Pith. sign in

REVIEW 5 major objections 7 minor 57 references

A Responsible Face Recognition Approach for Small and Mid-Scale Systems Through Personalized Neural Networks

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Replacing face templates with tiny per-person networks keeps accuracy and blocks gender inference.

desk verdict MOTE's per-identity classifier idea is genuinely worth a look, but the headline privacy claim rests on a broken attack adaptation and needs major rework before it convinces. read the letter →

arxiv 2505.19920 v1 pith:6NWXL5AB submitted 2025-05-26 cs.CV cs.AI

Sebastian Gro{ss} , Stefan Heindorf , Philipp Terhörst This is my paper · ORCID
classification cs.CVcs.AI
keywords facerecognitionmodeltemplatesprivacy-enhancingsoft-biometricprivacygenderfairnesskerneldensityestimationpersonalizedneuralnetworkssynthetictemplategeneration
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 proposes MOTE, a face recognition design in which each enrolled identity is stored not as a fixed vector template but as a small binary neural network trained to recognize that one person. The authors claim that this change preserves verification accuracy comparable to ArcFace and MagFace while making gender-inference attacks fail at random-guessing level and improving gender fairness. The practical motivation is that small- and mid-scale systems can now choose a representation that is harder to exploit for soft-biometric leakage, with a per-identity knob for fairness and easier explainability.

What carries the argument

The central object is the model-template (MOTE): a small feed-forward binary classifier (512->128->64->1 with ReLU and dropout) trained per identity with balanced genuine and imposter samples. The enabling mechanism is the template generation step: embeddings are centered by subtracting the identity centroid, a Gaussian KDE is fit separately to the normalized embeddings of male and female subjects, and synthetic templates are drawn from the KDE and added back to the centroid, allowing one reference image to seed a full training set.

What would settle it

Train MOTE on a single enrollment image per identity and test on a probe set with large pose or lighting variation (e.g., ColorFeret's pose subsets); if FNMR at $10^{-3}$ FMR rises far above the reported about 0.04 for ArcFace, the KDE synthesis is not capturing true intra-identity variation. Alternatively, train a supervised gender classifier on MOTE's per-identity output scores and check whether balanced accuracy stays near 50%.

Watch

Extended reading notes

Core claim

The central claim is that a personalized classifier per identity, trained from a single reference embedding plus synthetic samples generated from gender-specific kernel density estimates, can replace a stored template without a meaningful loss in verification performance, while eliminating the attribute leakage that vector templates carry. MOTE's decision boundary is not calibrated across identities, so an adversary comparing classifier outputs for different templates collapses to one class and the gender inference attack falls to about 50% balanced accuracy. The paper further reports that MOTE matches or exceeds the gender fairness of the underlying embedding model on FDR and iGARBE metrics, and that per-identity fairness can be tuned by choosing the gender balance of the synthetic training set.

Load-bearing premise

The method assumes that synthetic samples drawn from a global gender-specific noise distribution around a single reference embedding faithfully represent how that person's face varies in pose, lighting, and expression; if that assumption fails, the per-identity classifier overfits to the reference image and verification accuracy on new probes collapses.

Editorial extensions

If this is right

  • Verification accuracy drops by only a small margin relative to the underlying embedding model: on LFW with ArcFace, FNMR at $10^{-3}$ FMR stays at 0.0260 versus 0.0256 for the traditional system.
  • Gender-inference attacks that achieve 79.6 to 93.3% balanced accuracy on traditional templates and existing privacy-enhancement methods fall to 50.0 to 50.3% with MOTE.
  • Fairness metrics (FDR and iGARBE) are higher for MOTE than for the traditional baseline across most datasets and balancing factors, with the largest gain on LFW with MagFace where iGARBE rises from 0.692 to 0.997.
  • The per-identity nature of the classifier allows fairness to be adjusted at enrollment time by varying the male-to-female ratio of generated samples.
  • Explanations can be produced with standard classifier visualization tools such as Grad-CAM++, which is difficult for similarity-based embedding systems.

