Pith. sign in

REVIEW 4 major objections 4 minor 70 references

Component-Based Fairness in Face Attribute Classification with Bayesian Network-informed Meta Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A Bayesian Network-informed reweighting method cuts face-component prediction disparity on CelebA, and the resulting component fairness also narrows gender gaps.

desk verdict New fairness notion worth thinking about, but the key derivation drops the conditioning that the loss requires, and Table 3's ablation values are swapped relative to Table 1; as submitted, the claims are not supported. read the letter →

arxiv 2505.01699 v1 pith:BTWTWGQP submitted 2025-05-03 cs.CV cs.AI

Yifan Liu , Ruichen Yao , Yaokun Liu , Ruohan Zong , Zelin Li , Yang Zhang , Dong Wang This is my paper · ORCID
classification cs.CVcs.AI
keywords facecomponentfairnessattributeclassificationBayesianNetworkmetalearningsamplereweightingequalopportunitytruepositiveratedisparityCelebA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces face component fairness, a fairness notion defined over biological face parts such as lips, eyebrows, nose, chin, and beard, and argues that it can be optimized directly during face attribute classification. To do so, the authors propose Bayesian Network-informed Meta Reweighting (BNMR), which models the dependencies among face component attributes and the classifier's current predictions with a Bayesian Network, then uses that calibrated bias signal to reweight training samples through meta-learning. On CelebA, BNMR reports lower true-positive-rate disparity and disparate-impact gap than recent debiasing baselines for both smiling and attractiveness classification, with roughly unchanged accuracy. The paper also claims that improving component-level fairness improves gender fairness, suggesting that face component fairness could serve as a surrogate objective for demographic fairness.

What carries the argument

The load-bearing object is the Bayesian Network calibrator, a probabilistic graphical model over the chosen face component attributes plus a node for the classifier prediction $\hat{Y}$, whose structure is learned by exhaustive search with a K2 score, pruned by chi-square independence tests, and whose conditional probability tables are fit by maximum likelihood. During training, variable elimination queries on this network supply $P(A=a)$ and $P(A=a\mid \hat{Y}=1)$, which combine with the classifier's positive-class confidence into the Bayes-rule factor $Z$ that scales the fairness loss; the network is updated every $N$ training steps so that its belief about model bias tracks the evolving classifier. The second mechanism is the meta-learning reweighting loop: a temporary classifier update is computed on the task loss, the calibrated fairness loss is evaluated on small micro validation sets balanced per face component, and its gradient updates a per-sample weight vector, normalized by a temperature-controlled softmax, before the main classifier update. Together these components let the method optimize true-positive-rate disparity without enumerating exponentially many intersectional subpopulations and without requiring a globally fair exemplar training set.

What would settle it

Compute the empirical TPRD of Equation 1 on a held-out CelebA split, and compare it with the Bayesian-calibrated fairness loss of Equation 5 on the same batches; if the two rank samples differently or disagree in magnitude, the calibrator is not evaluating the objective being reported. A cleaner test: replace the calibrator with a direct empirical estimate of $P(\hat{Y}=1\mid A=a,Y=1)$ from the micro validation sets, retrain BNMR, and check whether the TPRD gains survive.

Watch

Extended reading notes

Core claim

The central claim is that a face classifier can be debiased with respect to biological face components—big lips, arched eyebrows, big nose, double chin, no beard—by treating those components as overlapping, interdependent sensitive groups rather than independent demographic attributes. BNMR replaces a balanced exemplar dataset with a Bayesian Network calibrator that estimates the joint distribution over face component attributes, the prediction target, and the classifier's predictions, and is updated online during training. The fairness loss is evaluated through this calibrator, and its gradient is used in a meta-learning loop to update per-sample weights, pushing the classifier toward equal true positive rates across each component attribute. The paper reports that BNMR consistently outperforms baselines on CelebA for smiling and attractiveness classification, and that the model with the best component-level fairness is also the fairest with respect to gender; the authors read this as evidence that, for their selected attributes, face component fairness can serve as a proxy for demographic fairness.

Load-bearing premise

The load-bearing premise is in Section 3.4.1: the Bayes-rule expansion of $P(\hat{Y}=1|A=a)$ is taken to evaluate a fairness metric defined with the extra conditioning $Y=1$; if that conditioning is not harmless, the reweighting gradient optimizes a different quantity than the reported TPRD.

