Pith. sign in

REVIEW 2 major objections 5 minor 38 references

Quantifying Training Membership Information in the Hyperspherical Embedding Geometry of Face Recognition Models

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A face recognition model's embedding geometry carries measurable membership information, and the amount is controlled mainly by the number of training identities rather than by backbone or loss head.

desk verdict Useful large-scale measurement study, but the same-domain monotonic claim overreaches at n_ids=all. read the letter →

arxiv 2607.15084 v1 pith:NUITA4UA submitted 2026-07-16 cs.CV

Ünsal Öztürk , Sébastien Marcel This is my paper · ORCID
classification cs.CV
keywords membershipinferencefacerecognitionembeddinggeometryhypersphereprivacyauditingangular-marginlossclusterstatisticstraining-setsize
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 asks how much a face recognition model's output embeddings reveal whether a given identity was part of its training data, when only the embeddings are available. It trains 180 models varying backbone size, loss head, training duration, and the number of training identities, and measures four cluster-geometry statistics that separate member from non-member identities. The dominant finding is that the number of training identities controls the separation far more than architecture or loss: on a same-domain held-out reference, the membership signal falls monotonically as more identities are added. Cross-domain non-member sets (different pose, quality, age, ethnicity) inflate the apparent signal, so privacy audits need same-domain references. Combining all four statistics with a learned classifier recovers membership information beyond any single statistic.

What carries the argument

The central object is the per-identity embedding cluster on the unit hypersphere, summarised by four statistics: pairwise cosine similarity (cluster tightness), von Mises-Fisher concentration kappa, a leave-one-out margin-penalised logit approximating the training loss, and a prototype softmax cross-entropy treating other identities as negatives. The argument runs on the assumption that angular-margin losses optimise only training identities, leaving non-member clusters with looser geometry; the statistics convert that geometry into scalar scores, and threshold-classifier AUC quantifies how much membership information survives. The factorial design over 180 models isolates which training-tim

What would settle it

Train a batch of models at 1K and 100K identities on a different large face dataset and check whether the same-domain held-out gap between member and non-member AUC still decreases monotonically; a non-monotonic or inverted trend would refute the claim that identity count is the dominant control. Alternatively, construct a held-out split that matches image count and quality per identity and see if the separation vanishes or changes.

Watch

Extended reading notes

Core claim

The paper establishes that the hyperspherical embedding clusters of training identities remain statistically distinguishable from those of unseen identities, and that the size of this distinguishability is governed chiefly by the number of training identities rather than by backbone or loss head. Using four per-identity statistics—mean pairwise cosine similarity, a von Mises-Fisher concentration estimate, a margin-penalised logit, and a prototype softmax cross-entropy—the authors measure threshold-classifier AUC separating members from non-members across 180 models. On a same-domain held-out split, AUC decreases monotonically as training identities grow from 1K to all available identities, w

Load-bearing premise

The same-domain held-out reference is built by randomly splitting one dataset into member and non-member identities, assuming the random split removes all domain shift; if the split leaves residual differences in image count, quality, or pose, the monotonic decrease and the cross-domain inflation conclusions are confounded.

Editorial extensions

If this is right

  • Privacy auditors should use same-domain held-out non-members rather than cross-domain benchmarks, since cross-domain sets overstate the membership leak.
  • Training on more identities is the only tested design choice that substantially reduces geometric membership leakage; backbone and loss head have minor effects.
  • Even at large training-set sizes where individual statistics are weak, combining geometry statistics with a classifier reveals additional membership information, so audits should fuse statistics.
  • The measured separability is a lower bound on membership information available from embeddings; richer models could extract more.
  • The monotonic decrease implies that models trained on small identity pools (e.g., niche or custom datasets) are the most vulnerable to this type of geometric membership inference.

