Pith. sign in

REVIEW 2 major objections 1 minor 44 references

What's Old is New Again: Classical Dimensionality Reduction for Efficient Saliency-Guided Biometric Attack Detection

T0 review · 2 major / 1 minor · reviewed 2026-06-27 · grok-4.3

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

Pith's one-line read Saliency maps generated from PCA and LDA on raw data enable effective training for biometric presentation attack detection without annotations or domain knowledge.

desk verdict PCA and LDA can generate usable saliency maps for biometric PAD from raw data alone, and the multi-domain tests show competitive results, but the zero-tuning claim rests on how the maps are actually extracted from the components. read the letter →

arxiv 2606.13528 v1 pith:WPRQHKMS submitted 2026-06-11 cs.CV

Samuel Webster , Walter Scheirer This is my paper · ORCID
classification cs.CV
keywords saliency-guidedtrainingpresentationattackdetectionPCALDAbiometricPADdimensionalityreductionirisfingerprint
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

The paper shows that classical dimensionality reduction can directly produce saliency maps from training images for use in saliency-guided model training. These maps require no human labels and apply across iris PAD, synthetic face detection, fingerprint PAD, fingerprint vein PAD, and ID card PAD. When incorporated into training, the resulting models outperform standard baselines and sometimes match or surpass existing saliency techniques while using zero extra resources. The approach removes the main practical obstacles that have limited saliency-guided training in biometric security.

What carries the argument

Saliency maps produced by applying PCA and LDA directly to raw training images, which serve as attention guides during model training.

What would settle it

A controlled test in which models trained with PCA or LDA saliency maps show no improvement over models trained with no saliency guidance or with random maps on a new biometric domain.

Watch

Extended reading notes

Core claim

Saliency maps derived from principal component analysis and linear discriminant analysis on raw biometric training data allow models to achieve higher robustness and generalization in presentation attack detection than baseline methods, and sometimes state-of-the-art saliency approaches, across five tested modalities.

Load-bearing premise

Saliency maps from PCA and LDA applied to raw training data capture the features that matter for detecting presentation attacks in different biometric types.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper claims that saliency maps derived from classical dimensionality reduction techniques (PCA and LDA) applied directly to raw training data can be used for saliency-guided training in biometric presentation attack detection (PAD). These maps require no human annotation or domain knowledge and lead to models that outperform baselines and sometimes state-of-the-art saliency methods across multiple domains: iris PAD, synthetic face detection, fingerprint PAD, fingerprint vein PAD, and ID card PAD.

Significance. If the central claims hold, this work would offer a highly scalable and zero-cost alternative to existing saliency acquisition methods, removing a key barrier to adopting saliency-guided training in biometric security applications. The multi-domain evaluation, including novel domains, strengthens the potential impact if the no-tuning aspect is confirmed.

major comments (2)
  1. [Abstract and Methods] The claim that the approach requires 'no ... domain knowledge' and 'without any resource investment or domain-specific tooling' (Abstract) is load-bearing for the scalability and attribution arguments, but the conversion from PCA/LDA outputs (global eigenvectors or class means) to localized per-pixel saliency maps requires at least one non-trivial step such as component selection, absolute loadings, reconstruction error, or thresholding; these steps are not demonstrated to be free of per-domain choices.
  2. [Abstract] The abstract asserts outperformance over baselines and SOTA without providing experimental details, data, or verification steps, making the central performance claim difficult to assess from available information.
minor comments (1)
  1. [Abstract] The distinction between 'saliency-explored domains' and 'saliency-novel domains' is introduced without an explicit definition or reference to prior work.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the presentation of our claims regarding scalability and verifiability. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Abstract and Methods] The claim that the approach requires 'no ... domain knowledge' and 'without any resource investment or domain-specific tooling' (Abstract) is load-bearing for the scalability and attribution arguments, but the conversion from PCA/LDA outputs (global eigenvectors or class means) to localized per-pixel saliency maps requires at least one non-trivial step such as component selection, absolute loadings, reconstruction error, or thresholding; these steps are not demonstrated to be free of per-domain choices.