Editorial extensions

If this is right

  • For the two tasks and five component attributes studied, reducing component-level disparity also reduces gender disparity: BNMR reports the lowest gender TPRD and DIG among all compared methods.
  • The method scales with the number of components: fairness gains grow from three to five face component attributes while accuracy stays flat or improves slightly.
  • Because the Bayesian Network replaces balanced sampling, training-time fairness evaluation avoids the exponential growth of attribute intersections and the label-scarcity problem faced by prior reweighting methods.
  • Bayesian calibration constrains disparity redistribution across correlated attributes: with the calibrated L1 loss, reducing bias on one component does not freely increase bias on another.
  • At inference the method adds no extra model parameters, so it remains as efficient as vanilla training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If component-level fairness is a reliable proxy for demographic fairness beyond gender, the same machinery could debias face models without collecting sensitive demographic labels, a privacy-preserving route that the paper only hints at; it tests gender only, so the proxy claim needs verification on race and age.
  • Equation 5 drops the $Y=1$ conditioning that appears in the TPRD definition; swapping in a direct empirical estimate of $P(\hat{Y}=1\mid A=a,Y=1)$ and re-running the experiments would show whether the reported gains come from the calibrated surrogate or from the reweighting scheme itself.
  • The same Bayesian Network-informed reweighting could transfer to other domains with fine-grained, correlated sensitive attributes, such as medical imaging or hiring, but that transfer is not tested here.
  • A synthetic experiment with a known attribute-dependency graph could separate the contribution of dependency modeling from the contribution of online calibration; the paper's ablation removes the calibrator entirely, conflating the two.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces 'face component fairness,' a group fairness notion defined over biological face attributes rather than demographic groups, and proposes Bayesian Network-informed Meta Reweighting (BNMR), which uses a Bayesian Network calibrator to estimate the joint distribution of face component attributes, the prediction target, and the classifier output, and then guides a meta-learning sample-reweighting procedure. Experiments on CelebA for attractiveness and smiling detection report improved True Positive Rate Disparity (TPRD) and Disparate Impact Gap (DIG) relative to several baselines, and the paper further claims that component-level debiasing improves gender fairness.

Significance. If the method and results were correct, the paper would open a useful new axis for fairness auditing and mitigation in face analysis: biological face components as sensitive attributes, complementing demographic categories. The Bayesian-network calibrator is a plausible mechanism for handling attribute dependencies and label scarcity, and the authors release publicly available code and make falsifiable empirical predictions. However, the central derivation of the fairness loss, the reported ablation table, and the proxy claim for demographic fairness each contain load-bearing inconsistencies that currently prevent accepting the empirical claims.

major comments (4)
  1. [Section 3.4.1, Eq. (5) and Definition 3.1/Eq. (3)] Definition 3.1 and Eq. (3) define the training objective with P(Ŷ=1|A=a,Y=1), i.e., the true positive rate per attribute group. Section 3.4.1 instead states that evaluating δ_m requires the conditional probability P(Ŷ=1|A=a), and Eq. (5) expands this unconditional probability via Bayes' rule. The conditioning on Y=1 is dropped without justification, and recovering P(Ŷ=1|A=a,Y=1) from Eq. (5) would require additional factors (e.g., P(Y=1|A=a) or P(Y=1|Ŷ=1,A=a)) that are not provided. Because L_fair is the gradient signal driving the entire reweighting mechanism, the paper does not show that the method optimizes the reported TPRD metric; either the implementation follows Eq. (5) and optimizes a different quantity, or it follows Eq. (3) and the Bayesian calibrator's role in the derivation is unverified.
  2. [Table 3 and Section 4.5] Several ablation rows in Table 3 are exact column swaps of baseline rows in Table 1. For example, the Attractiveness row 'w/o Reweighting' (80.13, DIG 16.84, TPRD 23.24) equals the Vanilla 5-attribute row with DIG and TPRD interchanged (80.13, 23.24, 16.84), and the Smiling row 'w/o Bayesian Calibration' (92.23, 4.17, 4.71) equals the FORML 5-attribute row with the two fairness columns exchanged (92.23, 4.71, 4.17). These coincidences strongly suggest a transcription error or that the ablations were not run as described, and they invalidate the conclusions of Section 4.5 about the individual contribution of each pipeline component.
  3. [Definition 4.1 and Section 4.2] Definition 4.1 defines DIG using P(Ŷ=1|A=a1,Y=1)/P(Ŷ=1|A=a2,Y=1), i.e., a ratio of true positive rates, yet the text in Section 4.2 says DIG compares 'ratios of positive outcomes,' which normally refers to the unconditional positive prediction rate. Eq. (5) computes an unconditional probability, so it is unclear whether the Bayesian calibrator feeds a conditional or an unconditional estimator. This ambiguity prevents the reader from verifying which fairness metric the method actually optimizes and how the numbers in Tables 1, 3, and 5 are computed.
  4. [Section 4.6 and Abstract] The paper concludes that 'face component fairness can serve as a reasonable proxy for demographic fairness,' but the experiments only show that BNMR improves component fairness and also has the best gender fairness in Table 5. There is no controlled test of whether the component-level debiasing transfers to gender, such as comparing a model debiased only on the five facial attributes against one debiased directly on gender, or ablating the connection. The mutual-information argument in Section 4.6 is heuristic and does not establish a proxy relationship; as stated in the abstract, this is a central claim and is currently unsupported.
minor comments (4)
  1. [Table 3 footnote] The abbreviation note defines DIG as 'Demographic Intersectional Gap,' whereas Section 4.2 and Definition 4.1 use 'Disparate Impact Gap'; please make the notation consistent.
  2. [References [66] and [67]] References [66] and [67] are listed as the same paper (Zeng et al., 'On adversarial robustness of demographic fairness in face attribute recognition'); one entry should be replaced with the intended distinct paper or removed.
  3. [Section 4.2, second paragraph] The number '50000' should be written as '50,000' for readability.
  4. [Section 4.8] The statement that L1-norm disparity loss 'does not inherently penalize disparity redistribution across attributes' is confusing: both L1 and L2 penalize per-attribute disparity; the intended point is that L1 does not couple attributes without the Bayesian calibration. Please rephrase for clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the BNMR fairness gains are measured on held-out test data and the Bayesian calibrator is estimated from data rather than being defined by the target metric; self-citations and the correlated-attribute choice are not load-bearing.

