pith. sign in

arxiv: 2604.11250 · v1 · submitted 2026-04-13 · 💻 cs.CV

Variational Latent Entropy Estimation Disentanglement: Controlled Attribute Leakage for Face Recognition

\"Unsal \"Ozt\"urk (1) , Vedrana Krivoku\'ca Hahn (1) , Sushil Bhattacharjee (1) , S\'ebastien Marcel (1 , 2) ((1) Idiap Research Institute , Martigny , Switzerland , (2) UNIL
show 2 more authors
Lausanne Switzerland)
This is my paper

Pith reviewed 2026-05-10 14:56 UTC · model grok-4.3

classification 💻 cs.CV
keywords face recognitiondisentanglementprivacyvariational autoencoderfairnessattribute leakageentropy estimationbias reduction
0
0 comments X

The pith

A variational autoencoder with entropy estimation separates gender and ethnicity from face recognition embeddings while preserving verification accuracy.

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

The paper presents Variational Latent Entropy Estimation Disentanglement (VLEED) as a post-hoc transformation applied to pretrained face embeddings. It trains a variational autoencoder to produce a latent representation in which a chosen categorical attribute is isolated from the information needed for identity verification. This matters because raw embeddings leak demographic details that downstream systems can exploit for privacy breaches or biased matching decisions. The method realizes the separation through a mutual information objective that estimates the entropy of the attribute within the latent space, yielding stable training and adjustable removal strength. Tests on IJB-C, RFW, and VGGFace2 demonstrate improved privacy-utility curves compared with prior approaches and measurable drops in recognition bias across demographic groups.

Core claim

VLEED feeds pretrained face embeddings into a variational autoencoder whose training objective combines reconstruction with an entropy-based estimate of mutual information between the latent code and the target categorical variable. The resulting distilled embeddings retain high verification performance on standard benchmarks yet exhibit sharply reduced predictability of the chosen attribute under both linear and nonlinear probes, with explicit control over the leakage level.

What carries the argument

The variational autoencoder whose latent entropy estimation objective controls how much categorical attribute information remains entangled with identity features.

If this is right

  • Verification utility on IJB-C remains competitive across a range of disentanglement strengths.
  • Predictability of the target attribute drops under both linear and nonlinear classifiers.
  • False-match-rate disparities across demographic groups decrease.
  • The transformation works as a lightweight post-processing step on any existing embedding model without retraining it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same entropy-estimation approach could be tested on continuous attributes or on embeddings from modalities other than faces.
  • Production systems might expose the disentanglement strength as a tunable privacy parameter for end users.
  • Repeated application across successive demographic attributes could reveal cumulative effects on downstream fairness metrics.

Load-bearing premise

The variational autoencoder and entropy objective can isolate the target categorical attribute from identity information without creating new instabilities or unintended correlations.

What would settle it

If a nonlinear classifier trained after VLEED still predicts gender or ethnicity from the transformed embeddings at rates comparable to the original embeddings, or if verification equal-error rates rise markedly above the pretrained baseline.

Figures

Figures reproduced from arXiv: 2604.11250 by 2) ((1) Idiap Research Institute, (2) UNIL, Lausanne, Martigny, S\'ebastien Marcel (1, Sushil Bhattacharjee (1), Switzerland, Switzerland), \"Unsal \"Ozt\"urk (1), Vedrana Krivoku\'ca Hahn (1).