Reading between the lines

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

  • If MOTE's privacy claim holds against stronger adversaries, the same design could be applied to other soft-biometric attributes: training the KDE per attribute (age, ethnicity) and balancing generated samples per identity could give a general mechanism for attribute obfuscation.
  • The method's accuracy depends on how well a global gender-specific KDE captures each individual's within-identity variation; a targeted experiment on pose- or lighting-diverse probe sets with a single enrollment image would reveal whether the synthetic samples generalize or merely memorize the reference.
  • MOTE could extend beyond faces to other biometric signals (iris, voice, gait) where template leakage is a concern, provided a suitable embedding space and attribute-labeled background data exist.
  • The lack of inter-identity calibration that blocks inference attacks may also make score-based fusion across identities or confidence thresholds less straightforward, a design trade-off worth exploring.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. This paper proposes MOTE, a face verification scheme that replaces vector templates with per-identity binary MLP classifiers. For each enrolled identity, one classifier is trained on the single reference embedding plus KDE-generated synthetic embeddings, with a 'balancing factor' that controls the male/female mix of the synthetic training set. The authors evaluate on Adience, ColorFeret, and LFW using ArcFace and MagFace embeddings, reporting verification accuracy close to the template baselines, improved fairness (FDR, iGARBE), gender-inference attack success around 50%, qualitative Grad-CAM++ explanations, and increased storage and enrollment costs.

Significance. If the privacy result were valid, the paper would offer a qualitatively different privacy mechanism: rather than removing soft-biometric information from a template, the stored object is a decision boundary that, as claimed, outputs uncalibrated scores that cannot be fused across identities for attribute inference. The per-identity fairness knob is also practically interesting for small enrollment sets. The paper is commendable for evaluating on three datasets, two embedding backbones, multiple fairness metrics, and for reporting storage and enrollment overhead. However, the headline privacy claim is currently supported only by an undefined adaptation of the Osorio-Roig attack, and the fairness gains are partly induced by the experimental protocol rather than by an intrinsic property of the method.

major comments (5)
  1. [5.2 / Table 3] The privacy claim rests on an attack adaptation that is not specified and, by the authors' own description, cannot support the conclusion. The Osorio-Roig attack is defined for vector templates compared by a similarity function; applying it to MOTE's sigmoid outputs requires a rule for aggregating per-identity classifier scores across the attack set, but no such rule is given. Section 5.2 states that 'all gender predictions collapse to a single class with unalgined comparison scores' and that 'personalized models are not calibrated among each other.' A balanced accuracy near 50% under these conditions indicates that the score aggregation is broken, not that gender information is absent. The adversary model must be specified (white-box access to stored models, query access, or only black-box decisions), and the attack should be evaluated on calibrated scores, on internal activations, or with a supervised classifier trained on MOTE outputs. Without this, Table 3 does not establish gender obfuscation.
  2. [5.3 / Table 4] The fairness comparison is partly by construction. The balancing factor directly sets the male/female ratio of synthetic training samples, so reporting FDR/iGARBE for MOTE at a chosen factor demonstrates that the designer can tune fairness, not that MOTE is intrinsically fairer. The traditional ArcFace and MagFace baselines are fixed systems that were not given the same gender-balancing intervention, making the comparison asymmetric. The paper should either compare MOTE against baselines retrained on equivalently gender-balanced data, or report accuracy and fairness jointly for every balancing factor (the accuracy corresponding to each fairness value is missing), so the cost of the fairness knob is transparent.
  3. [3.2, Eq. (4)] Equation (4) adds the identity centroid to samples drawn from a gender-specific KDE of normalized embeddings from unrelated identities. This assumes that the within-identity residual distribution of the enrolled subject is identical to the global gender-specific residual distribution. With only one reference sample per identity, there is no evidence that this assumption holds, and the personalized classifier may overfit to the single reference instead of learning identity-preserving variation. The manuscript provides no ablation on the number of synthetic samples, no comparison of synthetic versus real embedding distributions, and no analysis of verification performance as a function of synthetic sample diversity. This missing validation directly affects the accuracy and generalization claims in Table 2.
  4. [Table 1 / Section 5.1] The recognition accuracy numbers are not sufficiently defined. Table 1 reports 'Recognition Accuracy' without stating the pair construction, the threshold selection rule, or the operating point; values such as 99.956% on Adience are hard to reconcile with the FNMR of 0.0558 at FMR=10^-3 in Table 2 for the same configuration. If accuracy is computed at a different threshold, that threshold and its selection protocol must be given, and confidence intervals over subjects or splits should be reported. As written, the 'comparable accuracy' claim rests on an ambiguous metric.
  5. [5.4 / Figure 4] Grad-CAM++ is applied to a classifier that operates on 512-dimensional embeddings, yet Figure 4 displays heatmaps over facial images. The paper does not explain how importance weights over embedding dimensions are mapped back to image pixels. Without this mapping, the explainability analysis is not reproducible and does not support the claim that the MOTE models 'focus on crucial facial regions.' Please specify the visualization protocol or replace the qualitative figure with a quantitative explanation evaluation.
minor comments (7)
  1. [Throughout] There are numerous typos and grammatical errors, including 'unalgined', 'amoung', 'presentes', 'Approachs', 'Anaylsis', and 'Accurracy'; these should be corrected.
  2. [Section 2.1] The paper mentions the Osorio-Roig attack but does not describe the attack setting (e.g., which embeddings, which similarity, which datasets) when introducing it; please state this context in the related work.
  3. [Section 3.2] The KDE bandwidth is determined by cross-validation, but the cross-validation procedure (data splits, metric, and reference implementation) is not described; without this, the synthetic sample generation is not reproducible.
  4. [Section 4.1] The paper says subject-disjoint training and testing sets were used; please clarify whether the gender-specific KDEs are also trained only on the enrollment/training identities, since the KDEs are computed from embeddings of unrelated identities.
  5. [Table 5] The storage value of 39.2 KB for a 512-dimensional float template is not explained; a 512-d float32 vector is 2 KB, so the source of the additional overhead should be clarified.
  6. [Section 5.5] Inference time of 0.31 ms per embedding is reported without details on batch size, hardware state, or measurement methodology; please add these details.
  7. [Section 5.3] The claim of up to 30.5% improvement in iGARBE should be accompanied by the corresponding accuracy at that balancing factor, since fairness improvements at a large accuracy cost are less meaningful.

