{"id":"157855bf-396d-4468-8cfd-c65c26621dad","arxiv_id":"2505.18532","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A distributionally robust AUC-fairness method with TV-distance bounds that preserves group fairness when protected-group labels are noisy.","lead":"This paper proposes a method to keep classification models fair across demographic groups even when the recorded group labels are wrong or noisy, specifically when the goal is ranking quality (AUC) rather than plain accuracy. It adds a robustness layer, a way to estimate how noisy the group labels are using CLIP, and tests on tabular socioeconomic data and deepfake detection.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never establishes that the chosen γ is a valid TV upper bound: tabular experiments set γ to the per-sample flip rate, but Lemma 4.2 bounds pairwise mismatch probability, and for minority pairs the true TV can exceed γ; CLIP-based image estimates are uncalibrated.","rationale":"Good-faith review: Theorem 4.1 is a straightforward and correct conditional statement, and the DRO relaxation in Eq. (8) is a sensible adaptation of Wang et al. (2020); the proof's use of TV is legitimate because the constant term in h cancels. The reader's concern about CLIP-based γ estimation is well placed, and I extend it: the problem is not only that CLIP may be miscalibrated, but that in the tabular experiments γ is set to the per-sample flip rate, which is mathematically the wrong quantity relative to Lemma 4.2 and can be far below the true TV for minority group pairs. This means the headline 'fairness theoretical guarantees' is not actually demonstrated in any reported experiment. The issue is fixable—validate γ on data with known clean labels, use the pairwise mismatch probability, or state that γ is a heuristic—so the paper should not be rejected outright. The empirical results, especially the consistent violation reductions and the clean-label FF++ check in Table 7, give real support to the method as a heuristic. The verdict remains CONDITIONAL: the paper should add a direct validation of the γ bound and either adjust the tabular γ choice or temper the theoretical claim.","tokens_in":25327,"tokens_out":22565,"duration_ms":187716,"concrete_test":"Use a dataset with known clean protected labels (e.g., Bank's original labels as clean, or the human-corrected FF++ labels from Lin et al. 2025). Inject the same noise process as in §5.2 with γ=0.5 on Bank (and γ=0.02 on FF++), compute the true TV distance TV(p_{z,z'}, \\hat p_{z,z'}) for every group pair from the known flip mechanism, and compare with the γ used in Table 1 (or Eq. (10)). If TV exceeds γ for any group pair, the precondition of Theorem 4.1 is violated; the paper would need to use a valid upper bound such as the pairwise mismatch probability 1-(1-γ)^2 and re-run the experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee in Theorem 4.1 is conditional on γ_{z,z'} being a true upper bound on the total variation between the clean and noisy group-conditional distributions. The paper never establishes this condition. In the tabular experiments (§5.2), noise is injected by flipping a fraction γ of protected labels and the same γ is used as the DRO radius in Eq. (9). Under independent per-sample flips, Lemma 4.2 bounds TV by the pairwise mismatch probability 1-(1-γ)^2, which is larger than γ; for minority-group pairs the induced TV can be much larger still. For example, on Bank (group ratio 0.05:1) at γ=0.5, the posterior probability that a sample whose noisy label is the minority group truly belongs to that group is about 0.05, so if the group-conditional feature distributions are even moderately separated, the clean pair distribution is nearly orthogonal to the noisy-conditioned mixture, giving TV close to 1 - q_z q_{z'} ≈ 0.998 for the (minority, minority) pair—far above the chosen radius. The paper provides no computation showing TV ≤ γ for these pairs. On images, γ=0.02 is obtained by counting CLIP-predicted mismatches via Eq. (10); there is no calibration showing that CLIP's verdicts are correct or that the resulting count is an upper bound, and the sensitivity analysis in Fig. 3 selects γ by test violation rather than by verifying TV. Consequently, the theoretical guarantee is not instantiated in any experiment, and the empirical results support the method's practical value but not the claimed fairness guarantees.