Figure 1
Figure 1. Figure 1: Top: standard face recognition (FR) embeddings e e (enrolment) and e p (probe) yield high cosine similarity but leak sensitive attributes (gen￾der, ethnicity). Bottom: a disentanglement step produces privacy-preserving embeddings e e ′ , e p ′ that retain verification utility while reducing attribute leakage. VLEED, proposed in this paper, is one such disentanglement method. Values are illustrative. Disent… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VLEED architecture. The encoder maps the input [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of VLEED pipeline. Feature Extraction: A pretrained face recognition model produces fixed embeddings x. Disentanglement: VLEED (a VAE-based disentanglement module) is trained post-hoc to factorise embeddings into a residual latent zr (identity-relevant, minimises I(C; Zr)) and a class latent zc (demographic, class-conditional prior). Evaluation: zr is released for verification (high TMR) and shows… view at source ↗
Figure 4
Figure 4. Figure 4: Visual analysis of VLEED behaviour and comparison with prior methods. (a) t-SNE visualisation of how disentanglement strength affects the residual [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ROC curves (TMR vs. FMR) for VLEED across [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
read the original abstract

Face recognition embeddings encode identity, but they also encode other factors such as gender and ethnicity. Depending on how these factors are used by a downstream system, separating them from the information needed for verification is important for both privacy and fairness. We propose Variational Latent Entropy Estimation Disentanglement (VLEED), a post-hoc method that transforms pretrained embeddings with a variational autoencoder and encourages a distilled representation where the categorical variable of interest is separated from identity-relevant information. VLEED uses a mutual information-based objective realised through the estimation of the entropy of the categorical attribute in the latent space, and provides stable training with fine-grained control over information removal. We evaluate our method on IJB-C, RFW, and VGGFace2 for gender and ethnicity disentanglement, and compare it to various state-of-the-art methods. We report verification utility, predictability of the disentangled variable under linear and nonlinear classifiers, and group disparity metrics based on false match rates. Our results show that VLEED offers a wide range of privacy-utility tradeoffs over existing methods and can also reduce recognition bias across demographic groups.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes Variational Latent Entropy Estimation Disentanglement (VLEED), a post-hoc variational autoencoder method that transforms pretrained face recognition embeddings to separate categorical attributes (gender, ethnicity) from identity-relevant information via a mutual information objective realized through entropy estimation in the latent space. It claims stable training, fine-grained control over information removal, superior privacy-utility tradeoffs compared to existing methods, and reduced recognition bias across demographic groups, with evaluations on IJB-C, RFW, and VGGFace2 using verification utility, attribute predictability under linear/nonlinear probes, and false match rate disparities.

Significance. If the entropy-based disentanglement reliably achieves the claimed separation without residual leakage or new instabilities, the approach would provide a practical post-hoc tool for privacy-preserving and fairer face recognition systems, addressing key ethical challenges in biometric embeddings.

major comments (3)
  1. [§3.2] §3.2 (mutual information objective): the entropy estimation of the categorical attribute is central to minimizing leakage and generating the reported privacy-utility curves and bias reductions, yet no independent validation, error bounds, ablation on estimator choice (e.g., vs. MINE or InfoNCE), or analysis of bias/variance in the estimator is supplied; if the estimator fails to capture nonlinear dependencies, residual identity leakage could remain undetected by the linear/nonlinear probes.
  2. [Experiments] Experiments (results on IJB-C/RFW/VGGFace2): the central claim of 'wide range of privacy-utility tradeoffs' and bias reduction rests on quantitative comparisons, but the manuscript supplies no numerical values, error bars, statistical significance tests, or ablations on the VAE regularization weight (beta) and latent dimension, leaving the superiority over baselines unverifiable.
  3. [§4.1] §4.1 (evaluation of isolation): the assumption that the VAE isolates the categorical variable without introducing unintended correlations is load-bearing for the fairness claims, but is tested only via attribute predictability; no additional diagnostics (e.g., correlation with other factors or reconstruction of identity from the 'disentangled' latent) are reported.
minor comments (2)
  1. [Abstract] Abstract: 'various state-of-the-art methods' is vague; explicitly list the baselines compared (e.g., in a table) for reproducibility.
  2. [Notation] Notation: the precise form of the variational loss and entropy estimator (sampling-based, neural, etc.) should be written out with equation numbers for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive referee report. We address each major comment below and will incorporate revisions to strengthen the validation of the entropy estimator, improve experimental reporting, and add further diagnostics for disentanglement isolation.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (mutual information objective): the entropy estimation of the categorical attribute is central to minimizing leakage and generating the reported privacy-utility curves and bias reductions, yet no independent validation, error bounds, ablation on estimator choice (e.g., vs. MINE or InfoNCE), or analysis of bias/variance in the estimator is supplied; if the estimator fails to capture nonlinear dependencies, residual identity leakage could remain undetected by the linear/nonlinear probes.