Reading between the lines

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

  • The monotonic decrease with training-set size suggests a scaling law: leakage may continue to drop with even larger identity counts, potentially below practical attack thresholds, though the paper does not test that.
  • The finding that cross-domain benchmarks inflate leakage implies that prior membership-inference studies on face recognition may have overstated real-world risk for models trained on large public data; replicating those attacks with same-domain references could change reported accuracy numbers.
  • Because the statistics are cluster-level and require only a few probe images per identity, an auditor could probe any deployed model with a handful of photos; the fusion result suggests a practical auditing tool, but also means the same technique could be used adversarially.
  • A testable extension would train models on different data sources (the paper itself flags this as open) to see whether the monotonic trend and factor ranking hold across data distributions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper quantifies how much membership information about training identities is retained in the geometry of face-recognition embedding clusters. The authors train 180 face-recognition models in a fully crossed design over three IResNet backbones, three angular-margin loss heads, five training-set sizes (1K/10K/50K/100K/all WebFace4M identities), and four training durations. For each model they compute four per-identity cluster statistics (pairwise cosine similarity, vMF concentration, penalized logit, and prototype softmax CE) and evaluate member/non-member separability (AUC) on nine benchmarks, including a same-domain held-out WebFace4M split and eight cross-domain benchmarks. The central claims are: (1) the number of training identities has the largest effect on separability, with backbone and loss head contributing little; (2) on the same-domain reference, the membership signal decreases monotonically with n_ids; (3) cross-domain benchmarks inflate the apparent signal; and (4) fusing all four statistics with a learned classifier recovers additional membership information. The paper also reports verification EERs to position the models within a realistic quality range.

Significance. If the claims hold, this is a valuable and unusually large-scale measurement contribution to the privacy-auditing literature. The 180-model factorial design, the matched-pair cross-domain comparisons (CFP-FF vs. CFP-FP, LFW vs. XQLFW), and the 5-fold stratified fusion evaluation are concrete strengths. The paper's main value is in identifying training-set size and training duration as the dominant controls on geometric membership leakage and in demonstrating that cross-domain benchmarks conflate domain shift with training-attributable signal. These are actionable findings for practitioners designing privacy audits. The paper is also honest about its limitations (single training source, no defenses, statistical assumptions), which increases confidence in its overall framing.

major comments (2)
  1. [Section 4.3, Table 3] The same-domain held-out WebFace4M split is undefined at n_ids=all: Section 4.3 states 'at n_ids=all no non-members remain.' Yet Table 3 reports a marginal PairCos AUC of 0.71 for n_ids=all, and the abstract claims the geometric membership signal 'decreases monotonically as more identities are added to training' on the same-domain reference. This is internally inconsistent. The same-domain curve has no valid point at n_ids=all, so the monotonic decrease is only supported from 1K to 100K. If the 'all' value was computed from cross-domain non-members, then Table 3 mixes different non-member populations and the monotonic sequence is not a same-domain curve. The authors must either exclude n_ids=all from same-domain analyses (and revise the abstract/conclusion accordingly) or explicitly define and justify a same-domain non-member pool for n_ids=all. This is load-bearing for the headline clai
  2. [Section 5.2, Table 3, Abstract] The abstract says 'the number of training identities has the largest effect on member/non-member separability,' but Table 3 shows that training duration (Epoch) has a comparable partial eta-squared: for PairCos both n_ids and Epoch are 0.99, and for PenLogit n_ids is 0.99 vs. 0.98 for Epoch. The paper's own text in Section 5.2 acknowledges that training duration is a close second. Without confidence intervals or a formal statistical comparison on the eta-squared values, the claim of 'largest effect' is not well supported. Please qualify the statement to reflect that n_ids and training duration both have dominant effects, with n_ids slightly larger, or provide uncertainty quantification showing the difference is significant.
minor comments (5)
  1. [Table 3 and throughout] The notation for the pairwise cosine statistic appears as both 'PairCos' (Table 3) and 'PAIRCOS' (text and Figure 2). Please standardize.
  2. [General] There are typographical artifacts such as 'ANOV A' and 'IRESNET' in tables; these should be fixed.
  3. [Section 3.2] In the vMF concentration formula, \bar{R} is used before it is defined; reorder the definitions for readability.
  4. [Figure 2] The dashed 'Within-dist' curve is difficult to distinguish in grayscale; consider a different line style or a separate legend callout.
  5. [Section 5.3] The probit-transform translation of Cohen's d' to AUC change is presented without a worked example; a short numerical illustration would help readers interpret the +0.26 and +0.22 values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical measurement study whose claims rest on direct evaluation, not on a derivation that reduces to its inputs. The n_ids=all benchmark gap is a data-validity concern, not a circular step.

