Pith. sign in

REVIEW 3 major objections 4 minor 50 references

Counterfactually fair image classifiers do not automatically satisfy group fairness when a latent attribute is correlated with but not caused by the sensitive attribute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 02:44 UTC pith:C62AXVXH

load-bearing objection Solid new CF benchmarks + a clean theorem explaining why CF fails to imply GF on faces; the CTF validity assumption is soft but not fatal, and the work is worth engaging. the 3 major comments →

arxiv 2607.06603 v1 pith:C62AXVXH submitted 2026-07-07 cs.CV cs.AI

Do Counterfactually Fair Image Classifiers Satisfy Group Fairness? -- A Theoretical and Empirical Study

Sangwon Jung , Sumin Yu , Sanghyuk Chun , Taesup Moon This is my paper
classification cs.CV cs.AI
keywords counterfactual fairnessgroup fairnessimage classificationlatent attributesknowledge distillationfacial attribute editingequalized odds
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper establishes that, for image classifiers, achieving counterfactual fairness does not guarantee group fairness, unlike findings on tabular data. The reason is a latent attribute G that correlates with the sensitive attribute without being caused by it, such as hair length with secondary sex characteristics. The authors build two carefully filtered counterfactual face datasets from CelebA and LFW so both fairness notions can be measured on the same images, prove that group disparity is bounded by a model's sensitivity to G, and show that a simple distillation method that reduces that sensitivity lets counterfactually fair models also meet group fairness.

Core claim

In image classification a model that is counterfactually fair with respect to a sensitive attribute A can still violate equalized odds whenever a latent attribute G is correlated with A but not caused by it; the group-fairness gap is then upper-bounded by the model's sensitivity to G, and becomes small once that sensitivity is deliberately reduced.

What carries the argument

Theorem 4.1, which upper-bounds the disparity of equalized odds of any counterfactually fair classifier by the maximum change in its predictions when only the latent attribute G is altered.

Load-bearing premise

The edited images that survive human majority-vote filtering are true counterfactuals: only attributes caused by the sensitive attribute change while every non-descendant, including the target label and the latent G, stays fixed.

What would settle it

Train a counterfactually fair model whose rate of flipped predictions under controlled G edits is essentially zero, yet still observe a large equalized-odds gap on CelebA-CF or LFW-CF; that single observation would violate the bound in Theorem 4.1.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper constructs CelebA-CF and LFW-CF by editing test images of CelebA/LFW with InstructPix2Pix (sex as sensitive attribute) followed by multi-annotator filtering and reliability checks, enabling joint evaluation of counterfactual fairness (CF, via counterfactual disparity CD) and group fairness (GF, via disparity of equalized odds DEO). Empirically, CF-aware methods (CP, Scratch+aug, etc.) reduce CD relative to Scratch but leave DEO high or worse (Table 2). Under an image SCM that admits a latent G correlated with but not caused by A (Figure 2), Theorem 4.1 upper-bounds DEO of any CF classifier by its sensitivity to G. The authors introduce Counterfactual Knowledge Distillation (CKD) that combines CP with distillation from a Scratch teacher presumed more robust to G, and show on CIFAR-10B, CelebA-CF and LFW-CF that CKD simultaneously lowers CD and DEO (Tables 3–5, H.1). Code and datasets are released.

Significance. If the central claim holds, the work is a substantial contribution to fairness in computer vision: it supplies the first real-image CF evaluation benchmarks built on standard GF datasets, supplies a clean theoretical explanation (via G and the image SCM) for why the CF-implies-GF implication observed on tabular data fails for images, and supplies a simple, effective baseline (CKD) that restores the implication once reliance on G is reduced. Strengths that raise the contribution include a complete proof of Theorem 4.1 under the stated Markov assumptions (Appendix A), multi-seed experiments with standard deviations, ablations on teacher robustness and incomplete CTF samples, a fully controllable synthetic dataset (CIFAR-10B), and public release of both filtered CTF pairs and training code. These elements make the result falsifiable and reusable.