Circularity Check

2 steps flagged · score 6.0 of 10

Central privacy claim reduces to uncalibrated-output artifact; fairness is a tunable knob rather than a prediction.

  1. self definitional [Section 5.2 (Privacy Protection), Table 3]
    "However, our analysis revealed that MOTE achieves robust privacy protection by creating a decision boundary that effectively neutralizes gender inference attacks, reducing them to random guessing performance as personalized models are not calibrated amoung each other. ... This privacy protection stems from a fundamental difference in output characteristics: when attacking MOTE on a given dataset, all gender predictions collapse to a single class with unalgined comparison scores, effectively destroying any meaningful inference capability."

    The paper's evidence for privacy is the Osorio-Roig attack's near-50% balanced accuracy in Table 3. The paper itself explains that this outcome occurs because per-identity classifiers are trained independently and are not calibrated, making their raw sigmoid outputs incommensurable; the attack's score-averaging recipe then collapses all predictions to one class. A 50% balanced accuracy is therefore a mechanical consequence of applying a score-comparison attack to an output space that is by construction unaligned, not evidence that gender information has been removed.

  2. fitted input called prediction [Section 5.3 (Fairness Evaluation), Table 4]
    "We further analyzed how gender distribution in training data affects the system's fairness by systematically varying the ratio of male to female embeddings in the training set. We selected specific balancing factors (0.0, 0.4, 0.5, 0.6, and 1.0) ... A balancing factor of 0.0 has the consequence that only female templates are sampled, while a balancing factor of 1.0 has the consequence of only sampling male templates."

    The fairness metrics (FDR, iGARBE) are reported after choosing a balancing factor that directly determines the gender composition of the synthetic training data for each identity classifier. The paper claims that 'MOTE consistently outperforms the traditional methods in terms of gender fairness,' but the reported outcomes are largely forced by the selected input distribution; the balancing factor is a design knob, not an independently discovered property. This is a transparent tuning procedure rather than a prediction from first principles, but it still means the headline fairness improvement is partly by construction of the experimental setup.

full rationale

The recognition performance evaluation is self-contained and benchmarked against standard models, so no circularity is present in the accuracy results. The fairness results are transparently controlled by a balancing factor, and although they are reported as improvements, they are tunable outcomes rather than derived predictions; this is a methodological concern more than a hidden circularity. The central privacy claim, however, is not supported by evidence that gender information is absent. The paper's own text attributes the near-random attack success to the lack of calibration among independent per-identity classifiers, causing all predictions to collapse to a single class under the Osorio-Roig score-comparison attack. That makes the privacy result equivalent to the architecture's unaligned output space by construction, rather than a demonstrated neutralization of gender information. The self-citations to prior work by the same author (e.g., Osorio-Roig et al. [9]) are published external benchmarks and do not themselves form a circular argument, but they do not rescue the privacy claim because the attack's failure is an artifact of uncalibrated outputs. Overall, the paper's strongest contribution—privacy protection—reduces by construction, warranting a score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities beyond the model-template concept, which is a renamed per-identity classifier. The main free parameter is the balancing factor that controls fairness by construction. The key hidden assumption is that the KDE-based synthetic samples generalize to unseen within-identity variations, which is not independently validated.