full rationale

The paper does not claim to derive membership signal from first principles; it measures the separability of member/non-member embedding-cluster statistics across a factorial grid of 180 models. The four statistics (PairCos, vMF-kappa, PENLOGIT, PROTOCE) are defined a priori from cluster geometry; member/non-member labels are used only to evaluate AUC, not to compute the statistics, so no fitted value is renamed as a prediction. PENLOGIT and PROTOCE deliberately reuse the training loss scale s and margin m, but this is an explicit grey-box measurement choice (Section 3.2), not a way of encoding the membership answer into the input. The ANOVA and MLP fusion are descriptive/cross-validated analyses of measured quantities. There are no load-bearing self-citations: all cited prior work (loss functions, vMF estimation, benchmarks, attack methods) is external to the authors, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The paper even acknowledges that PairCos is equivalent to Li et al.'s intra-similarity, so it is not renaming a known result. The only notable issue is the same-domain held-out benchmark being undefined at n_ids=all: Section 4.3 states 'at n_ids=all no non-members remain,' yet Table 3 reports an 'all' PairCos AUC of 0.71 and the abstract claims monotonic decrease on the same-domain reference. This is an internal-consistency/validity gap in the largest factor level, not a circular reduction; it should be weighed as a correctness risk rather than as circularity. Section 5.5 also concedes the random-partition limitation, again a confound, not a circular step. Therefore the circularity score is 0.

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

The paper introduces no fitted parameters: the threshold sweep is summarized as a threshold-free AUC, the loss scale/margin are taken from prior art, and the MLP fusion hyperparameters are standard and cross-validated. The claims rest on several domain assumptions, most importantly the validity of the same-domain held-out control, which the authors themselves partially qualify.

assumptions (6)
  • domain assumption Angular-margin losses (ArcFace/CosFace/MagFace) optimize per-identity cluster compactness, so a geometric membership signal may exist.
    Motivates the entire study (Sections 1, 3.2).
  • domain assumption Random partitioning of WebFace4M identities into training set and held-out non-members removes all domain shift.
    The same-domain control in Section 4.3 is the basis for the monotonic-decrease and inflation claims; the authors acknowledge in Section 5.5 that membership boundaries may correlate with identity-level attributes.
  • ad hoc to paper The probit transform in Section 5.3 assumes approximately Gaussian member/non-member statistic distributions with stable variances.
    Used to convert matched-pair Cohen's d' into AUC changes; the assumption is stated in Section 5.5.
  • domain assumption The four cluster statistics capture enough of the geometry that the measured separability is a meaningful lower bound on membership information.
    Section 5.5 explicitly frames the measured signal as a lower bound, so the analysis is honest about scope.
  • domain assumption The varying per-identity probe count k (not controlled across benchmarks) does not non-uniformly bias the cluster-statistic comparisons.
    Section 5.5 lists 'the number of probe images k per identity varies across benchmarks and is not controlled for' as a limitation; the main conclusions assume this does not reverse the factor ranking.
  • domain assumption All models share a single training source (WebFace4M); results generalize to other datasets.
    Section 5.5 states generalisation to other training data is an open question.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying Training Membership Information in the Hyperspherical Embedding Geometry of Face Recognition Models." pith.science (2026). https://pith.science/paper/NUITA4UA