    Authors: We agree that explicit specification of the mapping procedure is necessary to substantiate the no-domain-knowledge claim. Our method uses a single, fixed, parameter-free pipeline applied identically to all five domains: PCA saliency is the normalized absolute loadings of the first principal component; LDA saliency is the normalized absolute values of the between-class mean difference in the leading discriminant direction. No component selection, per-domain thresholding, or reconstruction-error tuning occurs. Section 2.2 already describes this procedure, but to directly address the concern we will add a short subsection with pseudocode confirming the steps are domain-agnostic and require no choices or tooling. revision: partial

  2. Referee: [Abstract] The abstract asserts outperformance over baselines and SOTA without providing experimental details, data, or verification steps, making the central performance claim difficult to assess from available information.

    Authors: Abstracts conventionally summarize results at a high level; the full experimental protocol, datasets, metrics, and quantitative comparisons (including all baseline and SOTA numbers) appear in Sections 3–5 and Tables 1–5. The outperformance statements are therefore verifiable from the manuscript body. We do not believe the abstract requires experimental details, but if the editor prefers we can append a single sentence noting the five-domain, cross-dataset evaluation protocol. revision: no

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard DR applied directly to training data

full rationale

The paper applies classical PCA and LDA to raw training images to produce saliency maps, with no equations, fitted parameters renamed as predictions, or self-citation chains that reduce the central claim to its own inputs. The method is presented as a direct, annotation-free use of existing dimensionality reduction on the training set itself; no load-bearing uniqueness theorems, ansatzes smuggled via citation, or self-definitional loops appear in the abstract or described approach. The performance claims rest on empirical comparison across domains rather than any derivation that collapses by construction.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The approach rests on standard machine learning assumptions about the value of saliency guidance and the applicability of classical dimensionality reduction to image data for feature highlighting.

assumptions (1)
  • domain assumption Saliency-guided training improves model robustness and generalization in biometric presentation attack detection
    Described as a paradigm with shown benefits in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of What's Old is New Again: Classical Dimensionality Reduction for Efficient Saliency-Guided Biometric Attack Detection." pith.science (2026). https://pith.science/paper/WPRQHKMS

@misc{pith2026260613528,
  author       = {Pith},
  title        = {Pith review of: What's Old is New Again: Classical Dimensionality Reduction for Efficient Saliency-Guided Biometric Attack Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPRQHKMS}},
  note         = {Machine review of arXiv:2606.13528}
}
read the original abstract

Saliency-guided training is a paradigm in visual recognition that encourages models to focus on the most relevant image regions during learning. While its application in biometric presentation attack detection (PAD) has shown strong benefits in robustness and generalization, adoption is often limited by the high cost, domain specificity, and limited scalability of existing saliency acquisition methods, such as human annotations over a limited dataset. We present a novel, cost-efficient, and highly-scalable approach to saliency acquisition using maps inspired by classical dimensionality reduction techniques: PCA and LDA. Our proposed methods generate saliency maps directly from raw training data, requiring no human annotation nor domain knowledge. We contextualize the effectiveness of these saliency sources in three saliency-explored domains (iris PAD, synthetic face detection, fingerprint PAD) and demonstrate its scalability in two saliency-novel domains (fingerprint vein PAD and ID card PAD). Across all domains tested, models trained using dimensionality reduction-sourced saliency maps exceed baseline and sometimes SOTA saliency methods without any resource investment or domain-specific tooling. Our findings overcome an important yet unaddressed barrier to saliency-guided training for biometric attack detection and beyond.

Figures

Figures reproduced from arXiv: 2606.13528 by the authors.