    Authors: We appreciate the referee's emphasis on rigorous validation of the entropy estimator. Our method realizes the mutual information objective via variational latent entropy estimation within the VAE. While the submission prioritizes end-to-end results, we agree additional checks are warranted. In revision we will add an ablation comparing our estimator against MINE and InfoNCE, report bias/variance analysis on controlled synthetic data, and include error bounds derived from the variational approximation. These will confirm capture of nonlinear dependencies and that residual leakage is not missed by the probes. revision: yes

  2. Referee: [Experiments] Experiments (results on IJB-C/RFW/VGGFace2): the central claim of 'wide range of privacy-utility tradeoffs' and bias reduction rests on quantitative comparisons, but the manuscript supplies no numerical values, error bars, statistical significance tests, or ablations on the VAE regularization weight (beta) and latent dimension, leaving the superiority over baselines unverifiable.

    Authors: We acknowledge that the manuscript relies primarily on figures for the privacy-utility curves and bias metrics without accompanying numerical tables, error bars, or statistical tests. We will revise to include detailed numerical results in tables, error bars from multiple random seeds, statistical significance testing (e.g., paired t-tests against baselines), and ablations varying the beta regularization weight and latent dimension. These changes will render the claimed wide tradeoffs and superiority verifiable. revision: yes

  3. Referee: [§4.1] §4.1 (evaluation of isolation): the assumption that the VAE isolates the categorical variable without introducing unintended correlations is load-bearing for the fairness claims, but is tested only via attribute predictability; no additional diagnostics (e.g., correlation with other factors or reconstruction of identity from the 'disentangled' latent) are reported.

    Authors: The referee correctly identifies that isolation is currently assessed via linear and nonlinear attribute predictability. To more thoroughly support the assumption of no unintended correlations and minimal identity leakage, the revision will add correlation analyses with additional demographic factors and experiments attempting identity reconstruction from the disentangled latent. These diagnostics will provide stronger evidence for the fairness claims. revision: yes

Circularity Check

0 steps flagged

No circularity in VLEED variational objective or evaluations

full rationale

The paper describes VLEED as a post-hoc VAE transformation of pretrained embeddings with a mutual-information objective implemented via entropy estimation of categorical attributes (gender/ethnicity) in the latent space. This is a standard variational formulation with no equations or claims that reduce by construction to fitted inputs, self-citations, or renamed known results. Reported privacy-utility tradeoffs, predictability under linear/nonlinear probes, and bias metrics are obtained from empirical evaluation on IJB-C, RFW, and VGGFace2 rather than forced by the method's own definitions. The derivation chain is self-contained against external benchmarks and does not invoke load-bearing self-citations or uniqueness theorems from prior author work.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that a variational latent space can be shaped to separate identity from a single categorical attribute via entropy minimization; standard VAE regularization weights and latent dimensionality are likely free parameters tuned on the target datasets.

free parameters (1)
  • VAE regularization weight (beta) and latent dimension
    Typical hyperparameters in variational autoencoders that control the disentanglement objective and must be chosen or fitted for each attribute and dataset.
axioms (1)
  • domain assumption The entropy of the categorical attribute estimated in the latent space is a faithful proxy for mutual information with the original embedding.
    Central to the mutual-information-based objective described in the abstract.

pith-pipeline@v0.9.0 · 5552 in / 1256 out tokens · 58534 ms · 2026-05-10T14:56:07.782968+00:00 · methodology

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

33 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    On soft-biometric information stored in biometric face embeddings,

    P. Terh ¨orst, D. F ¨ahrmann, N. Damer, F. Kirchbuchner, and A. Kuijper, “On soft-biometric information stored in biometric face embeddings,” IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 3, no. 4, pp. 519–534, 2021. 13

    work page 2021

  2. [2]

    Beyond identity: What information is stored in biometric face templates?