@misc{pith2026260715084,
  author       = {Pith},
  title        = {Pith review of: Quantifying Training Membership Information in the Hyperspherical Embedding Geometry of Face Recognition Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NUITA4UA}},
  note         = {Machine review of arXiv:2607.15084}
}
read the original abstract

Face recognition models represent each face as an embedding vector on the unit hypersphere by clustering embeddings of the same identity while pushing different identities apart through angular-margin losses. Because these losses act only on training identities, non-member identities may form clusters with different geometric properties. In this paper, we quantify the magnitude of this difference and what training-time factors control it. We compute four statistics based on cluster geometry across 180 face recognition models in a factorial design over IResNet backbone size, loss head, training duration, and the number of training identities, and evaluate each configuration on nine benchmarks. Our results indicate that the number of training identities has the largest effect on member/non-member separability, while backbone and loss head contribute far less, and that, on a same-domain held-out reference, the geometric membership signal decreases monotonically as more identities are added to training. We provide an analysis of cross-domain (pose, age, quality, ethnicity) non-member benchmarks and report that these inflate the apparent membership signal. Finally, we fuse all four statistics with a learned classifier to reveal additional membership information beyond the best individual statistic.

Figures

Figures reproduced from arXiv: 2607.15084 by the authors.

Figure 1
Figure 1. Measurement pipeline. Per-identity statistics ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. PAIRCOS (s¯) kernel density estimates. The bold black curve (with shading) shows member identities (present in the training set). All other curves show non-member identities: the nine thin solid lines correspond to cross-distribution benchmarks whose domain differs from the training data (one colour per dataset; see legend), while the dashed line (“Within-dist”) shows non-members from the held-out WebFace4M split, w… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 3 linked inside Pith

  1. [1]

    Abadi, A

    M. Abadi, A. Chu, I. J. Goodfellow, H. B. McMahan, I. Mironov, K. Talwar, and L. Zhang. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Se- curity, CCS 2016, Vienna, Austria, October 24-28, 2016, pages 308–318. ACM, 2016. 3, 8

    2016

  2. [2]

    X. An, J. Deng, J. Guo, Z. Feng, X. Zhu, J. Yang, and T. Liu. Killing two birds with one stone: Efficient and robust train- ing of face recognition cnns by partial FC. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 4032–4041. IEEE, 2022. 6

    2022

  3. [3]

    Banerjee, I

    A. Banerjee, I. S. Dhillon, J. Ghosh, and S. Sra. Clustering on the unit hypersphere using von mises-fisher distributions. J. Mach. Learn. Res., 6:1345–1382, 2005. 3

    2005

  4. [4]

    Bourtoule, V

    L. Bourtoule, V . Chandrasekaran, C. A. Choquette-Choo, H. Jia, A. Travers, B. Zhang, D. Lie, and N. Papernot. Ma- chine unlearning. In42nd IEEE Symposium on Security and Privacy, SP 2021, San Francisco, CA, USA, May 23-27, 2021, pages 141–159. IEEE, 2021. 3

    2021

  5. [5]

    Carlini, S

    N. Carlini, S. Chien, M. Nasr, S. Song, A. Terzis, and F. Tram`er. Membership inference attacks from first prin- ciples. In43rd IEEE Symposium on Security and Privacy, SP 2022, San Francisco, CA, USA, May 22-26, 2022, pages 1897–1914. IEEE, 2022. 2

    2022

  6. [6]

    G. Chen, Y . Zhang, and F. Song. SLMIA-SR: Speaker- level membership inference attacks against speaker recog- nition systems. In31st Annual Network and Distributed Sys- tem Security Symposium, NDSS 2024, San Diego, CA, USA, February 26 - March 1, 2024. The Internet Society, 2024. 2