Figure 1
Figure 1. A brief comparison between existing saliency strate￾gies and our proposed method. While existing pipelines con￾tain high-expense human annotation collection studies or specific domain technologies, our methods building upon classical dimen￾sionality reduction have virtually no barrier to application in any domain. However, acquisition of saliency maps are a major bottleneck for applying saliency-guided methods to ne… view at source ↗
Figure 2
Figure 2. Examples of proposed dimensionality reduction-based saliency types across all explored biometric attack domains. Each row provides a single sample and four distinct saliency maps from the evaluated datasets. are composed to create the final LDA saliency. The saliency is identically processed with absolute value and normalization to {0-255}. All saliency types are visually represented for all four domains explored in… view at source ↗
Figure 3
Figure 3. Changes to accuracy when varying Eigenface M-% across all biometric attack detection domains. Each domain’s solid line denotes mean and transparent region denotes ±1 stan￾dard deviation. There is a subtle global preference toward lower error percentiles and a negative reaction to mid-range values in ID Document and Fingerprint Vein PAD. methods achieving an AUC of 1.0 as well as near 0.0 error in bonafide classifica… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Changes to AUC, APCER, and BPCER when varying Eigenface M-% across all biometric attack detection domains. Due to space constraints and limited expressivity of AUC and BPCER in Fingerprint Vein PAD, we prioritized inclusion of the Accuracy figure in the main paper. How…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 canonical work pages

  1. [1]

    http://www.cbsr.ia.ac.cn/china/Iris%20Databases%20CH.asp

    Chinese Academy of Sciences Institute of Automation. http://www.cbsr.ia.ac.cn/china/Iris%20Databases%20CH.asp. Accessed: 03-12-2021. 3

    2021

  2. [2]

    Banerjee, J

    S. Banerjee, J. S. Bernhard, W. J. Scheirer, K. W. Bowyer, and P. J. Flynn. Srefi: Synthesis of realistic example face images. In2017 IEEE International Joint Conference on Biometrics (IJCB), pages 37–45. IEEE, 2017. 3

    2017

  3. [3]

    P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman. Eigen- faces vs. fisherfaces: Recognition using class specific linear projection.IEEE Transactions on pattern analysis and ma- chine intelligence, 19(7):711–720, 1997. 2, 3, 4, 7, 8

    1997

  4. [4]

    Bhattacharjee, D

    S. Bhattacharjee, D. Geissbuhler, G. Clivaz, K. Kotwal, and S. Marcel. Vascular biometrics experiments on candy – a new contactless finger-vein dataset. InProceedings of the In- ternational Conference on Pattern Recognition (ICPR), Dec

  5. [5]

    Boned, M

    C. Boned, M. Talarmain, N. Ghanmi, G. Chiron, S. Biswas, A. M. Awal, and O. Ramos Terrades. Synthetic dataset of id and travel documents.Scientific data, 11(1):1356, 2024. 2

    2024

  6. [6]

    A. Boyd, K. W. Bowyer, and A. Czajka. Human-aided saliency maps improve generalization of deep learning. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 2735–2744, 2022. 1, 2, 3, 4

    2022

  7. [7]

    A. Boyd, Z. Fang, A. Czajka, and K. W. Bowyer. Iris presen- tation attack detection: Where are we now?Pattern Recog- nition Letters, 138:483–489, 2020. 3

    2020

  8. [8]

    A. Boyd, P. Tinsley, K. W. Bowyer, and A. Czajka. Cy- borg: Blending human saliency into the loss improves deep learning-based synthetic face detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 6108–6117, 2023. 1, 2, 3, 4, 5, 11

    2023

Show all 44 references
  1. [9]

    Bulatov, E

    K. Bulatov, E. Emelianova, D. Tropin, N. Skoryukina, Y . Chernyshova, Z. Ming, J.-C. Burie, and M. M. Luqman. Midv-2020: A comprehensive benchmark dataset for iden- tity document analysis.Computer Optics, 46(2):252–270,

    2020

  2. [10]

    Casula, M

    R. Casula, M. Micheletto, G. Orr `u, R. Delussu, S. Concas, A. Panzino, and G. L. Marcialis. Livdet 2021 fingerprint liveness detection competition-into the unknown. In2021 IEEE international joint conference on biometrics (IJCB), pages 1–6. IEEE, 2021. 3

    2021

  3. [11]

    Y . Choi, Y . Uh, J. Yoo, and J.-W. Ha. Stargan v2: Diverse image synthesis for multiple domains. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188–8197, 2020. 3

    2020

  4. [12]

    C. R. Crum, A. Boyd, K. Bowyer, and A. Czajka. Teach- ing ai to teach: Leveraging limited human salience data into unlimited saliency-based training.arXiv preprint arXiv:2306.05527, 2023. 1, 2, 3, 4