major comments (3)
  1. [Section 2, Table 1] Section 2 and Table 1: The reliability check only reports majority-vote accuracy on the sensitive flip and three non-sensitive attributes (blond hair, gray hair, smiling). The ten-attribute filter used for curation does not exhaust non-descendants of A; residual generative artifacts (makeup removal bias explicitly noted by the authors, unintended identity/lighting shifts, or subtle changes to hair length itself) can remain. Because CD is measured exclusively on these filtered pairs and the CF-does-not-imply-GF diagnosis rests on the gap between CD and DEO, imperfect counterfactuals can systematically understate true CD and confound the subsequent success of CKD. Stronger validation (additional attributes, automated identity-preservation metrics, or a sensitivity analysis that injects controlled residual edits) is required for the empirical half of the central claim.
  2. [Theorem 4.1, §4.2] Theorem 4.1 (and the surrounding discussion in §4.2): The derived inequality is only an upper bound on DEO by max d_θ,XY(XG,X′G). It correctly shows that a CF classifier that is sensitive to G can exhibit large DEO in the worst case, but does not establish that the observed high DEO of CP (Table 2) is caused by G-sensitivity rather than other unmodeled factors. The equality case requires dθ,XY ≡ 0 for all XY; the paper never verifies this condition. While RFP experiments (Figure 3, Table 3) are consistent with the bound, a tighter causal link (e.g., an intervention that isolates G while holding all other non-descendants fixed) is needed before the theorem can be invoked as the explanation for the image-versus-tabular discrepancy.
  3. [§5.1, Table C.2] Section 5.1 and the construction of training CTFs: CF-aware methods (including CKD) are trained on unfiltered IP2P edits while evaluation uses the human-filtered CelebA-CF/LFW-CF pairs. The high filtering failure rates (Table C.2: 0.57–0.80) imply a non-trivial distribution shift between train and test counterfactuals. This shift can inflate the apparent benefit of CKD (or of any method that sees the same generative artifacts at train time) and weakens the claim that reduced reliance on G alone restores GF. Either the training CTFs should also be filtered, or an ablation that quantifies the effect of the train–test CTF mismatch is required.
minor comments (4)
  1. [§4] Section 4 title and running text contain the typo “Theoritical”; correct to “Theoretical”.
  2. [Figure 1, A.1] Figure 1 and A.1 captions should explicitly state the Image-CFG / Text-CFG values used for the displayed examples so that readers can reproduce the visual quality.
  3. [Eq. (5)] Equation (5) defines the target vector fTi as the average of teacher representations; a short remark clarifying whether stop-gradient is applied (standard in KD) would remove ambiguity for re-implementers.
  4. [Table H.1] Table H.1 and the main comparison tables omit the standard-deviation columns that appear only in the appendix (G.1); for a journal version the main tables should include them or at least a note pointing to the appendix.

Circularity Check

0 steps flagged

No load-bearing circularity: Theorem 4.1 is a standard Markov/SCM bound independent of measured DEO/CD; CKD and empirical claims rest on held-out CTF pairs and an independently measured Scratch teacher, not on fitted self-predictions.

full rationale

The central theoretical claim (Theorem 4.1) derives an upper bound on DEO for any counterfactually fair classifier from the Markov properties of the postulated image SCM (Figure 2) and the existence of a non-descendant latent G connected to A by an unblocked path. The proof (Appendix A) expands the DEO difference via Bayes, cancels terms by d-separation (Ŷ ⊥ A | XY,XG and XY ⊥ XG,A | Y), and applies Jensen; the bound is expressed solely in terms of the model’s own sensitivity dθ,XY(XG,X′G). Nothing in the derivation embeds the numerical CD or DEO values later reported, nor does it redefine DEO as the quantity being bounded. Empirically, CD is measured on newly constructed, human-filtered held-out pairs (CelebA-CF/LFW-CF) that are never used for training the evaluated models; DEO and RFP are measured on the original test sets and on separately edited hair-length images, respectively. CKD simply distills representations from a vanilla Scratch teacher whose lower RFP is measured independently; the teacher is not optimized for the reported GF metric. Self-citations (e.g., the authors’ prior MFD work) appear only as ordinary baselines and are not used to justify uniqueness or to force the main inequality. The validity of the IP2P counterfactuals is an external modeling assumption, not a circular reduction of the derivation to its own inputs. Consequently the paper contains at most the ordinary, non-load-bearing self-reference common to any research program and scores 1.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The load-bearing claim rests on a standard SCM Markov assumption plus three domain-specific modeling choices (image SCM topology, existence of a non-descendant G, and validity of diffusion-edited counterfactuals). Free parameters are ordinary training hyper-parameters and the synthetic correlation alpha; none are fitted to the final DEO/CD numbers. The latent G is the sole invented entity and is given a partial external handle via hair-length editing and RFP.