    2024

  7. [7]

    M. Chen, Z. Zhang, T. Wang, M. Backes, and Y . Zhang. FACE-AUDITOR: Data auditing in facial recognition sys- tems. In32nd USENIX Security Symposium, USENIX Se- curity 2023, Anaheim, CA, USA, August 9-11, 2023, pages 7195–7212. USENIX Association, 2023. 2

    2023

  8. [8]

    DeAlcala, G

    D. DeAlcala, G. Mancera, A. Morales, J. Fi ´errez, R. Tolosana, and J. Ortega-Garcia. A comprehensive analy- sis of factors impacting membership inference. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2024, Seattle, WA, USA, June 17-18, 2024, pages 3585–3593. IEEE, 2024. 2, 3

    2024

Show all 38 references
  1. [9]

    DeAlcala, A

    D. DeAlcala, A. Morales, J. Fierrez, G. Mancera, R. Tolosana, and J. Ortega-Garcia. Is my data in your AI? Membership inference test (MINT) applied to face biomet- rics.IEEE Access, 13:163805–163819, 2025. 2

    2025

  2. [10]

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InIEEE Con- ference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690–

    2019

  3. [11]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In9th International Conference on Learnin...

    2021

  4. [12]

    V . Feldman. Does learning require memorization? A short tale about a long tail. InProceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing, STOC 2020, Chicago, IL, USA, June 22-26, 2020, pages 954–959. ACM, 2020. 3

    2020

  5. [13]

    Fredrikson, S

    M. Fredrikson, S. Jha, and T. Ristenpart. Model inversion at- tacks that exploit confidence information and basic counter- measures. InProceedings of the 22nd ACM SIGSAC Confer- ence on Computer and Communications Security, CCS 2015, Denver, CO, USA, October 12-16, 2015, pages...

    2015

  6. [14]

    Official Journal of the European Union, L 119, 1–88, 2016

    Regulation (EU) 2016/679 of the European Parliament and of the Council (general data protection regulation). Official Journal of the European Union, L 119, 1–88, 2016. Art. 17 (Right to erasure). 2

    2016

  7. [15]

    Golatkar, A

    A. Golatkar, A. Achille, and S. Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2020, Seattle, WA, USA, June 13- 19, 2020, pages 9301–9309. Computer Vision Found...

    2020

  8. [16]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Com- puter Vision and Pattern Recognition, CVPR 2016, Las Ve- gas, NV , USA, June 27-30, 2016, pages 770–778. IEEE Com- puter Society, 2016. 4

    2016

  9. [17]

    G. E. Hinton, O. Vinyals, and J. Dean. Distilling the knowl- edge in a neural network.CoRR, abs/1503.02531, 2015. NeurIPS Deep Learning Workshop. 8

    2015 arXiv

  10. [18]

    H. Hu, Z. Salcic, L. Sun, G. Dobbie, P. S. Yu, and X. Zhang. Membership inference attacks on machine learning: A sur- vey.ACM Comput. Surv., 54(11s):1–37, 2022. 2

    2022

  11. [19]

    G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Re- port 07-49, University of Massachusetts, Amherst, 2007. 4, 5

    2007

  12. [20]

    Huang, H

    Y . Huang, H. Chen, Y . Wang, and L. Wang. Inference attacks against face recognition model without classification layers. CoRR, abs/2401.13719, 2024. 2, 3

    2024 arXiv

  13. [21]

    Knoche, S

    M. Knoche, S. H ¨ormann, and G. Rigoll. Cross-quality LFW: A database for analyzing cross- resolution image face recog- nition in unconstrained environments. In16th IEEE Interna- tional Conference on Automatic Face and Gesture Recogni- tion, FG 2021, Jodhpur, India, December 1...