full rationale

The central claim is not circular. The fairness loss in Eq. 3 is evaluated on micro fairness validation sets drawn from a validation split, and the reported DIG/TPRD results in Tables 1 and 5 are computed on the CelebA test split, so the improvements are not forced by construction. The Bayesian Network calibrator is learned from training-data attribute annotations via K2-score structural search and MLE parameter estimation; it is not defined in terms of the test fairness metric, and no fitted parameter is renamed as a prediction. The self-citations in Sec. 4.2 (refs 65-67, which share co-authors Yang Zhang and Dong Wang) only set the experimental protocol: the 50k training subsample, smile/attractiveness targets, and lightCNN backbone. These are not load-bearing for the method or for the surrogate claim, so they do not constitute circularity. The demographic-surrogate conclusion is scoped to the selected attributes, some of which (No Beard, Arched Eyebrow) are strongly gender-correlated; that is a selection/confounding concern rather than a definitional reduction. Two non-circular correctness issues should be weighed separately. First, Sec. 3.4.1/Eq. 5 drops the Y=1 conditioning: Definition 3.1 and Eq. 3 require P(Yhat=1|A=a,Y=1), while the Bayes expansion computes P(Yhat=1|A=a), so the written derivation does not show that the calibrator gradient optimizes the claimed TPRD. Second, Table 3's 'w/o Reweighting' and 'w/o Bayesian Calibration' rows appear to duplicate Table 1 baseline rows with DIG and TPRD column values interchanged (e.g., Attractiveness w/o Reweighting shows 16.84 DIG and 23.24 TPRD versus Vanilla's 23.24 DIG and 16.84 TPRD), weakening the ablation narrative. These are derivation and data-integrity concerns, not circular reductions.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim depends on the BN structure and probability estimates learned from CelebA labels, on the micro fairness validation sets, and on the assumption that TPRD is the right metric; none of these are derived from first principles, so they are inventoried as assumptions or tuned parameters.