free parameters (3)
  • Balancing factor for synthetic gender distribution = Varied (0.0, 0.4, 0.5, 0.6, 1.0)
    The ratio of male to female synthetic samples is a design choice that directly controls the fairness outcome. It is not fitted to a validation metric but varies per experiment, and the paper reports results across this range.
  • Decision threshold τ = Not specified numerically, optimized to balance FMR and FNMR
    Section 3.4 says τ is optimized to balance false matches and false non-matches. This is a fitted parameter that affects reported accuracy and fairness numbers.
  • KDE bandwidth h = Determined by cross-validation
    Section 3.2 states the bandwidth is determined through cross-validation, so it is fitted to the data and influences the synthetic sample distribution.
assumptions (4)
  • domain assumption KDE-based synthetic samples preserve identity while adding realistic within-identity variation
    Equation (4) generates new templates by adding the identity centroid to noise drawn from a global per-gender KDE. This assumes the global gender-specific variation is a good proxy for the variation of any single individual, which is not established in the paper.
  • domain assumption Gender is the only soft-biometric attribute relevant for fairness in this evaluation
    The paper only considers gender fairness and gender inference privacy. It does not address age, ethnicity, or other attributes, so the reported improvements are limited to gender.
  • domain assumption The Osorio-Roig inference attack is the most successful privacy attack for face templates
    Section 4.2 states this attack is used because it is the most successful attack the authors are aware of, but no comparison to other recent attacks is provided.
  • standard math Standard mathematical and algorithmic background (KDE, neural network training, threshold optimization)
    The paper relies on standard kernel density estimation, backpropagation, and binary classification, which are accepted within the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Responsible Face Recognition Approach for Small and Mid-Scale Systems Through Personalized Neural Networks." pith.science (2026). https://pith.science/paper/6NWXL5AB

@misc{pith2026250519920,
  author       = {Pith},
  title        = {Pith review of: A Responsible Face Recognition Approach for Small and Mid-Scale Systems Through Personalized Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NWXL5AB}},
  note         = {Machine review of arXiv:2505.19920}
}
read the original abstract

Traditional face recognition systems rely on extracting fixed face representations, known as templates, to store and verify identities. These representations are typically generated by neural networks that often lack explainability and raise concerns regarding fairness and privacy. In this work, we propose a novel model-template (MOTE) approach that replaces vector-based face templates with small personalized neural networks. This design enables more responsible face recognition for small and medium-scale systems. During enrollment, MOTE creates a dedicated binary classifier for each identity, trained to determine whether an input face matches the enrolled identity. Each classifier is trained using only a single reference sample, along with synthetically balanced samples to allow adjusting fairness at the level of a single individual during enrollment. Extensive experiments across multiple datasets and recognition systems demonstrate substantial improvements in fairness and particularly in privacy. Although the method increases inference time and storage requirements, it presents a strong solution for small- and mid-scale applications where fairness and privacy are critical.

Figures

Figures reproduced from arXiv: 2505.19920 by the authors.

Figure 2
Figure 2. ), and separate KDE models are trained for each attribute (Step 4), ˆf(x) = 1 nh X i K  x − x˜i h  (3) where K is a Gaussian kernel and h is the bandwidth pa￾rameter determined through cross-validation. Finally, as depicted in Step 5 of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Analysing Verification Performance in Terms Of ROC Curves - The FNMR is reported over various FMR, comparing the proposed MOTE approach against traditional (Trd.) face recognition methods (either ArcFace or MagFace) across multiple datasets. Over all dataset and face recognition systems, a weaker but comparable performance is observable. while both the traditional models and existing privacy￾enhancing techniques (PF… view at source ↗
Figure 4
Figure 4. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 33 canonical work pages

  1. [1]

    S. Z. Li and A. K. Jain, Eds., Handbook of Face Recognition, 2nd Edition. Springer, 2011, ISBN : 978- 0-85729-931-4. DOI: 10.1007/978-0-85729- 932-1

    doi:10.1007/978-0-85729- 2011

  2. [2]