free parameters (4)
  • CKD strength mu and CP strength lambda
    Grid-searched regularization weights that control the trade-off among CE, distillation and pairing losses; selected by lowest CD subject to an accuracy floor.
  • CIFAR-10B correlation alpha
    Controllable skew between A and G used to generate the synthetic diagnostic set; varied from 0.5 to 0.9 to test the theory.
  • IP2P Image-CFG / Text-CFG and denoising steps
    Generation hyper-parameters fixed at 1.8 / 7.5 / 50; chosen by visual inspection before human filtering.
  • Accuracy lower-bound for model selection (98 % of Scratch)
    Ad-hoc threshold that discards overly unfair models; affects which checkpoint is reported for every method.
axioms (4)
  • standard math Markov assumption of the Structural Causal Model: d-separation implies statistical independence
    Invoked throughout Section 4 and the proof of Theorem 4.1 (Appendix A).
  • domain assumption Image SCM topology of Figure 2: latent attributes (Y, A, G) generate image components that compose the observed image X; no unblocked path from A to Y
    Stated in Section 4.1; required for the independence steps that turn the DEO difference into a sum over G-sensitivity.
  • ad hoc to paper G is a non-descendant of A and is connected to A by an unblocked path
    Explicit hypothesis of Theorem 4.1; without it the upper bound on DEO collapses and the CF-GF gap disappears.
  • domain assumption IP2P edits (after human majority-vote filtering) change only attributes caused by A while leaving non-descendants of A intact
    Foundation of the entire evaluation pipeline (Section 2); reliability check reports high but imperfect agreement.
invented entities (1)
  • latent attribute G (e.g., hair length) independent evidence
    purpose: Explains why a CF-fair model can still exhibit large DEO; supplies the quantity whose sensitivity upper-bounds DEO in Theorem 4.1
    Postulated to exist in facial images; given a partial external handle by SDEdit hair-length edits and RFP measurement, but not independently measured for every possible correlate.

pith-pipeline@v1.1.0-grok45 · 32786 in / 3145 out tokens · 47658 ms · 2026-07-11T02:44:14.755456+00:00 · methodology

0 comments
read the original abstract

The notion of algorithmic fairness has been actively explored from various aspects of fairness, such as counterfactual fairness (CF) and group fairness (GF). However, the exact relationship between CF and GF remains to be unclear, especially in image classification tasks; the reason is because we often cannot collect counterfactual samples regarding a sensitive attribute, essential for evaluating CF, from the existing images (\eg, a photo of the same person but with different secondary sex characteristics). In this paper, we construct new image datasets for evaluating CF by using a high-quality image editing method and carefully labeling with human annotators. Our datasets, \oursceleb and \ourslfw, build upon the popular image GF benchmarks; hence, we can evaluate CF and GF simultaneously. We empirically observe that CF does not imply GF in image classification, whereas previous studies on tabular datasets observed the opposite. We theoretically show that it could be due to the existence of a latent attribute $G$ that is correlated with, but not caused by, the sensitive attribute (\eg, secondary sex characteristics are highly correlated with hair length). From this observation, we propose a simple baseline, Counterfactual Knowledge Distillation (CKD), to mitigate such correlation with the sensitive attributes. Extensive experimental results on \oursceleb and \ourslfw demonstrate that CF-achieving models satisfy GF if we successfully reduce the reliance on $G$ (\eg, using CKD).

Figures

Figures reproduced from arXiv: 2607.06603 by Sanghyuk Chun, Sangwon Jung, SuMin Yu, Taesup Moon.

Figure 1
Figure 1. Figure 1: CelebA-CF examples. The counterfactual (CTF) images regarding the “sex attribute” are shown. IP2P can generate high-quality CTF samples by simply adjusting the text instructions without any model retraining. As the first step, we edit the test images of two popular facial image datasets, CelebA [25] and LFW [13]. We choose the “sex” of a person in an image as the sensitive attribute3 and edit the sex-relat… view at source ↗
Figure 2
Figure 2. Figure 2: Image SCM. Blue, gray, and yellow circles represent latent at￾tributes, components of an image and a whole image, respectively. Directed edges indicate a causal relationship from the source to the target. The blue region indicates that there can be any direction of edges between blue nodes. Structural Causal Models (SCMs) are represented as directed acyclic graphs satisfying the conditions specified in [30… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of the correlation of G and A. α indicates how A and G are correlated on CIFAR-10B. skewnesses among (Y , A) and (A, G), respectively; the former ratio is the spurious correlation between Y and A, and the latter one is the correlation between A and G. We then construct the CIFAR-10B dataset by randomly injecting Gaussian or Contrast noise to each CIFAR-10 image at given ratios, as illustrated in Fig… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 7 internal anchors