free parameters (6)
  • lambda (fairness-task trade-off) = not reported
    Trade-off weight in Eq. (2) between task loss and fairness loss. The actual algorithm replaces this with learned sample weights; no value, grid range, or sensitivity analysis is reported.
  • tau (softmax temperature) = 0.9 (grid-searched over [0.1, 0.2, ..., 1.0])
    Temperature in the softmax normalization of the reweighting vector (Eq. 6). Selected by grid search; the paper notes optimal values are between 0 and 2.
  • Prior sample number for Bayesian Network updates = 80 (grid-searched over [40, 80, 160])
    Controls the frequency and sample size of the maximum-likelihood belief update for the prediction node in the BN. Listed in Table 2 as a hyperparameter; its relationship to the every N steps update is not fully specified.
  • Learning rate = 1e-4 (grid-searched over [1e-5, 1e-4, 1e-3])
    Global learning rate for the face classifier in the meta-training loop.
  • Fairness validation size = 20
    Number of samples in each micro fairness validation set. The TPRD loss gradient is estimated from very small validation sets, which is a high-variance design choice not discussed.
  • Chi-square pruning p threshold = p < 0.05
    Edges not meeting statistical significance at p < 0.05 are pruned from the learned Bayesian Network structure; this threshold is a free modeling choice.
assumptions (6)
  • standard math Bayes' rule
    Used in Eq. (5) to expand P(Ŷ=1|A=a); the expansion itself is standard, though it is applied to a conditional that does not match the TPRD definition.
  • domain assumption The joint distribution of face component attributes and predictions is faithfully represented by a DAG learned with K2Score and chi-square pruning on CelebA
    Section 3.4 and Appendix B assume the learned Bayesian Network structure captures the true dependencies among face component attributes and the prediction variable; the calibrator's probability estimates inherit this assumption.
  • ad hoc to paper The five chosen binary attributes are a valid operationalization of face component fairness
    Section 4.2 selects five attributes as the operationalization of face component fairness; the paper acknowledges that real deployment should inform attribute selection with fairness auditing, making this an ad hoc modeling choice.
  • domain assumption Equal Opportunity (TPRD) is the appropriate fairness metric for face component fairness
    Section 3.1 chooses Equal Opportunity as the fairness notion for all face component attributes without comparing alternatives such as demographic parity or calibration; the reported DIG metric is a further ad hoc variant.
  • ad hoc to paper The optimal predictor f*(X) captures all relevant information about Y, so MI-based reasoning about debiasing difficulty applies
    Section 4.6's mutual-information argument assumes the optimal classifier encodes all relevant information about Y, so that attribute-Y correlations determine debiasing difficulty; this idealization supports the surrogate conclusion.
  • ad hoc to paper The prediction node can be appended to the BN with uniform initialization and updated with MLE from the classifier's own outputs during training
    Section 3.4 appends the classifier's own prediction as a node in the BN and updates it with MLE during training, so the calibrator's estimate of model bias depends on the very model being trained; this self-referential loop is assumed stable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Component-Based Fairness in Face Attribute Classification with Bayesian Network-informed Meta Learning." pith.science (2026). https://pith.science/paper/BTWTWGQP

@misc{pith2026250501699,
  author       = {Pith},
  title        = {Pith review of: Component-Based Fairness in Face Attribute Classification with Bayesian Network-informed Meta Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BTWTWGQP}},
  note         = {Machine review of arXiv:2505.01699}
}
read the original abstract

The widespread integration of face recognition technologies into various applications (e.g., access control and personalized advertising) necessitates a critical emphasis on fairness. While previous efforts have focused on demographic fairness, the fairness of individual biological face components remains unexplored. In this paper, we focus on face component fairness, a fairness notion defined by biological face features. To our best knowledge, our work is the first work to mitigate bias of face attribute prediction at the biological feature level. In this work, we identify two key challenges in optimizing face component fairness: attribute label scarcity and attribute inter-dependencies, both of which limit the effectiveness of bias mitigation from previous approaches. To address these issues, we propose \textbf{B}ayesian \textbf{N}etwork-informed \textbf{M}eta \textbf{R}eweighting (BNMR), which incorporates a Bayesian Network calibrator to guide an adaptive meta-learning-based sample reweighting process. During the training process of our approach, the Bayesian Network calibrator dynamically tracks model bias and encodes prior probabilities for face component attributes to overcome the above challenges. To demonstrate the efficacy of our approach, we conduct extensive experiments on a large-scale real-world human face dataset. Our results show that BNMR is able to consistently outperform recent face bias mitigation baselines. Moreover, our results suggest a positive impact of face component fairness on the commonly considered demographic fairness (e.g., \textit{gender}). Our findings pave the way for new research avenues on face component fairness, suggesting that face component fairness could serve as a potential surrogate objective for demographic fairness. The code for our work is publicly available~\footnote{https://github.com/yliuaa/BNMR-FairCompFace.git}.

Figures

Figures reproduced from arXiv: 2505.01699 by the authors.