    P. Terh ¨orst, D. F ¨ahrmann, N. Damer, F. Kirchbuchner, and A. Kuijper, “Beyond identity: What information is stored in biometric face templates?” in2020 IEEE International Joint Conference on Biometrics (IJCB). IEEE Press, 2020, p. 1–10. [Online]. Available: https: //doi.org/10.1109/IJCB48548.2020.9304874

    work page doi:10.1109/ijcb48548.2020.9304874 2020

  3. [3]

    An attack on facial soft-biometric privacy enhancement,

    D. Osorio-Roig, C. Rathgeb, P. Drozdowski, P. Terh ¨orst, V . ˇStruc, and C. Busch, “An attack on facial soft-biometric privacy enhancement,” IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 4, no. 2, pp. 263–275, 2022

    work page 2022

  4. [4]

    Jointly de-biasing face recognition and demographic attribute estimation,

    S. Gong, X. Liu, and A. K. Jain, “Jointly de-biasing face recognition and demographic attribute estimation,” inComputer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 330–347

    work page 2020

  5. [5]

    Emerging Properties in Self-Supervised Vision Transformers , booktitle =

    P. Dhar, J. Gleason, A. Roy, C. D. Castillo, and R. Chellappa, “PASS: Protected Attribute Suppression System for Mitigating Bias in Face Recognition,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV). Los Alamitos, CA, USA: IEEE Computer Society, Oct. 2021, pp. 15 067–15 076. [Online]. Available: https://doi.ieeecomputersociety.org/10.11...

    work page doi:10.1109/iccv48922.2021.01481 2021

  6. [6]

    Suppressing gender and age in face templates using incremental variable elimination,

    P. Terh ¨orst, N. Damer, F. Kirchbuchner, and A. Kuijper, “Suppressing gender and age in face templates using incremental variable elimination,” in2019 International Conference on Biometrics (ICB), 2019, pp. 1–8

    work page 2019

  7. [7]

    Raman, Demetri Terzopoulos, and Kyunghyun Sung

    P. Melzi, H. O. Shahreza, C. Rathgeb, R. Tolosana, R. Vera- Rodriguez, J. Fierrez, S. Marcel, and C. Busch, “Multi-IVE: Privacy Enhancement of Multiple Soft-Biometrics in Face Embeddings,” in 2023 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW). Los Alamitos, CA, USA: IEEE Computer Society, Jan. 2023, pp. 323–331. [Online]....

    work page doi:10.1109/w 2023

  8. [8]

    Learning privacy-enhancing face representations through feature disentanglement,

    B. Bortolato, M. Ivanovska, P. Rot, J. Kri ˇzaj, P. Terh ¨orst, N. Damer, P. Peer, and V . ˇStruc, “Learning privacy-enhancing face representations through feature disentanglement,” in2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020). IEEE Press, 2020, p. 495–502. [Online]. Available: https: //doi.org/10.1109/FG47...

    work page doi:10.1109/fg47880.2020.00007 2020

  9. [9]

    Aspecd: Adaptable soft- biometric privacy-enhancement using centroid decoding for face ver- ification,

    P. Rot, P. Terh ¨orst, P. Peer, and V . ˇStruc, “Aspecd: Adaptable soft- biometric privacy-enhancement using centroid decoding for face ver- ification,” in2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), 2024, pp. 1–11

    work page 2024

  10. [10]

    Slerpface: face template protection via spherical linear interpolation,

    Z. Zhong, Y . Mi, Y . Huang, J. Xu, G. Mu, S. Ding, J. Zhang, R. Guo, Y . Wu, and S. Zhou, “Slerpface: face template protection via spherical linear interpolation,” inProceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educ...

    work page doi:10.1609/aaai.v39i10.33162 2025

  11. [11]