    Facenet: A unified embedding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding for face recognition and clustering,” inIEEE Conference on Computer Vi- sion and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, IEEE Computer Society, 2015, pp. 815–823. DOI: 10.1109/CVPR.2015. 7298682

    doi:10.1109/cvpr.2015 2015

  3. [3]

    Arcface: Addi- tive angular margin loss for deep face recognition,

    J. Deng, J. Guo, and S. Zafeiriou, “Arcface: Addi- tive angular margin loss for deep face recognition,” CoRR, vol. abs/1801.07698, 2018. arXiv: 1801 . 07698

    arXiv 2018

  4. [4]

    Face transformer for recog- nition,

    Y . Zhong and W. Deng, “Face transformer for recog- nition,” CoRR, vol. abs/2103.14803, 2021. arXiv: 2103.14803

    arXiv 2021

  5. [5]

    Part-based face recognition with vision transformers,

    Z. Sun and G. Tzimiropoulos, “Part-based face recognition with vision transformers,” in33rd British Machine Vision Conference 2022, BMVC 2022, Lon- don, UK, November 21-24, 2022 , BMV A Press, 2022, p. 611

    work page 2022

  6. [6]

    On soft-biometric information stored in biometric face embeddings,

    P. Terh ¨orst, D. F ¨ahrmann, N. Damer, F. Kirchbuch- ner, and A. Kuijper, “On soft-biometric information stored in biometric face embeddings,” IEEE Trans. Biom. Behav. Identity Sci., vol. 3, no. 4, pp. 519–534,

    work page

  7. [7]

    Beyond identity: What informa- tion is stored in biometric face templates?

    P. Terh ¨orst, D. F ¨ahrmann, N. Damer, F. Kirchbuch- ner, and A. Kuijper, “Beyond identity: What informa- tion is stored in biometric face templates?” In 2020 IEEE International Joint Conference on Biometrics, IJCB 2020, Houston, TX, USA, September 28 - Octo- ber 1, 2020, IEEE, 2020, pp. 1–10. DOI: 10.1109/ IJCB48548.2020.9304874

    arXiv 2020

  8. [8]

    What else does your biometric data reveal? A survey on soft biomet- rics,

    A. Dantcheva, P. Elia, and A. Ross, “What else does your biometric data reveal? A survey on soft biomet- rics,”IEEE Trans. Inf. Forensics Secur., vol. 11, no. 3, pp. 441–467, 2016. DOI: 10.1109/TIFS.2015. 2480381

    work page doi:10.1109/tifs.2015 2016

Show all 57 references
  1. [9]

    An attack on facial soft- biometric privacy enhancement,

    D. O. Roig, C. Rathgeb, P. Drozdowski, P. Terh ¨orst, V . Struc, and C. Busch, “An attack on facial soft- biometric privacy enhancement,” IEEE Trans. Biom. Behav. Identity Sci., vol. 4, no. 2, pp. 263–275, 2022. DOI: 10.1109/TBIOM.2022.3172724

    2022

  2. [10]

    Privacy-enhancing face biometrics: A comprehensive survey,

    B. Meden et al., “Privacy-enhancing face biometrics: A comprehensive survey,” IEEE Trans. Inf. Foren- sics Secur., vol. 16, pp. 4147–4183, 2021. DOI: 10. 1109/TIFS.2021.3096024

    2021

  3. [11]

    Face recog- nition vendor test part 3: Demographic effects,

    P. Grother, M. Ngan, and K. Hanaoka, “Face recog- nition vendor test part 3: Demographic effects,”NIST Interagency/Internal Report (NISTIR) - 8280), 2019

    2019

  4. [12]

    A Comprehensive Study on Face Recognition Biases Beyond Demographics,

    P. Terh ¨orst et al., “A Comprehensive Study on Face Recognition Biases Beyond Demographics,” IEEE Transactions on Technology and Society , vol. 3, no. 1, pp. 16–30, Mar. 2022, ISSN : 2637-6415. DOI: 10.1109/TTS.2021.3111823

    2022

  5. [13]

    Evaluating Pro- posed Fairness Models for Face Recognition Al- gorithms,

    J. J. Howard, E. J. Laird, R. E. Rubin, Y . B. Sirotin, J. L. Tipton, and A. R. Vemury, “Evaluating Pro- posed Fairness Models for Face Recognition Al- gorithms,” in Pattern Recognition, Computer Vi- sion, and Image Processing. ICPR 2022 Interna- tional Workshops and Challenge...