Figure 1
Figure 1. Illustration of face component dependency rela [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our method, which learns a weight [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Weighting sensitivity analysis on 𝜏 in re-weighting. The value 𝜏 = 0.9 provides the best fairness result. A value of 𝜏 that is either excessively high or low diminishes the effectiveness of bias mitigation efforts. redistribution is restricted by the learned dependency structure, meaning the increase in 𝐴2 disparity is bounded. As a result, the total fairness loss ensures fairness adjustments remain probabilisti￾cal… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Bayesian Network structure learned for all face [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Bayesian Network structure learned on training set [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 39 canonical work pages

  1. [1]

    Irene Blair, Charles Judd, Melody Sadler, and Christopher Jenkins. 2002. The Role of Afrocentric Features in Person Perception: Judging by Features and Component-Based Fairness in Face Attribute Classification with Bayesian Network-informed Meta Learning FAccT ’25, June 23–26, 2025, Athens, Greece Categories. Journal of Personality and Social Psychology 8...

    work page doi:10.1037/0022-3514.83.1.5 2002

  2. [2]

    Meredith Broussard. 2023. More than a Glitch: Confronting Race, Gender, and Ability Bias in Tech. The MIT Press, Cambridge, MA. https://mitpress.mit.edu/ 9780262047654/more-than-a-glitch/

    work page 2023

  3. [3]

    Junyi Chai, Taeuk Jang, and Xiaoqian Wang. 2022. Fairness without demograph- ics through knowledge distillation. Advances in Neural Information Processing Systems 35 (2022), 19152–19164

    work page 2022

  4. [4]

    Zhisheng Chen. 2023. Ethics and discrimination in artificial intelligence-enabled recruitment practices. Humanities and Social Sciences Communications 10 (09 2023). https://doi.org/10.1057/s41599-023-02079-x

    doi:10.1057/s41599-023-02079-x 2023

  5. [5]

    Andrew Cotter, Heinrich Jiang, Maya Gupta, Serena Wang, Taman Narayan, Seungil You, and Karthik Sridharan. 2019. Optimization with non-differentiable constraints with applications to fairness, recall, churn, and other goals. Journal of Machine Learning Research 20, 172 (2019), 1–59

    work page 2019

  6. [6]

    Andrew Cotter, Heinrich Jiang, and Karthik Sridharan. 2019. Two-player games for efficient non-convex constrained optimization. InAlgorithmic Learning Theory. PMLR, 300–332

    work page 2019

  7. [7]

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. 2019. Class- balanced loss based on effective number of samples. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9268–9277

    2019

  8. [8]

    Mostafa Dehghani, Arash Mehrjou, Stephan Gouws, Jaap Kamps, and Bernhard Scholkopf. 2018. FIDELITY-WEIGHTED LEARNING. (2018)

    work page 2018

Show all 70 references
  1. [9]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Im- ageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition . 248–255. https://doi.org/10.1109/CVPR. 2009.5206848 ISSN: 1063-6919

    2009

  2. [10]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference . 214–226

    2012

  3. [11]

    Jennifer Eberhardt, Phillip Goff, Valerie Purdie, and Paul Davies. 2004. Seeing Black: Race, Crime, and Visual Processing. Journal of Personality and Social Psychology 87 (12 2004), 876–93. https://doi.org/10.1037/0022-3514.87.6.876

    2004 doi

  4. [12]

    Eberhardt, Paul G

    Jennifer L. Eberhardt, Paul G. Davies, Valerie J. Purdie-Vaughns, and Sheri Lynn Johnson. 2006. Looking Deathworthy: Perceived Stereotypicality of Black Defendants Predicts Capital-Sentencing Outcomes. Psychological Science 17, 5 (2006), 383–386. https://doi.org/10.1111/j.1467...

    2006

  5. [13]

    Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. 2019. A guide to deep learning in healthcare. Nature Medicine 25, 1 (Jan. 2019), 24–29. https://doi.org/10.1038/s4...

    2019 doi

  6. [14]

    Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. 2015. Certifying and removing disparate impact. In pro- ceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining. 259–268

    2015

  7. [15]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. In International conference on machine learning. PMLR, 1126–1135

    2017

  8. [16]

    Neil Zhenqiang Gong and Bin Liu. 2018. Attribute Inference Attacks in Online Social Networks. ACM Transactions on Privacy and Security 21, 1 (Feb. 2018), 1–30. https://doi.org/10.1145/3154793

    2018 doi

  9. [17]

    Aditya Gulati, Bruno Lepri, and Nuria Oliver. 2024. Lookism: The overlooked bias in computer vision. arXiv:2408.11448 [cs.CV] https://arxiv.org/abs/2408.11448