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributionally robust optimization (DRO) approach for AUC fairness when protected group labels are noisy. The authors define a group-level AUC fairness metric, formulate a constrained optimization problem with total variation (TV) distance constraints, and prove a theorem bounding the clean-group fairness violation by the TV distance between clean and noisy group-conditional distributions. They estimate the TV bound using a CLIP-based mismatch counter for images and use the known flip rate for tabular data. They design an SGDA algorithm with sharpness-aware minimization and evaluate on three tabular datasets and four deepfake detection datasets, reporting lower AUC fairness violation than several baselines. The central claim is that this is the first robust AUC fairness method with theoretical guarantees under noisy protected groups.","tokens_in":25679,"tokens_out":13753,"duration_ms":106812,"significance":"If the theoretical guarantee were instantiated correctly, the paper would be a useful contribution to the fairness literature, extending DRO-based robustness from classification losses to pairwise AUC metrics. The paper addresses a practical and underexplored problem, provides a clear formal adaptation of Wang et al. (2020), and includes extensive experiments covering multiple noise levels, high noise, multi-group settings, and a clean-label image benchmark. The code is made publicly available. However, the central guarantee is not actually connected to the experimental setup as described, and several assumption checks are missing. The empirical results are promising but the theoretical claims are currently overstated.","major_comments":[{"comment":"The definition of g_{z,z'}(θ) in Eq. (5) implicitly assumes that every pair consists of one positive and one negative sample; without this paired-sampling interpretation, E[h(θ)|Z=z,Z'=z'] does not equal AUC_{z,z'}(θ) - AUC(θ) because the conditional expectation includes the label indicators and the group proportions. The paper should state this sampling assumption explicitly, as the theoretical guarantee in Theorem 4.1 and the empirical objective in Eq. (9) both rely on it.","section":"Section 4.1, Eq. (5)"},{"comment":"The tabular experiments set the DRO radius γ to the per-sample flip rate, but Lemma 4.2 bounds TV by the conditional pairwise mismatch probability, which under independent flips is 1-(1-γ)^2 = 2γ-γ^2, strictly larger than γ. Moreover, Lemma 4.2 requires the marginal pairwise group probabilities to be unchanged by noise, which is not satisfied under uniform random flipping when the group prior is imbalanced (e.g., Bank has group ratio 0.05:1). As a result, Theorem 4.1 does not apply to the reported tabular results, and the central theoretical guarantee is not instantiated in the experiments.","section":"Section 5.1-5.2 and Lemma 4.2"},{"comment":"The CLIP-based estimate of γ in Eq. (10) is used as the TV upper bound for the image experiments, but the paper provides no evidence that CLIP's verdicts yield an upper bound on the true mismatch probability; indeed, the text concedes that CLIP predictions are not perfectly reliable. The sensitivity analysis in §5.3 selects γ by test-set violation, which is post-hoc and does not validate the bound. Without a calibration or a conservative correction, the guarantee in Theorem 4.1 cannot be claimed for the deployed image models.","section":"Section 4.3, Eq. (10)"},{"comment":"The image benchmark results in Table 2 evaluate fairness using the noisy protected labels (as stated in Appendix F.2), whereas the paper's objective is clean-group fairness. The reported violations on DFDC, DFD, and Celeb-DF may therefore not reflect true clean-group disparity; only the FF++ clean-label evaluation in Table 7 addresses this, and it covers a single dataset. The cross-domain empirical claims in §5.2 are consequently not fully supported.","section":"Section 5.2 and Appendix F.2"}],"minor_comments":[{"comment":"The phrase 'optimizing AUC over accuracy' should be 'optimizing AUC instead of accuracy' to avoid ambiguity.","section":"Abstract"},{"comment":"For the Default dataset at noise level 0.1, 'reducing it by 5.14%' is ambiguous; the absolute reduction from 0.0701 to 0.0187 is 5.14 percentage points, which corresponds to about a 73% relative reduction. Please rephrase.","section":"Section 5.2"},{"comment":"The sentence 'We added experiments with one more baseline to address your concern' is a reviewer-response artifact and should be removed from the manuscript.","section":"Appendix F.2"},{"comment":"The Min/Max column header uses '↓' instead of '↑', which is inconsistent with the other tables.","section":"Table 7"},{"comment":"The reported reductions of 3.40% and 1.70% do not match the gap values shown in the figure (e.g., 0.095 to 0.075 is about a 21% relative reduction); please reconcile the numbers.","section":"Figure 3 and Section 5.3"},{"comment":"The sentence 'we regard the group label of the test image as clean' should read 'training image', since the estimation is performed on the training set.