    2023

  5. [13]

    C. R. Crum, S. Webster, and A. Czajka. Grains of saliency: Optimizing saliency-based training of biometric attack de- tection models. In2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–9. IEEE, 2024. 1, 2, 3, 4, 5, 6, 7, 11

    2024

  6. [14]

    P. Das, J. Mcfiratht, Z. Fang, A. Boyd, G. Jang, A. Mo- hammadi, S. Purnapatra, D. Yambay, S. Marcel, M. Trok- ielewicz, P. Maciejewicz, K. Bowyer, A. Czajka, S. Schuck- ers, J. Tapia, S. Gonzalez, M. Fang, N. Damer, F. Boutros, A. Kuijper, R. Sharma, C. Chen, and A. Ross. Iri...

    2020

  7. [15]

    Galbally, J

    J. Galbally, J. Ortiz-Lopez, J. Fierrez, and J. Ortega-Garcia. Iris liveness detection based on quality related features. In 2012 5th IAPR Int. Conf. on Biometrics (ICB), pages 271– 276, New Delhi, India, March 2012. IEEE. 3

    2012

  8. [16]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. corr abs/1512.03385 (2015), 2015. 5

    2015 arXiv

  9. [17]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Wein- berger. Densely connected convolutional networks. InPro- ceedings of the IEEE conference on computer vision and pat- tern recognition, pages 4700–4708, 2017. 5

    2017

  10. [18]

    Hubert, J

    M. Hubert, J. Raymaekers, and P. J. Rousseeuw. Robust dis- criminant analysis.Wiley Interdisciplinary Reviews: Com- putational Statistics, 16(5):e70003, 2024. 8

    2024

  11. [19]

    Standard, International Or- ganization for Standardization, Geneva, CH, 2023

    Information technology – Biometric presentation attack de- tection – Testing and reporting. Standard, International Or- ganization for Standardization, Geneva, CH, 2023. 5

    2023

  12. [20]

    Karras, T

    T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Vari- ation.arXiv preprint arXiv:1710.10196, 2017. 3

    2017 arXiv

  13. [21]

    Karras, M

    T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, and T. Aila. Training generative adversarial networks with lim- ited data. InProc. NeurIPS, 2020. 3

    2020

  14. [22]

    Karras, M

    T. Karras, M. Aittala, S. Laine, E. H ¨ark¨onen, J. Hellsten, J. Lehtinen, and T. Aila. Alias-free generative adversarial networks.Proc. NeurIPS, 2021. 3

    2021

  15. [23]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila. Analyzing and improving the image quality of stylegan. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119,

  16. [24]

    Kohli, D

    N. Kohli, D. Yadav, M. Vatsa, and R. Singh. Revisiting iris recognition with color cosmetic contact lenses. In2013 International Conference on Biometrics (ICB), pages 1–7. IEEE, 2013. 3

    2013

  17. [25]

    Kohli, D

    N. Kohli, D. Yadav, M. Vatsa, R. Singh, and A. Noore. De- tecting medley of iris spoofing attacks using desist. In2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2016. 3

    2016

  18. [26]

    S. J. Lee, K. R. Park, Y . J. Lee, K. Bae, and J. H. Kim. Multifeature-based fake iris detection method.Optical En- gineering, 46(12):1 – 10, 2007. 3

    2007

  19. [27]

    Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face at- tributes in the wild. InProceedings of the IEEE international conference on computer vision, pages 3730–3738, 2015. 3

    2015

  20. [28]

    V . Mura, L. Ghiani, G. L. Marcialis, F. Roli, D. A. Yambay, and S. A. Schuckers. Livdet 2015 fingerprint liveness de- tection competition 2015. In2015 IEEE 7th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6, 2015. 3

    2015

  21. [29]

    V . Mura, G. Orr `u, R. Casula, A. Sibiriu, G. Loi, P. Tuveri, L. Ghiani, and G. L. Marcialis. Livdet 2017 fingerprint live- ness detection competition 2017. In2018 international con- ference on biometrics (ICB), pages 297–302. IEEE, 2018. 3

    2017

  22. [30]