    2024 arXiv

  10. [18]

    Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equality of opportunity in supervised learning. Advances in neural information processing systems 29 (2016)

    2016

  11. [19]

    Tatsunori Hashimoto, Megha Srivastava, Hongseok Namkoong, and Percy Liang

  12. [20]

    Ayanna Howard and Jason Borenstein. 2018. The Ugly Truth About Ourselves and Our Robot Creations: The Problem of Bias and Social Inequity. Science and Engineering Ethics 24, 5 (Oct. 2018), 1521–1536. https://doi.org/10.1007/s11948- 017-9975-2

    2018 doi

  13. [21]

    Heinrich Jiang and Ofir Nachum. 2019. Identifying and Correcting Label Bias in Machine Learning. arXiv:1901.04966 [cs.LG] https://arxiv.org/abs/1901.04966

    2019 arXiv

  14. [22]

    Kahn and A

    H. Kahn and A. W. Marshall. 1953. Methods of Reducing Sample Size in Monte Carlo Computations. Journal of the Operations Research Society of America 1, 5 (1953), 263–278. https://www.jstor.org/stable/166789 Publisher: INFORMS

    1953

  15. [23]

    Jian Kang, Tiankai Xie, Xintao Wu, Ross Maciejewski, and Hanghang Tong. 2022. Infofair: Information-theoretic intersectional fairness. In 2022 IEEE International Conference on Big Data (Big Data) . IEEE, 1455–1464

    2022

  16. [24]

    Michael Kearns, Seth Neel, Aaron Roth, and Zhiwei Steven Wu. 2018. Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness. In Proceedings of the 35th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80) , Jen...

    2018

  17. [25]

    Michael P Kim, Amirata Ghorbani, and James Zou. 2019. Multiaccuracy: Black-box post-processing for fairness in classification. InProceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society . 247–254

    2019

  18. [26]

    Daphne Koller and Nir Friedman. 2010. Probabilistic graphical models: principles and techniques (nachdr. ed.). MIT Press, Cambridge, Mass

    2010

  19. [27]

    Camila Kolling, Victor Araujo, Adriano Veloso, and Soraia Raupp Musse. 2023. Mitigating bias in facial analysis systems by incorporating label diversity. Com- puters & Graphics 116 (2023), 173–184. https://doi.org/10.1016/j.cag.2023.08.021

    2023 doi

  20. [28]

    Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. 2017. Counterfac- tual fairness. Advances in neural information processing systems 30 (2017)

    2017

  21. [29]

    Preethi Lahoti, Alex Beutel, Jilin Chen, Kang Lee, Flavien Prost, Nithum Thain, Xuezhi Wang, and Ed Chi. 2020. Fairness without demographics through adver- sarially reweighted learning. Advances in neural information processing systems 33 (2020), 728–740

    2020

  22. [30]

    Nova Hadi Lestriandoko, Raymond Veldhuis, and Luuk Spreeuwers. 2022. The contribution of different face parts to deep face recognition.Frontiers in Computer Science 4 (2022). https://doi.org/10.3389/fcomp.2022.958629

    2022

  23. [31]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2020. Focal Loss for Dense Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 42, 2 (2020), 318–327. https://doi.org/10.1109/TPAMI.2018. 2858826

    2020 doi

  24. [32]

    Davis, and Emily M

    Bryson Lingenfelter, Sara R. Davis, and Emily M. Hand. 2022. A Quantitative Analysis of Labeling Issues in the CelebA Dataset. In Advances in Visual Com- puting, George Bebis, Bo Li, Angela Yao, Yang Liu, Ye Duan, Manfred Lau, Rajiv Khadka, Ana Crisan, and Remco Chang (Eds.). ...

    2022

  25. [33]

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. 2015. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision. 3730–3738

    2015

  26. [34]

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. 2015. Deep Learning Face Attributes in the Wild. In 2015 IEEE International Conference on Computer Vision (ICCV). 3730–3738. https://doi.org/10.1109/ICCV.2015.425

    2015 doi

  27. [35]

    Keith Maddox. 2004. Perspectives on Racial Phenotypicality Bias. Personality and social psychology review : an official journal of the Society for Personality and Social Psychology, Inc8 (02 2004), 383–401. https://doi.org/10.1207/s15327957pspr0804_4

    2004 doi

  28. [36]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2021. A Survey on Bias and Fairness in Machine Learning.ACM Comput. Surv. 54, 6, Article 115 (July 2021), 35 pages. https://doi.org/10.1145/3457607