  1. [1]

    J. R. Anthis and V . Veitch. Causal context connects counterfactual fairness to robust prediction and group fairness. InAdv. Neural Inform. Process. Syst. (NeurIPS), 2023. 1, 2, 5, 6, 21

    work page 2023

  2. [2]

    Brooks, A

    T. Brooks, A. Holynski, and A. A. Efros. Instructpix2pix: Learning to follow image editing instructions. InIEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 18392–18402, 2023. 2, 9, 13, 18, 29

    work page 2023

  3. [3]

    Buolamwini and T

    J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. InConf. Fairness, Accountability and Transparency (FAccT), pages 77–91. PMLR, 2018. 1

    work page 2018

  4. [4]

    S. Dash, V . N. Balasubramanian, and A. Sharma. Evaluating and mitigating bias in image classifiers: A causal perspective using counterfactuals. InIEEE/CVF Winter Conf. App. Comput. Vis. (WACV), pages 915–924, 2022. 2, 5

    work page 2022

  5. [5]

    D’Incà, C

    M. D’Incà, C. Tzelepis, I. Patras, and N. Sebe. Improving fairness using vision-language driven image augmentation. InIEEE/CVF Winter Conf. App. Comput. Vis. (WACV), pages 4695–4704, 2024. 2

    work page 2024

  6. [6]

    Dutta, D

    S. Dutta, D. Wei, H. Yueksel, P.-Y . Chen, S. Liu, and K. Varshney. Is there a trade-off between fairness and accuracy? A perspective using mismatched hypothesis testing. InInt. Conf. Mach. Learn. (ICML), pages 2803–2813. PMLR, 2020. 4

    work page 2020

  7. [7]

    S. Garg, V . Perot, N. Limtiaco, A. Taly, E. H. Chi, and A. Beutel. Counterfactual fairness in text classification through robustness. InAAAI/ACM Conf. AI, Ethics, and Society (AIES), pages 219–226,

    work page

  8. [8]

    Goldblum, L

    M. Goldblum, L. Fowl, S. Feizi, and T. Goldstein. Adversarially robust distillation. InProc. of the AAAI Conf. Artificial Intelligence (AAAI), volume 34, pages 3996–4003, 2020. 22

    work page 2020

  9. [9]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial nets.Adv. Neural Inform. Process. Syst. (NeurIPS), 27, 2014. 2 10

    work page 2014

  10. [10]

    Hardt, E

    M. Hardt, E. Price, and N. Srebro. Equality of opportunity in supervised learning. InAdv. Neural Inform. Process. Syst. (NeurIPS), volume 29, 2016. 5

    work page 2016

  11. [11]

    Harlan and O

    E. Harlan and O. Schnuck. Objective or biased – the questionable use of artificial intelligence in job applications.bayerischer rundfunk, 2021. URL https://interaktiv.br.de/ki-bewerbung/en/. 1

    work page 2021

  12. [12]

    Distilling the Knowledge in a Neural Network

    G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. 7, 22, 25

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  13. [13]

    G. B. Huang, V . Jain, and E. Learned-Miller. Unsupervised joint alignment of complex images. InInt. Conf. Comput. Vis. (ICCV), 2007. 3, 13, 18, 27

    work page 2007

  14. [14]

    B. Y . Idrissi, M. Arjovsky, M. Pezeshki, and D. Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. InConf. Causal Learning and Reasoning (CLeaR), pages 336–351. PMLR, 2022. 9, 19, 24

    work page 2022

  15. [15]

    Jiang and O

    H. Jiang and O. Nachum. Identifying and correcting label bias in machine learning. InInt. Conf. Artificial Intelligence and Statistics (AISTATS), pages 702–712. PMLR, 2020. 9, 19, 20, 24

    work page 2020

  16. [16]

    S. Jung, D. Lee, T. Park, and T. Moon. Fair feature distillation for visual recognition. InIEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12115–12124, 2021. 7, 9, 19, 20, 24

    work page 2021

  17. [17]