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a revised version with embedded reviewer responses in the appendix; such text should be removed. The central technical issue is that the theoretical guarantee is not connected to the experimental setup: the chosen γ is not a valid TV upper bound in the tabular experiments, and the CLIP-based estimate in the image experiments is not validated. The authors need to either revise the experiments to satisfy the theorem's conditions or substantially weaken the theoretical claims. The paper's topic and scope are appropriate, but the claim of being 'first' with guarantees is currently overstated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nHere's my take on arXiv:2505.18532. The genuinely new thing is the adaptation of the DRO-with-noisy-groups framework to pairwise AUC fairness, and the experiments are broad and consistent. That part is worth a careful look. But the paper's central theoretical guarantee is not actually instantiated anywhere: the γ used as the DRO radius is never shown to be a valid upper bound on the TV distance between clean and noisy group-conditional distributions.\n\nOn the tabular side, the paper injects noise by flipping a fraction γ of protected labels and then uses that same γ as the radius in Eq. (9). Lemma 4.2 bounds TV by the pairwise mismatch probability P[(Z,Z')≠(bZ,bZ') | (bZ,bZ')=(z,z')], which is exactly the quantity that gets inflated by pairwise coupling. For independent per-sample flips at rate γ, that probability is 1-(1-γ)^2, which is larger than γ. More importantly, the lemma's assumption that the pairwise group marginals are unaffected by noise is false for imbalanced groups under random flips. So the guarantee in Theorem 4.1 does not apply to the tabular experiments as run.\n\nOn the image side, γ=0.02 comes from counting CLIP-predicted mismatches via Eq. (10). There's no calibration showing that CLIP's verdicts bound the true mismatch rate, and the sensitivity analysis chooses γ on the test set by looking at fairness violation, not by verifying the TV bound. So the same problem: the theory is sound conditional on a valid γ, but the γ is a free hyperparameter in practice.\n\nI want to give credit where it's due. The problem is real and previously unaddressed for AUC. The formulation in Eq. (9) is clean, the minimax/SGDA algorithm is a sensible extension, and the empirical results are consistent: the method reliably lowers fairness violation on tabular and deepfake benchmarks compared to strong baselines. The paper also ships code and is honest in its limitation section about not having a tabular noise estimator. The CLIP-based estimator is a nice practical idea, though the framing that CLIP is 'not used for classification' is misleading: deciding whether a label is clean or noisy is itself a classification decision.\n\nMinor issues: Table 2 has no error bars, SGDA convergence is not analyzed, and the sensitivity analysis selects γ on the test set. These are addressable.\n\nBottom line: this deserves a serious referee, not a desk rejection. The main revision request should be an honest treatment of γ — either a calibrated estimate with a proof that it's an upper bound, or a statement that γ is tuned as a hyperparameter and the theoretical guarantee is only conditional. As it stands, the paper demonstrates practical value but not the promised fairness guarantee.","headline":"The DRO-to-AUC extension is plausible and the experiments are consistent, but the TV bound γ is never shown to be valid, so the stated fairness guarantee doesn't bind.","tokens_in":26216,"tokens_out":4465,"would_cite":true,"duration_ms":37195,"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":"Noisy protected-group labels need not break AUC fairness","keywords":["AUC fairness","noisy protected groups","distributionally robust optimization","total variation distance","group fairness","deepfake detection","pairwise ranking fairness"],"falsifier":"On a deepfake test set with human-corrected demographic labels, count how often the vision-language model calls a pair clean when the protected label is actually flipped; if that disagreement exceeds the margin by which the true TV distance surpasses the selected $\\gamma$, the Theorem 4.1 guarantee ceases to apply to the deployed model.","tokens_in":25127,"feed_emoji":"⚖️","tokens_out":11592,"duration_ms":92627,"temperature":0.7,"pith_summary":"The paper tackles a gap in AUC-based fairness: existing methods assume protected attributes like gender and race are accurate, but real labels are often noisy, whether from survey response bias or from demographic attributes inferred on AI-generated faces. Its central claim is that a model trained with noisy protected-group