    2021 doi

  29. [37]

    Parkhi, A

    O. Parkhi, A. Vedaldi, and A. Zisserman. 2015. Deep face recognition. BMVC 2015 - Proceedings of the British Machine Vision Conference 2015 (2015). https: //ora.ox.ac.uk/objects/uuid:a5f2e93f-2768-45bb-8508-74747f85cad1 Publisher: British Machine Vision Association

    2015

  30. [38]

    Tian Qiu, Arjun Nichani, Rasta Tadayon, and Haewon Jeong. 2024. Gone With the Bits: Benchmarking Bias in Facial Phenotype Degradation Under Low-Rate Neural Compression. In ICML 2024 Next Generation of AI Safety Workshop . https: //openreview.net/forum?id=zIrvyQdIG4

    2024

  31. [39]

    Novi Quadrianto, Viktoriia Sharmanska, and Oliver Thomas. 2019. Discover- ing Fair Representations in the Data Domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  32. [40]

    Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. 2018. Learning to reweight examples for robust deep learning. In International conference on machine learning. PMLR, 4334–4343

    2018

  33. [41]

    Willy E Rice. 1996. Race, gender, redlining, and the discriminatory access to loans, credit, and insurance: An historical and empirical analysis of consumers who sued lenders and insurers in federal and state courts, 1950-1995. San Diego L. Rev. 33 (1996), 583

    1996

  34. [42]

    Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner

    Joseph P. Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner. 2020. Face Recognition: Too Bias, or Not Too Bias?. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    2020

  35. [43]

    Ignacio Serna, Aythami Morales, Julian Fierrez, Manuel Cebrian, Nick Obradovich, and Iyad Rahwan. 2020. Algorithmic Discrimination: Formulation and Explo- ration in Deep Learning-based Face Biometrics. In AAAI Workshop on Artificial Intelligence Safety (SafeAI) (CEUR-WS, Vol. ...

    2020

  36. [45]

    Sijie Shen, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. 2017. Fooling Neural Networks in Face Attractiveness Evaluation: Adversarial Ex- amples with High Attractiveness Score But Low Subjective Score. In 2017 IEEE Third International Conference on Multimedia Big D...

    2017 doi

  37. [46]

    Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. 2019. Meta-weight-net: Learning an explicit mapping for sample weighting. Advances in neural information processing systems 32 (2019). FAccT ’25, June 23–26, 2025, Athens, Greece Liu et al

    2019

  38. [47]

    Changjian Shui, Gezheng Xu, Qi Chen, Jiaqi Li, Charles X Ling, Tal Arbel, Boyu Wang, and Christian Gagné. 2022. On learning fairness and accuracy on multiple subgroups. Advances in Neural Information Processing Systems 35 (2022), 34121– 34135

    2022

  39. [48]

    parts and wholes

    James W. Tanaka and Diana Simonyi. 2016. The “parts and wholes” of face recogni- tion: a review of the literature.Quarterly journal of experimental psychology (2006) 69, 10 (Oct. 2016), 1876–1889. https://doi.org/10.1080/17470218.2016.1146780

    2006

  40. [49]

    TIME Magazine. 2025. Are Face Detection Cameras Racist? https://time.com/ archive/6906847/are-face-detection-cameras-racist/ Accessed: 2025-03-18

    2025

  41. [50]

    Sahil Verma and Julia Rubin. 2018. Fairness definitions explained. In Proceedings of the International Workshop on Software Fairness . ACM, Gothenburg Sweden, 1–7. https://doi.org/10.1145/3194770.3194776

    2018

  42. [51]

    Esteban Villalobos, Domingo Mery, and Kevin Bowyer. 2022. Fair Face Verification by Using Non-Sensitive Soft-Biometric Attributes. IEEE Access 10 (2022), 30168– 30179. https://doi.org/10.1109/ACCESS.2022.3158967

    2022

  43. [52]

    Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez

  44. [53]

    Zhibo Wang, Xiaowei Dong, Henry Xue, Zhifei Zhang, Weifeng Chiu, Tao Wei, and Kui Ren. 2022. Fairness-aware adversarial perturbation towards bias miti- gation for deployed deep models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10379–10388

    2022

  45. [54]

    Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. 2020. Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2020

  46. [55]

    Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Lai Jian-Huang, and Tie-Yan Liu. 2018. Learning to teach with dynamic loss functions. Advances in neural information processing systems 31 (2018)