    2022

  6. [14]

    Demographic bias in biometrics: A survey on an emerging challenge,

    P. Drozdowski, C. Rathgeb, A. Dantcheva, N. Damer, and C. Busch, “Demographic bias in biometrics: A survey on an emerging challenge,” IEEE Transac- tions on Technology and Society, vol. 1, no. 2, pp. 89– 103, 2020

    2020

  7. [15]

    Deep face recognition: A survey,

    M. Wang and W. Deng, “Deep face recognition: A survey,” Neurocomputing, vol. 429, pp. 215–244,

  8. [16]

    SLICE: stabilized LIME for consistent explanations for image classi- fication,

    R. P. Bora, P. Terh ¨orst, R. N. J. Veldhuis, R. Ra- machandra, and K. B. Raja, “SLICE: stabilized LIME for consistent explanations for image classi- fication,” in IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, ...

    2024

  9. [17]

    1016 / J

    DOI: 10 . 1016 / J . NEUCOM . 2020 . 10 . 081

    2020

  10. [18]

    Ef- ficient explainable face verification based on similar- ity score argument backpropagation,

    M. Huber, A. T. Luu, P. Terh ¨orst, and N. Damer, “Ef- ficient explainable face verification based on similar- ity score argument backpropagation,” in IEEE/CVF Winter Conference on Applications of Computer Vi- sion, WACV 2024, Waikoloa, HI, USA, January 3-8, 2024, IEEE, 2024, ...

    2024

  11. [19]

    Explain- able face recognition,

    J. R. Williford, B. B. May, and J. Byrne, “Explain- able face recognition,” in Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, Au- gust 23-28, 2020, Proceedings, Part XI , A. Vedaldi, H. Bischof, T. Brox, and J. Frahm, Eds., ser. Lecture Notes in Computer ...

    2020 doi

  12. [20]

    A. K. Jain, P. Flynn, and A. A. Ross, Handbook of biometrics. Springer Science & Business Media, 2007

    2007

  13. [21]

    Pixel-level face im- age quality assessment for explainable face recogni- tion,

    P. Terh ¨orst, M. Huber, N. Damer, F. Kirchbuchner, K. B. Raja, and A. Kuijper, “Pixel-level face im- age quality assessment for explainable face recogni- tion,” IEEE Trans. Biom. Behav. Identity Sci., vol. 5, no. 2, pp. 288–297, 2023. DOI: 10.1109/TBIOM. 2023.3263186

    2023

  14. [22]

    Unsupervised privacy-enhancement of face representations using similarity-sensitive noise trans- formations,

    P. Terh ¨orst, N. Damer, F. Kirchbuchner, and A. Kuijper, “Unsupervised privacy-enhancement of face representations using similarity-sensitive noise trans- formations,” Appl. Intell., vol. 49, no. 8, pp. 3043– 3060, 2019. DOI: 10 . 1007 / S10489 - 019 - 01432-5

    2019

  15. [23]

    Magface: A universal representation for face recognition and quality assessment,

    Q. Meng, S. Zhao, Z. Huang, and F. Zhou, “Magface: A universal representation for face recognition and quality assessment,” in IEEE Conference on Com- puter Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021 , Computer Vision Foun- dation / IEEE, 2021, pp. 14...

    2021

  16. [24]

    PE-MIU: A training-free privacy- enhancing face recognition approach based on min- imum information units,

    P. Terh ¨orst et al., “PE-MIU: A training-free privacy- enhancing face recognition approach based on min- imum information units,” IEEE Access , vol. 8, pp. 93 635–93 647, 2020. DOI: 10 . 1109 / ACCESS.2020.2994960

    2020

  17. [25]

    Unsupervised enhancement of soft- biometric privacy with negative face recognition,

    P. Terh ¨orst, M. Huber, N. Damer, F. Kirchbuchner, and A. Kuijper, “Unsupervised enhancement of soft- biometric privacy with negative face recognition,” CoRR, vol. abs/2002.09181, 2020. arXiv: 2002 . 09181

    2002 arXiv

  18. [26]

    Multi-ive: Privacy enhancement of multiple soft-biometrics in face embeddings,

    P. Melzi et al., “Multi-ive: Privacy enhancement of multiple soft-biometrics in face embeddings,” inPro- ceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision (WACV) Workshops , 2023, pp. 323–331