labels can still guarantee fairness on the true, clean groups, up to a slack equal to the total variation distance between the clean and noisy group-conditional distributions. To achieve this, it casts AUC fairness as a distributionally robust optimization problem, enforces the fairness constraint over every distribution inside a TV-distance ball, and estimates the ball radius from data via a lemma relating the bound to the probability that clean and noisy group labels disagree. The paper presents the resulting algorithm and reports that it attains the lowest AUC fairness violation on tabular benchmarks with simulated label flips and on deepfake-detection benchmarks with naturally noisy demographic annotations.","feed_headline":"Fair AUC survives noisy protected-group labels","feed_subtitle":"A DRO objective keeps AUC fair when protected labels are wrong, within a provable shift bound.","key_machinery":"The load-bearing object is the pairwise AUC fairness function $g_{z,z'}(\\theta)=AUC_{z,z'}(\\theta)-AUC(\\theta)$, which measures how far the ranking of positive examples from group $z$ against negative examples from group $z'$ deviates from the overall ranking. The method wraps this function in a distributionally robust constraint, requiring the violation to stay non-positive for every pairwise distribution $\\tilde{p}_{z,z'}$ inside the $\\ell^{1,1}$ ball $\\{\\tilde{p}:\\|\\tilde{p}-\\hat{p}_{z,z'}\\|_{1,1}\\le 2\\gamma_{z,z'}\\}$. A minimax Lagrangian objective over the model parameters, the multipliers $\\lambda_{z,z'}$, and the adversarial pairwise distributions turns the constraint into a trainable objective. The radius $\\gamma_{z,z'}$ comes from Lemma 4.2, which bounds the total variation distance between clean and noisy group-conditional distributions by the conditional probability that the clean and noisy group labels disagree; in image experiments, this probability is estimated by prompting a vision-language model with `a photo of {group}` versus `a photo without {group}` and comparing cosine similarities.","core_discovery":"The central claim is that training under noisy protected groups does not have to sacrifice AUC fairness. Theorem 4.1 shows: if a model satisfies the fairness constraints $\\hat{g}_{z,z'}(\\theta)\\le 0$ for all noisy pairwise groups $(\\hat{Z},\\hat{Z}')=(z,z')$, then the clean-group fairness violation obeys $g_{z,z'}(\\theta)\\le\\gamma_{z,z'}$, where $\\gamma_{z,z'}$ is any upper bound on the total variation distance $TV(p_{z,z'},\\hat{p}_{z,z'})$ between the clean and noisy group-conditional pairwise distributions. Lemma 4.2 turns that bound into a countable quantity, namely the probability that clean and noisy group labels disagree on a positive-negative pair, and the paper estimates it for images by prompting a vision-language model with group-specific phrases. On tabular benchmarks with simulated label flips and on deepfake benchmarks with naturally noisy demographic annotations, the method reports the lowest AUC fairness violation among compared methods, and it keeps the advantage on a human-corrected clean test set.","pith_inferences":["Because Lemma 4.2 assumes equal group priors between clean and noisy labels, a dataset whose noise mechanism changes group base rates, such as annotation errors concentrated in one demographic, would need a larger $\\gamma$ than the estimated mismatch rate; that case is not tested in the paper.","A cheap calibration extension would be to compare the vision-language model's clean/noisy verdicts against a few hundred human-checked labels and inflate $\\gamma$ by the observed disagreement, making the image guarantee robust to the estimator's own error.","The paper's tabular experiments choose $\\gamma$ by grid search rather than by its own noise-estimation pipeline, and the paper states this as a limitation; a direct tabular analogue of the prompt-based estimator would let the same guarantee chain apply outside images.","Since the pairwise fairness function bounds every group pair against the overall AUC, the method implicitly controls both intra-group and inter-group disparities; an explicit two-metric bound could let practitioners trade them off separately."],"forward_implications":["A model trained with noisy group labels can carry a certificate: for each pair of groups, the clean-group fairness violation is at most the estimated TV-ball radius $\\gamma$.","Because the constraint is enforced for every distribution inside the TV ball, the method should keep the violation low even when the exact noise rate is unknown, as long as $\\gamma$ covers the true shift; the paper's high-noise runs up to a 0.9 flip rate are consistent with this.","On deepfake detection, the method yields the lowest group-AUC gap among