    Kamiran and T

    F. Kamiran and T. Calders. Data preprocessing techniques for classification without discrimination. Knowledge and Information Systems (KAIS), 33(1):1–33, 2012. 9, 19, 24

    work page 2012

  18. [18]

    Kennaway

    R. Kennaway. When causation does not imply correlation: Robust violations of the faithfulness axiom. In The Interdisciplinary Handbook of Perceptual Control Theory, pages 49–72. Elsevier, 2020. 21

    work page 2020

  19. [19]

    H. Kim, S. Shin, J. Jang, K. Song, W. Joo, W. Kang, and I.-C. Moon. Counterfactual fairness with disentangled causal effect variational autoencoder. InProc. of the AAAI Conf. Artificial Intelligence (AAAI), volume 35, pages 8128–8136, 2021. 2

    work page 2021

  20. [20]

    D. P. Kingma and P. Dhariwal. Glow: Generative flow with invertible 1x1 convolutions.Adv. Neural Inform. Process. Syst. (NeurIPS), 31, 2018. 5

    work page 2018

  21. [21]

    D. P. Kingma and M. Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  22. [22]

    CausalGAN: Learning Causal Implicit Generative Models with Adversarial Training

    M. Kocaoglu, C. Snyder, A. G. Dimakis, and S. Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training.arXiv preprint arXiv:1709.02023, 2017. 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    M. J. Kusner, J. R. Loftus, C. Russell, and R. Silva. Counterfactual fairness. InAdv. Neural Inform. Process. Syst. (NeurIPS), 2017. 1

    work page 2017

  24. [24]

    Liang, P

    H. Liang, P. Perona, and G. Balakrishnan. Benchmarking algorithmic bias in face recognition: An experimental approach using synthetic faces and human evaluation. InInt. Conf. Comput. Vis. (ICCV), pages 4977–4987, 2023. 2

    work page 2023

  25. [25]

    Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. InInt. Conf. Comput. Vis. (ICCV), pages 3730–3738, 2015. 3, 13, 18, 27

    work page 2015

  26. [26]

    Louizos, U

    C. Louizos, U. Shalit, J. M. Mooij, D. Sontag, R. Zemel, and M. Welling. Causal effect inference with deep latent-variable models.Adv. Neural Inform. Process. Syst. (NeurIPS), 30, 2017. 2

    work page 2017

  27. [27]

    A. S. Luccioni, C. Akiki, M. Mitchell, and Y . Jernite. Stable bias: Analyzing societal representations in diffusion models.arXiv preprint arXiv:2303.11408, 2023. 29

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021. 8, 20

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  29. [29]

    L. S. Nguyen and D. Gatica-Perez. Hirability in the wild: Analysis of online conversational video resumes. IEEE Trans. Multimedia, 18(7):1422–1437, 2016. 1

    work page 2016

  30. [30]

    Pearl et al

    J. Pearl et al. Models, reasoning and inference.Cambridge, UK: CambridgeUniversityPress, 19(2):3, 2000. 5

    work page 2000

  31. [31]

    Peychev, A

    M. Peychev, A. Ruoss, M. Balunovi´c, M. Baader, and M. Vechev. Latent space smoothing for individually fair representations. InEur. Conf. Comput. Vis. (ECCV), pages 535–554. Springer, 2022. 4, 5, 19, 20, 24

    work page 2022

  32. [32]

    S. R. Pfohl, T. Duan, D. Y . Ding, and N. H. Shah. Counterfactual reasoning for fair clinical risk prediction. InMachine Learning for Healthcare Conference (MLHC), pages 325–358. PMLR, 2019. 2 11

    work page 2019

  33. [33]

    Pinto, A

    M. Pinto, A. V . Carreiro, P. Madeira, A. Lopez, and H. Gamboa. The matrix reloaded: Towards counterfac- tual group fairness in machine learning.Journal of Data-centric Machine Learning Research (DMLR),

    work page

  34. [34]

    V . V . Ramaswamy, S. S. Kim, and O. Russakovsky. Fair attribute classification through latent space de-biasing. InIEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9301–9310, 2021. 2

    work page 2021

  35. [35]

    Rosenblatt and R

    L. Rosenblatt and R. T. Witter. Counterfactual fairness is basically demographic parity. InProc. of the AAAI Conf. Artificial Intelligence (AAAI), volume 37, pages 14461–14469, 2023. 1, 2, 5, 21

    work page 2023

  36. [36]