    2023

  19. [27]

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

    P. Terh ¨orst, N. Damer, F. Kirchbuchner, and A. Kui- jper, “Suppressing gender and age in face templates using incremental variable elimination,” in 2019 In- ternational Conference on Biometrics, ICB 2019, Crete, Greece, June 4-7, 2019 , IEEE, 2019, pp. 1–

    2019

  20. [28]

    DOI: 10.1109/ICB45273.2019.8987331

    2019

  21. [29]

    Aspecd: Adaptable soft-biometric privacy-enhancement us- ing centroid decoding for face verification,

    P. Rot, P. Terh ¨orst, P. Peer, and V . ˇStruc, “Aspecd: Adaptable soft-biometric privacy-enhancement us- ing centroid decoding for face verification,” in 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), 2024, pp. 1–11. DOI: 10.1109/FG592...

    2024

  22. [30]

    Sensitivenets: Learning agnostic repre- sentations with application to face images,

    A. Morales, J. Fi ´errez, R. Vera-Rodr ´ıguez, and R. Tolosana, “Sensitivenets: Learning agnostic repre- sentations with application to face images,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 43, no. 6, pp. 2158–2164, 2021. DOI: 10 . 1109 / TPAMI . 2020.3015420

    2021

  23. [31]

    Learning privacy-enhancing face representations through feature disentanglement,

    B. Bortolato et al., “Learning privacy-enhancing face representations through feature disentanglement,” in 15th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2020, Buenos Aires, Argentina, November 16-20, 2020 , IEEE, 2020, pp. 495–502. DOI: 10 . ...

    2020

  24. [32]

    Instance- consistent fair face recognition,

    Y . Li, Y . Sun, Z. Cui, P. Shen, and S. Shan, “Instance- consistent fair face recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1– 17, 2025. DOI: 10 . 1109 / TPAMI . 2025 . 3545781

    2025

  25. [33]

    Fairface: Face attribute dataset for balanced race, gender, and age,

    K. K ¨arkk¨ainen and J. Joo, “Fairface: Face attribute dataset for balanced race, gender, and age,” CoRR, vol. abs/1908.04913, 2019. arXiv: 1908.04913

    1908 arXiv

  26. [34]

    La- bellessface: Fair metric learning for face recogni- tion without attribute labels,

    T. Ohki, Y . Sato, M. Nishigaki, and K. Ito, “La- bellessface: Fair metric learning for face recogni- tion without attribute labels,” in 2024 IEEE Interna- tional Joint Conference on Biometrics (IJCB), 2024, pp. 1–8. DOI: 10 . 1109 / IJCB62174 . 2024 . 10744511

    2024

  27. [35]

    Post-comparison mitigation of de- mographic bias in face recognition using fair score normalization,

    P. Terh ¨orst, J. N. Kolf, N. Damer, F. Kirchbuchner, and A. Kuijper, “Post-comparison mitigation of de- mographic bias in face recognition using fair score normalization,” Pattern Recognit. Lett. , vol. 140, pp. 332–338, 2020. DOI: 10.1016/J.PATREC. 2020.11.007

    2020 doi

  28. [36]

    Demographic fairness transformer for bias mitigation in face recognition,

    K. Kotwal and S. Marcel, “Demographic fairness transformer for bias mitigation in face recognition,” in 2024 IEEE International Joint Conference on Bio- metrics (IJCB) , 2024, pp. 1–10. DOI: 10 . 1109 / IJCB62174.2024.10744457

    2024

  29. [37]

    Mitigating demographic bias in face recognition via regularized score calibra- tion,

    K. Kotwal and S. Marcel, “Mitigating demographic bias in face recognition via regularized score calibra- tion,” in WACV (Workshops), IEEE, 2024, pp. 1150– 1159

    2024

  30. [38]

    Grad-cam: Visual explana- tions from deep networks via gradient-based local- ization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explana- tions from deep networks via gradient-based local- ization,” in IEEE International Conference on Com- puter Vision, ICCV 2017, Venice, Italy, October 22- 29, 2017 , IEEE Comp...

    2017

  31. [39]

    FALCON: fair face recognition via local optimal feature normalization,

    R. Al-Refai, P. Hempel, C. Biagi, and P. Terh ¨orst, “FALCON: fair face recognition via local optimal feature normalization,” in IEEE/CVF Winter Con- ference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, IEEE, 2025, pp. 3416–3426....