compared methods on four benchmarks, and it remains best on a human-corrected clean FF++ test set, indicating the improvement is not an artifact of noisy evaluation labels.","Removing the robustness constraint raises the violation, and removing the sharpness-aware minimization term raises it further, so both the DRO objective and the flatness-seeking optimization contribute to the reported fairness performance."],"supporting_citations":[{"why":"Supplies the distributionally robust optimization formulation and the total-variation ball over which the fairness constraint is made robust.","marker":"(Duchi & Namkoong, 2021)"},{"why":"Provides the template for bounding fairness violation under noisy protected groups by total variation distance, which Theorem 4.1 adapts from non-ranking losses to pairwise AUC fairness.","marker":"(Wang et al., 2020)"},{"why":"Defines the group-level AUC formulation and the MinimaxFairAUC baseline that is the direct comparator, as well as the stratified sampling operator used in the optimization.","marker":"(Yang et al., 2023)"},{"why":"Supplies the vision-language model whose text and image encoders are used in Eq. (10) to estimate the label-mismatch rate $\\gamma$.","marker":"(Radford et al., 2021)"},{"why":"Provides the deepfake-detection setup, the FF++/DFDC/DFD/Celeb-DF evaluation protocol, and the PG-FDD baseline, along with the human-corrected clean FF++ test set used in the robustness check.","marker":"(Lin et al., 2024)"},{"why":"Prior work on fair classification with noisy protected attributes that motivates the setting and supports the comparison against noisy-group baselines.","marker":"(Celis et al., 2021)"}],"fun_headline_variants":["AUC fairness survives when protected groups are noisy","First robust AUC fairness method for noisy groups","Provable AUC fairness despite mislabeled demographics","DRO keeps AUC fair even with wrong group labels","Noise-robust AUC fairness with theoretical bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee holds only if the estimated radius $\\gamma$ really upper-bounds the total variation distance between clean and noisy group-conditional distributions; in the image experiments, $\\gamma=0.02$ is computed from a vision-language model's agreement with group-specific prompts, with no proof that the model's verdicts bound the true mismatch probability and no human calibration.","fun_headline_variants_meta":{"raw":{"variants":["AUC fairness survives when protected groups are noisy","First robust AUC fairness method for noisy groups","Provable AUC fairness despite mislabeled demographics","DRO keeps AUC fair even with wrong group labels","Noise-robust AUC fairness with theoretical bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000265,"raw_usage":{"total_tokens":1594,"prompt_tokens":916,"completion_tokens":678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":607}},"tokens_in":532,"tokens_out":678,"duration_ms":5797,"temperature":1.0,"reasoning_tokens":607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:29:47.779376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a deepfake test set with human-corrected demographic labels, count how often the vision-language model calls a pair clean when the protected label is actually flipped; if that disagreement exceeds the margin by which the true TV distance surpasses the selected $\\gamma$, the Theorem 4.1 guarantee ceases to apply to the deployed model.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the distributionally robust optimization formulation and the total-variation ball over which the fairness constraint is made robust."},{"cited_title":"Robust optimization for fairness with noisy protected groups","cited_arxiv_id":null,"evidence_quote":"Provides the template for bounding fairness violation under noisy protected groups by total variation distance, which Theorem 4.1 adapts from non-ranking losses to pairwise AUC fairness."},{"cited_title":"L., Varshney, K","cited_arxiv_id":null,"evidence_quote":"Defines the group-level AUC formulation and the MinimaxFairAUC baseline that is the direct comparator, as well as the stratified sampling operator used in the optimization."},{"cited_title":"Preserving fairness generalization in deepfake detection","cited_arxiv_id":null,"evidence_quote":"Provides the deepfake-detection setup, the FF++/DFDC/DFD/Celeb-DF evaluation protocol, and the PG-FDD baseline, along with the human-corrected clean FF++ test set used in the robustness check."},{"cited_title":"E., Huang, L., Keswani, V., and Vishnoi, N","cited_arxiv_id":null,"evidence_quote":"Prior work on fair classification with noisy protected attributes that motivates the setting and supports the comparison against noisy-group baselines."}],"review_version":1}