    Russell, M

    C. Russell, M. J. Kusner, J. Loftus, and R. Silva. When worlds collide: integrating different counterfactual assumptions in fairness.Adv. Neural Inform. Process. Syst. (NeurIPS), 30, 2017. 2, 4, 5, 6, 9, 19, 20, 22, 24, 25

    work page 2017

  37. [37]

    Scimeca, S

    L. Scimeca, S. J. Oh, S. Chun, M. Poli, and S. Yun. Which shortcut cues will dnns choose? a study from the parameter-space perspective. InInt. Conf. Learn. Represent. (ICLR), 2022. 7

    work page 2022

  38. [38]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  39. [39]

    M. Wang, W. Deng, J. Hu, X. Tao, and Y . Huang. Racial faces in the wild: Reducing racial bias by information maximization adaptation network. InInt. Conf. Comput. Vis. (ICCV), pages 692–702, 2019. 1

    work page 2019

  40. [40]

    H. Wu, G. Bezold, M. Günther, T. Boult, M. C. King, and K. W. Bowyer. Consistency and accuracy of celeba attribute values. InIEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3257–3265, 2023. 3

    work page 2023

  41. [41]

    D. Xu, S. Yuan, L. Zhang, and X. Wu. Fairgan: Fairness-aware generative adversarial networks. In2018 IEEE International Conference on Big Data (Big Data), pages 570–575. IEEE, 2018. 2

    work page 2018

  42. [42]

    SenSeI: Sensitive Set Invariance for Enforcing Individual Fairness

    M. Yurochkin and Y . Sun. Sensei: Sensitive set invariance for enforcing individual fairness.arXiv preprint arXiv:2006.14168, 2020. 4, 5, 19, 20, 24

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  43. [43]

    M. B. Zafar, I. Valera, M. G. Rogriguez, and K. P. Gummadi. Fairness constraints: Mechanisms for fair classification. InInt. Conf. Artificial Intelligence and Statistics (AISTATS), pages 962–970. PMLR, 2017. 1, 9, 19, 20, 24

    work page 2017

  44. [44]

    Zhang, K

    F. Zhang, K. Kuang, L. Chen, Y . Liu, C. Wu, and J. Xiao. Fairness-aware contrastive learning with partially annotated sensitive attributes. InInt. Conf. Learn. Represent. (ICLR), 2022. 2

    work page 2022

  45. [45]

    B. Zi, S. Zhao, X. Ma, and Y .-G. Jiang. Revisiting adversarial robustness distillation: Robust soft labels make student better. InInt. Conf. Comput. Vis. (ICCV), pages 16443–16452, 2021. 7 12 Checklist The checklist follows the references. Please read the checklist guidelines carefully for information on how to answer these questions. For each question, ...

    work page 2021

  46. [46]

    (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] Our claims are reflected accurately

    For all authors... (a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] Our claims are reflected accurately. (b) Did you describe the limitations of your work? [Yes] See Section B (c) Did you discuss any potential negative societal impacts of your work? [Yes] See Section B (d) Have you ...

    work page

  47. [47]

    (a) Did you state the full set of assumptions of all theoretical results? [Yes] See Section 4 (b) Did you include complete proofs of all theoretical results? [Yes] See Section A

    If you are including theoretical results... (a) Did you state the full set of assumptions of all theoretical results? [Yes] See Section 4 (b) Did you include complete proofs of all theoretical results? [Yes] See Section A

    work page

  48. [48]

    for benchmarks)

    If you ran experiments (e.g. for benchmarks)... (a) Did you include the code, data, and instructions needed to reproduce the main experi- mental results (either in the supplemental material or as a URL)? [Yes] See Section D (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section D (c) Did ...

    work page

  49. [49]

    (a) If your work uses existing assets, did you cite the creators? [Yes] We cited Liu et al

    If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... (a) If your work uses existing assets, did you cite the creators? [Yes] We cited Liu et al. [25], Huang et al. [13] and Brooks et al. [2]. (b) Did you mention the license of the assets? [Yes] See Section C.3 (c) Did you include any new assets either in the supp...

    work page

  50. [50]

    sex attribute

    If you used crowdsourcing or conducted research with human subjects... (a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] See Figure C.1, C.2, and C.3. (b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] (c) Did you i...

    work page arXiv