness","fair projection","Sinkhorn distance","threshold-free fairness metric","demographic parity","predictive equality"],"falsifier":"Train Im-FairAD on two groups using only normal data, then evaluate on a test set where anomalies for one group are deliberately placed much farther from the normal manifold than anomalies for the other group; if the between-group difference in flagging rates on these anomalies is large while the training normal ADPD is near zero, the transferability assumption fails.","tokens_in":34070,"feed_emoji":"⚖️","tokens_out":6208,"duration_ms":60639,"temperature":0.7,"pith_summary":"The paper's thesis is that group fairness in unsupervised anomaly detection can be achieved by construction, not by adding fairness penalties. It proposes learning one projection that maps the normal training data of every demographic group onto the same simple, compact target distribution; because the anomaly score is the norm of the projected point, identical projected distributions across groups imply identical flagging rates at every threshold. The paper proves this implication (Propositions 1 and 2) and states two assumptions under which the guarantee extends from normal training data to unseen anomalies, yielding equal opportunity or demographic parity. A threshold-free fairness metric, ADPD, is introduced so fairness can be evaluated across all thresholds without choosing a cutoff. On six real-world benchmarks, including balanced and skewed group splits, the method reports a better accuracy-fairness trade-off than existing fair anomaly detectors.","feed_headline":"One shared projection makes anomaly detection fair by construction","feed_subtitle":"FairAD maps each group's normal data to the same compact distribution, so equal scores mean equal flagging rates.","key_machinery":"The central object is a fair projection $h_\\phi$, an encoder trained to minimize, for every protected group $s$, the Sinkhorn distance between $h_\\phi(X_{S=s})$ and a common target distribution $D_z$, together with a reconstruction term that preserves the information needed for detection. The target distribution is a truncated isotropic Gaussian, chosen because its compactness gives a reliable density-based decision boundary and its identity across groups makes the induced score function group-fair. The argument runs through the chain Proposition 2 to Proposition 1: zero total distribution distance implies $h(x)$ is independent of the protected attribute, which implies that any threshold-based decision function on top of $h$ yields equal flagging probabilities across groups.","core_discovery":"The central claim is that a single learned projection $h$ can simultaneously serve as the anomaly score and as the fairness mechanism. The paper shows that if the normal data of every demographic group are mapped to the same simple, compact target distribution (a truncated isotropic Gaussian), then the projected features are independent of the protected attribute, so for any threshold the flagging probability is equal across groups. More precisely, Proposition 2 states that when the summed Sinkhorn distance between each group's projected distribution and the common target is zero, equation (9) holds and $h(x)$ is independent of the sensitive attribute; Proposition 1 then guarantees that any score function built on $h$ satisfies predictive equality on the training data. Under the stated transferability assumptions, the same reasoning extends to equal opportunity or demographic parity on unseen anomalies. The paper also introduces ADPD, a threshold-free fairness metric, and reports experiments on six benchmarks showing that the proposed Im-FairAD and Ex-FairAD achieve a better trade-off between detection accuracy and fairness than existing fair anomaly detection baselines.","pith_inferences":["The common-target mechanism could be transferred to other one-class problems, such as fraud or medical screening, where the cost of a false alarm differs by group, by choosing the target distribution to reflect those costs.","A direct extension would estimate the ratio of fairness violation on unseen anomalies to that on training normal data, turning the paper's transferability assumption into an empirically checkable quantity rather than a postulate.","Because the fairness guarantee is stated at the level of the learned representation, the same projection could be paired with any downstream thresholding rule, inviting a compositional view of fair anomaly detection."],"forward_implications":["If the projection is learned well, the resulting detector satisfies predictive equality on training normal data for any threshold, without any explicit fairness regularization term.","Under the transferable fairness assumption, the same score distribution implies equal opportunity on unseen anomalies, so fairness is not limited to the training distribution.","The threshold-free ADPD metric gives a single number summarizing fairness over all thresholds, resolving the sensitivity of fairness-ratio evaluations.","Skewed group sizes do not by themselves inflate unfairness, because each group is matched to the same target distribution; the reported ADPD stays nearly flat from balanced to skewed splits.","The method is applicable beyond tabular data, with experiments on image data showing the same shared-projection mechanism works."],"supporting_citations":[{"why":"Supplies the Sinkhorn distance used to measure how far each group's projected data is from the common target distribution.","marker":"[Cuturi, 2013]"},{"why":"Provides Deep SVDD, the deep one-class baseline whose fairness is studied and which FairAD is compared against.","marker":"[Ruff et al., 2018]"},{"why":"Defines the adversarial Deep Fair SVDD baseline and the threshold-based fairness ratio the paper contrasts with ADPD.","marker":"[Zhang and Davidson, 2021]"},{"why":"Provides FairOD, a fairness-regularized autoencoder baseline, and the equal-opportunity discussion in unsupervised anomaly detection.","marker":"[Shekhar et al., 2021]"},{"why":"Provides CFAD, a counterfactual-fairness baseline, and the assumption that anomalies arise as perturbed normal samples.","marker":"[Han et al., 2023]"},{"why":"Defines predictive equality, the training-level fairness notion guaranteed by the shared target distribution.","marker":"[Chouldechova, 2017]"},{"why":"Defines equal opportunity, the fairness notion recovered when transferable fairness holds.","marker":"[Hardt et al., 2016]"},{"why":"Provides LOF, the density-based local outlier baseline used in all benchmark comparisons.","marker":"[Breunig et al., 2000]"}],"fun_headline_variants":["One projection to a common distribution gives fair anomaly scores","FairAD maps each group to the same compact target for fairness","Throw away thresholds: new metric gauges anomaly fairness globally","Better accuracy and fairness trade-off via shared projection","Equal flagging rates by design: FairAD's single projection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method never sees anomalies during training, so the guarantee that fair flagging rates on normal data carry over to anomalies rests on the assumption that the fairness violation on unseen anomalies is at most a small constant times the violation on training normal data.","fun_headline_variants_meta":{"raw":{"variants":["One projection to a common distribution gives fair anomaly scores","FairAD maps each group to the same compact target for fairness","Throw away thresholds: new metric gauges anomaly fairness globally","Better accuracy and fairness trade-off via shared projection","Equal flagging rates by design: FairAD's single projection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3300,"prompt_tokens":945,"completion_tokens":2355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2289}},"tokens_in":561,"tokens_out":2355,"duration_ms":16126,"temperature":1.0,"reasoning_tokens":2289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:58:44.209472+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Im-FairAD on two groups using only normal data, then evaluate on a test set where anomalies for one group are deliberately placed much farther from the normal manifold than anomalies for the other group; if the between-group difference in flagging rates on these anomalies is large while the training normal ADPD is near zero, the transferability assumption fails.","supporting_citations":[{"cited_title":"Towards fair deep anomaly detection","cited_arxiv_id":null,"evidence_quote":"Defines the adversarial Deep Fair SVDD baseline and the threshold-based fairness ratio the paper contrasts with ADPD."},{"cited_title":"Fairod: Fairness-aware outlier detection","cited_arxiv_id":null,"evidence_quote":"Provides FairOD, a fairness-regularized autoencoder baseline, and the equal-opportunity discussion in unsupervised anomaly detection."}],"review_version":1}