    Netrapalli, U

    P. Netrapalli, U. N. Niranjan, S. Sanghavi, A. Anandkumar, and P. Jain. Non-convex robust pca.Advances in neural information processing systems, 27, 2014. 8

    2014

  23. [31]

    Orr `u, R

    G. Orr `u, R. Casula, P. Tuveri, C. Bazzoni, G. Dessalvi, M. Micheletto, L. Ghiani, and G. L. Marcialis. Livdet in action-fingerprint liveness detection competition 2019. In 2019 international conference on biometrics (ICB), pages 1–

    2019

  24. [32]

    P. J. Phillips, P. J. Flynn, and K. W. Bowyer. Lessons from collecting a million biometric samples.Image and Vision Computing, 58:96–107, 2017. 3

    2017

  25. [33]

    Szegedy, S

    C. Szegedy, S. Ioffe, V . Vanhoucke, and A. Alemi. Inception- v4, inception-resnet and the impact of residual connections on learning. InProceedings of the AAAI conference on arti- ficial intelligence, volume 31, 2017. 5

    2017

  26. [34]

    Trokielewicz, A

    M. Trokielewicz, A. Czajka, and P. Maciejewicz. Assess- ment of iris recognition reliability for eyes affected by ocular pathologies. In2015 IEEE 7th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2015. 3

    2015

  27. [35]

    Trokielewicz, A

    M. Trokielewicz, A. Czajka, and P. Maciejewicz. Post- mortem iris recognition with deep-learning-based image seg- mentation.Image and Vision Computing, 94:103866, 2020. 3

    2020

  28. [36]

    M. A. Turk, A. Pentland, et al. Face recognition using eigen- faces. InCVPR, volume 91, pages 586–591, 1991. 2, 3, 4

    1991

  29. [37]

    Webster and A

    S. Webster and A. Czajka. Saliency-guided training for fin- gerprint presentation attack detection. In2025 IEEE Inter- national Joint Conference on Biometrics (IJCB), pages 1–10,

  30. [38]

    1, 2, 3, 4, 5, 6, 7, 11

  31. [39]

    Webster, W

    S. Webster, W. Scheirer, and A. Czajka. Psychophysically- guided training for fingerprint presentation attack detection. IEEE Transactions on Biometrics, Behavior, and Identity Science, pages 1–1, 2026. 2, 3, 5, 11

    2026

  32. [40]

    Z. Wei, T. Tan, and Z. Sun. Synthesis of large realistic iris databases using patch-based sampling. In2008 19th Interna- tional Conference on Pattern Recognition, pages 1–4. IEEE,

  33. [41]

    Yambay, B

    D. Yambay, B. Becker, N. Kohli, D. Yadav, A. Czajka, K. W. Bowyer, S. Schuckers, R. Singh, M. Vatsa, A. Noore, et al. Livdet iris 2017—iris liveness detection competition 2017. In2017 IEEE International Joint Conference on Biometrics (IJCB), pages 733–741. IEEE, 2017. 3

    2017

  34. [42]

    Yambay, B

    D. Yambay, B. Becker, N. Kohli, D. Yadav, A. Czajka, K. W. Bowyer, S. Schuckers, R. Singh, M. Vatsa, A. Noore, et al. Livdet iris 2017—iris liveness detection competition 2017. In2017 IEEE International Joint Conference on Biometrics (IJCB), pages 733–741. IEEE, 2017. 3 Appendix

    2017

  35. [43]

    Supplementary Eigenface M-% Ablation Charts AUC APCER BPCER 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Reconstruction Error (M-%) 0.2 0.1 0.0 0.1 0.2 AUC (vs. Baseline) Eigenface M-% Ablation: AUC Iris PAD Synthetic Face Detection Fingerprint PAD Fingerprint Vein PAD Identification D...

  36. [44]

    All other rows are trained using saliency-guidance with the specified saliency type, applying the established CYBORG loss formulation [8] and alpha tuning scheme [38]

    Unabridged Training Results All configurations denoted as ‘Baseline’ and having gray rows are trained using cross entropy loss and without any saliency guidance, following existing baseline approaches [8, 13, 37, 38]. All other rows are trained using saliency-guidance with the...

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.