    2021

  14. [22]

    G. Li, S. Rezaei, and X. Liu. User-level membership in- ference attack against metric embedding learning.CoRR, abs/2203.02077, 2022. 2, 3

    2022 arXiv

  15. [23]

    H. Liu, J. Jia, W. Qu, and N. Z. Gong. EncoderMI: Mem- bership inference against pre-trained encoders in contrastive learning. InProceedings of the 2021 ACM SIGSAC Confer- ence on Computer and Communications Security, CCS 2021, Virtual Event, Republic of Korea, November 15-19,...

    2021

  16. [24]

    Mancera, D

    G. Mancera, D. DeAlcala, A. Morales, R. Tolosana, and J. Fierrez. Membership inference test: Auditing training data in object classification models. InDeployable AI Workshop (DAI 2025), co-located with the 39th AAAI Conference on Artificial Intelligence (AAAI-25), 2025. 2, 3

    2025

  17. [25]

    B. Maze, J. C. Adams, J. A. Duncan, N. D. 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, Gold Coast, Australia, February...

    2018

  18. [26]

    Q. Meng, S. Zhao, Z. Huang, and F. Zhou. Magface: A uni- versal representation for face recognition and quality assess- ment. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14225–14234. Computer Vision Foundation / IEE...

    2021

  19. [27]

    Moschoglou, A

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kot- sia, and S. Zafeiriou. Agedb: The first manually collected, in-the-wild age database. In2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 20...

    2017

  20. [28]

    Rezaei and X

    S. Rezaei and X. Liu. On the difficulty of membership in- ference attacks. InIEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2021, virtual, June 19- 25, 2021, pages 7892–7900. Computer Vision Foundation / IEEE, 2021. 3

    2021

  21. [29]

    Salem, Y

    A. Salem, Y . Zhang, M. Humbert, P. Berrang, M. Fritz, and M. Backes. ML-Leaks: Model and data independent mem- bership inference attacks and defenses on machine learning models. In26th Annual Network and Distributed System Se- curity Symposium, NDSS 2019, San Diego, CA, USA, ...

    2019

  22. [30]

    Sengupta, J

    S. Sengupta, J. Chen, C. D. Castillo, V . M. Patel, R. Chel- lappa, and D. W. Jacobs. Frontal to profile face verification in the wild. In2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, Lake Placid, NY, USA, March 7-10, 2016, pages 1–9. IEEE Computer...

    2016

  23. [31]

    Shokri, M

    R. Shokri, M. Stronati, C. Song, and V . Shmatikov. Member- ship inference attacks against machine learning models. In 2017 IEEE Symposium on Security and Privacy, SP 2017, San Jose, CA, USA, May 22-26, 2017, pages 3–18. IEEE Computer Society, 2017. 2

    2017

  24. [32]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J ´egou. Training data-efficient image transformers & distillation through attention. In M. Meila and T. Zhang, ed- itors,Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 Jul...

    2021

  25. [33]

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu. Cosface: Large margin cosine loss for deep face recognition. In2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265–5274. Comput...

    2018

  26. [34]

    M. Wang, W. Deng, J. Hu, X. Tao, and Y . Huang. Racial faces in the wild: Reducing racial bias by information max- imization adaptation network. In2019 IEEE/CVF Interna- tional Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 692–

    2019

  27. [35]

    Zheng and W

    T. Zheng and W. Deng. Cross-pose LFW: A database for studying cross-pose face recognition in unconstrained envi- ronments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018. 4, 5

    2018

  28. [36]

    Z. Zhu, G. Huang, J. Deng, Y . Ye, J. Huang, X. Chen, J. Zhu, T. Yang, D. Du, J. Lu, and J. Zhou. Webface260m: A bench- mark for million-scale deep face recognition.IEEE Trans. Pattern Anal. Mach. Intell., 45(2):2627–2644, 2023. 1, 4

    2023

  29. [2021]

    OpenReview.net, 2021. 6

    2021

  30. [4699]

    Computer Vision Foundation / IEEE, 2019. 1, 3, 4

    2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.