    2018

  47. [56]

    Xiang Wu, Ran He, Zhenan Sun, and Tieniu Tan. 2018. A light CNN for deep face representation with noisy labels. IEEE transactions on information forensics and security 13, 11 (2018), 2884–2896

    2018

  48. [57]

    Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. 2018. Fairgan: Fairness- aware generative adversarial networks. In 2018 IEEE international conference on big data (big data) . IEEE, 570–575

    2018

  49. [58]

    Bobby Yan*, Skyler Seto*, and Nicholas Apostoloff. 2022. FORML: Learning to Reweight Data for Fairness. In ICML Workshop. https://arxiv.org/abs/2202.01719

    2022 arXiv

  50. [59]

    Forest Yang, Mouhamadou Cisse, and Sanmi Koyejo. 2020. Fairness with Overlap- ping Groups; a Probabilistic Perspective. In Advances in Neural Information Pro- cessing Systems, Vol. 33. Curran Associates, Inc., 4067–4078. https://proceedings. neurips.cc/paper/2020/hash/29c0605a...

    2020

  51. [60]

    Ruichen Yao, Ziteng Cui, Xiaoxiao Li, and Lin Gu. 2022. Improving Fairness in Image Classification via Sketching. In Workshop on Trustworthy and Socially Responsible Machine Learning, NeurIPS 2022

    2022

  52. [61]

    Seyma Yucer, Amir Atapour Abarghouei, Noura Al Moubayed, and Toby P. Breckon. 2024. Disentangling Racial Phenotypes: Fine-Grained Control of Race- related Facial Phenotype Characteristics. In 2024 International Joint Conference on Neural Networks (IJCNN) . 1–10. https://doi.or...

    2024

  53. [62]

    Seyma Yucer, Furkan Tektas, Noura Al Moubayed, and Toby Breckon. 2022. Measuring Hidden Bias within Face Recognition via Racial Phenotypes. 3202–

    2022

  54. [63]

    Seyma Yucer, Furkan Tektas, Noura Al Moubayed, and Toby Breckon. 2024. Racial Bias within Face Recognition: A Survey. ACM Comput. Surv. 57, 4, Article 105 (Dec. 2024), 39 pages. https://doi.org/10.1145/3705295

    2024 doi

  55. [64]

    Bianca Zadrozny. 2004. Learning and evaluating classifiers under sample selection bias. In Proceedings of the twenty-first international conference on Machine learning (ICML ’04) . Association for Computing Machinery, New York, NY, USA, 114. https://doi.org/10.1145/1015330.1015425

    2004

  56. [65]

    Huimin Zeng, Zhenrui Yue, Ziyi Kou, Yang Zhang, Lanyu Shang, and Dong Wang. 2023. Fairness-aware training of face attribute classifiers via adversarial robustness. Knowledge-Based Systems 264 (March 2023), 110356. https://doi.org/ 10.1016/j.knosys.2023.110356

    2023

  57. [67]

    Huimin Zeng, Zhenrui Yue, Lanyu Shang, Yang Zhang, and Dong Wang. 2023. On adversarial robustness of demographic fairness in face attribute recognition. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (Macao, P.R.China) (IJCAI ’23)...

    2023 doi

  58. [68]

    Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. 2018. Mitigating un- wanted biases with adversarial learning. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society . 335–340

    2018

  59. [69]

    arched eyebrows

    Guanhua Zhang, Yihua Zhang, Yang Zhang, Wenqi Fan, Qing Li, Sijia Liu, and Shiyu Chang. 2022. Fairness reprogramming. Advances in Neural Information Processing Systems 35 (2022), 34347–34362. Component-Based Fairness in Face Attribute Classification with Bayesian Network-infor...

    2022

  60. [2018]

    InProceed- ings of the 35th International Conference on Machine Learning

    Fairness Without Demographics in Repeated Loss Minimization. InProceed- ings of the 35th International Conference on Machine Learning . PMLR, 1929–1938. https://proceedings.mlr.press/v80/hashimoto18a.html ISSN: 2640-3498

    1929

  61. [2019]

    In 2019 IEEE/CVF International Conference on Computer Vision (ICCV)

    Balanced Datasets Are Not Enough: Estimating and Mitigating Gender Bias in Deep Image Representations. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV) . IEEE, Seoul, Korea (South), 5309–5318. https://doi. org/10.1109/ICCV.2019.00541

    2019

  62. [3211]

    https://doi.org/10.1109/WACV51458.2022.00326

    2022

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.