    Privacy-preserving Adversarial Facial Features,

    Z. Wang, H. Wang, S. Jin, W. Zhang, J. Hut, Y . Wang, P. Sun, W. Yuan, K. Liu, and K. Rent, “Privacy-preserving Adversarial Facial Features,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Los Alamitos, CA, USA: IEEE Computer Society, Jun. 2023, pp. 8212–8221

    work page 2023

  12. [12]

    An overview of privacy-enhancing technologies in biometric recognition,

    P. Melzi, C. Rathgeb, R. Tolosana, R. Vera-Rodriguez, and C. Busch, “An overview of privacy-enhancing technologies in biometric recognition,”ACM Comput. Surv., vol. 56, no. 12, Oct. 2024. [Online]. Available: https://doi.org/10.1145/3664596

    work page doi:10.1145/3664596 2024

  13. [13]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”

    work page

  14. [14]

    Auto-Encoding Variational Bayes

    [Online]. Available: https://arxiv.org/abs/1312.6114

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    beta-V AE: Learning basic visual concepts with a constrained variational framework,

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-V AE: Learning basic visual concepts with a constrained variational framework,” inInternational Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/forum?id=Sy2fzU9gl

    work page 2017

  16. [16]

    Isolating sources of disentanglement in vaes,

    R. T. Q. Chen, X. Li, R. Grosse, and D. Duvenaud, “Isolating sources of disentanglement in vaes,” inProceedings of the 32nd International Conference on Neural Information Processing Systems, ser. NIPS’18. Red Hook, NY , USA: Curran Associates Inc., 2018, p. 2615–2625

    work page 2018

  17. [17]

    Disentangling by factorising,

    H. Kim and A. Mnih, “Disentangling by factorising,” inProceedings of the 35th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. PMLR, 10–15 Jul 2018, pp. 2649–2658. [Online]. Available: https://proceedings.mlr.press/v80/kim18b.html

    work page 2018

  18. [18]

    Disentangling factors of variation in deep representation using adversarial training,

    M. F. Mathieu, J. J. Zhao, J. Zhao, A. Ramesh, P. Sprechmann, and Y . LeCun, “Disentangling factors of variation in deep representation using adversarial training,” inAdvances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc.,

    work page

  19. [19]

    Available: https://proceedings.neurips.cc/paper files/ paper/2016/file/ef0917ea498b1665ad6c701057155abe-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2016/file/ef0917ea498b1665ad6c701057155abe-Paper.pdf

    work page 2016

  20. [20]

    Flexibly fair representation learning by disentanglement,

    E. Creager, D. Madras, J.-H. Jacobsen, M. Weis, K. Swersky, T. Pitassi, and R. Zemel, “Flexibly fair representation learning by disentanglement,” inProceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR, 09–15 Jun 2019, pp. 1436–1445. [Onli...

    work page 2019

  21. [21]

    Locatello, G

    F. Locatello, G. Abbati, T. Rainforth, S. Bauer, B. Sch ¨olkopf, and O. Bachem,On the fairness of disentangled representations. Red Hook, NY , USA: Curran Associates Inc., 2019

    work page 2019

  22. [22]

    Mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y . Bengio, A. Courville, and D. Hjelm, “Mutual information neural estimation,” inProceedings of the 35th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. PMLR, 10–15 Jul 2018, pp. 531–540. [Online]. Available: https://procee...

    work page 2018

  23. [23]

    CLUB: A contrastive log-ratio upper bound of mutual information,

    P. Cheng, W. Hao, S. Dai, J. Liu, Z. Gan, and L. Carin, “CLUB: A contrastive log-ratio upper bound of mutual information,” inProceedings of the 37th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, H. D. III and A. Singh, Eds., vol. 119. PMLR, 13–18 Jul 2020, pp. 1779–1788. [Online]. Available: https://proceedin...

    work page 2020

  24. [24]

    Face-CPFNet: Leveraging Disentangled Representations for Dual-Level Soft-Biometric Privacy- Enhancement,