    2025

  32. [40]

    On black-box explana- tion for face verification,

    D. Mery and B. Morris, “On black-box explana- tion for face verification,” in IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2022, Waikoloa, HI, USA, January 3-8, 2022, IEEE, 2022, pp. 1194–1203. DOI: 10 . 1109 / WACV51458.2022.00126

    2022

  33. [41]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Con- ference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y . Bengio and Y . LeCun, Eds., 2015

    2015

  34. [42]

    OneCycleLR — PyTorch 2.5 documentation

  35. [43]

    Grad-cam++: Generalized gradient-based visual explanations for deep convolu- tional networks,

    A. Chattopadhyay, A. Sarkar, P. Howlader, and V . N. Balasubramanian, “Grad-cam++: Generalized gradient-based visual explanations for deep convolu- tional networks,” CoRR, vol. abs/1710.11063, 2017. arXiv: 1710.11063

    2017 arXiv

  36. [44]

    Dropout: A sim- ple way to prevent neural networks from overfitting,

    N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A sim- ple way to prevent neural networks from overfitting,” J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958,

    1929

  37. [45]

    Age and gen- der estimation of unfiltered faces,

    E. Eidinger, R. Enbar, and T. Hassner, “Age and gen- der estimation of unfiltered faces,” IEEE Trans. Inf. Forensics Secur., vol. 9, no. 12, pp. 2170–2179, 2014. DOI: 10.1109/TIFS.2014.2359646

    2014

  38. [46]

    The FERET database and evaluation procedure for face-recognition algorithms,

    P. J. Phillips, H. Wechsler, J. Huang, and P. J. Rauss, “The FERET database and evaluation procedure for face-recognition algorithms,” Image Vis. Comput. , vol. 16, no. 5, pp. 295–306, 1998. DOI: 10.1016/ S0262-8856(97)00070-X

    1998

  39. [47]

    Labeled faces in the wild: A database forstudying face recognition in unconstrained envi- ronments,

    G. Huang, M. Mattar, T. Berg, and E. Learned- Miller, “Labeled faces in the wild: A database forstudying face recognition in unconstrained envi- ronments,” Tech. rep., Oct. 2008

    2008

  40. [48]

    L. N. Smith and N. Topin, Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates , May 2018. DOI: 10 . 48550 / arXiv . 1708 . 07120 . arXiv: 1708 . 07120 [cs]

    2018

  41. [49]

    BCEWithLogitsLoss — PyTorch 2.5 documentation

  42. [50]

    Fairness in Bio- metrics: A Figure of Merit to Assess Biometric Veri- fication Systems,

    T. de Freitas Pereira and S. Marcel, “Fairness in Bio- metrics: A Figure of Merit to Assess Biometric Veri- fication Systems,” IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 4, no. 1, pp. 19– 29, Jan. 2022, ISSN : 2637-6407. DOI: 10 . 1109 / TBIOM.2021.3102862

    2022

  43. [51]

    Evaluating proposed fairness models for face recognition 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 recognition algorithms,” in Pattern Recognition, Computer Vision, and Image Processing. ICPR 2022 International Workshops and Challenges - Mont...

    2022

  44. [52]

    On the ”illusion

    P. J. Kurz, H. Wu, K. W. Bowyer, and P. Terh¨orst, “On the ”illusion” of gender bias in face recognition: Explaining the fairness issue through non-demographic attributes,” CoRR, vol. abs/2501.12020, 2025. DOI: 10 . 48550 / ARXIV.2501.12020. arXiv: 2501.12020

    2025 doi arXiv

  45. [53]

    Information technology — Biometric performance testing and reporting — Part 1: Principles and frame- work (ISO/IEC 19795-1)

    “Information technology — Biometric performance testing and reporting — Part 1: Principles and frame- work (ISO/IEC 19795-1).” (2021)

    2021

  46. [54]

    FRONTEX, 2016

    European Agency for the Management of Opera- tional Cooperation at the External Borders of the Member States of the European Union, Guidelines for Processing of Third Country Nationals through Automated Border Control. FRONTEX, 2016. DOI: doi/10.2819/39041

    2016 doi

  47. [626]

    DOI: 10.1109/ICCV.2017.74

    2017 doi

  48. [2014]

    DOI: 10.5555/2627435.2670313

  49. [2021]

    DOI: 10.1109/TBIOM.2021.3093920

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.