    Z. Chen, Z. Yao, B. Jin, J. Ning, and M. Lin, “Face-CPFNet: Leveraging Disentangled Representations for Dual-Level Soft-Biometric Privacy- Enhancement,”IEEE Transactions on Dependable and Secure Comput- ing, vol. 22, no. 06, pp. 7060–7076, Nov. 2025. [Online]. Available: https://doi.ieeecomputersociety.org/10.1109/TDSC.2025.3594681

    work page doi:10.1109/tdsc.2025.3594681 2025

  25. [25]

    Privacy preservation in face soft biometrics via attribute disentanglement,

    Y . Wang, B. Jin, Z. Chen, J. Lin, and Z. Yao, “Privacy preservation in face soft biometrics via attribute disentanglement,”Expert Systems with Applications, vol. 312, p. 131520, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0957417426004331

    work page 2026

  26. [26]

    SensitiveNets: Learning Agnostic Representations with Application to Face Images,

    A. Morales, J. Fierrez, R. Vera-Rodriguez, and R. Tolosana, “SensitiveNets: Learning Agnostic Representations with Application to Face Images,”IEEE Transactions on Pattern Analysis & Machine Intel- ligence, vol. 43, no. 06, pp. 2158–2164, Jun. 2021. [Online]. Available: https://doi.ieeecomputersociety.org/10.1109/TPAMI.2020.3015420

    work page doi:10.1109/tpami.2020.3015420 2021

  27. [27]

    Null it out: Guarding protected attributes by iterative nullspace projection,

    S. Ravfogel, Y . Elazar, H. Gonen, M. Twiton, and Y . Goldberg, “Null it out: Guarding protected attributes by iterative nullspace projection,” inProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds. Online: Association for Computational Linguistics, Jul. 2020, pp...

    work page 2020

  28. [28]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4685–4694

    work page 2019

  29. [29]

    In: 2018 IEEE Symposium on Security and Privacy (SP)

    Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman, “VGGFace2: A Dataset for Recognising Faces across Pose and Age,” in2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018). Los Alamitos, CA, USA: IEEE Computer Society, May 2018, pp. 67–74. [Online]. Available: https://doi.ieeecomputersociety.org/10.1109/FG.2018.00020

    work page doi:10.1109/fg.2018.00020 2018

  30. [30]

    Iarpa janus benchmark - c: Face dataset and protocol,

    B. Maze, J. Adams, J. A. Duncan, N. Kalka, T. Miller, C. Otto, A. K. Jain, W. T. Niggel, J. Anderson, J. Cheney, and P. Grother, “Iarpa janus benchmark - c: Face dataset and protocol,” in2018 International Conference on Biometrics (ICB), 2018, pp. 158–165

    work page 2018

  31. [31]

    Racial faces in the wild: Reducing racial bias by information maximization adaptation network,

    M. Wang, W. Deng, J. Hu, X. Tao, and Y . Huang, “Racial faces in the wild: Reducing racial bias by information maximization adaptation network,” in2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 692–702

    work page 2019

  32. [32]

    [Online]

    ISO/IEC, “ISO/IEC 19795-10:2024 — Information technology — Biometric performance testing and reporting — Part 10: Quantifying biometric system performance variation across demographic groups,” 2024, international Organization for Standardization, Geneva, Switzer- land. [Online]. Available: https://www.iso.org/standard/81223.html

    work page 2024

  33. [33]

    Evaluating proposed fairness models for face recogni- tion algorithms,

    J. J. Howard, E. J. Laird, R. E. Rubin, Y . B. Sirotin, J. L. Tipton, and A. R. Vemury, “Evaluating proposed fairness models for face recogni- tion algorithms,” inPattern Recognition, Computer Vision, and Image Processing. ICPR 2022 International Workshops and Challenges, J.-J. Rousseau and B. Kapralos, Eds. Cham: Springer Nature Switzerland, 2023, pp. 431–447

    work page 2022