pith. sign in

arxiv: 1907.07270 · v1 · pith:BS2LTQQCnew · submitted 2019-07-16 · 💻 cs.CV

Style Transfer Applied to Face Liveness Detection with User-Centered Models

Israel A. Laurensi R. , Luciana T. Menon , Manoel Camillo O. Penna N. , Alessandro L. Koerich , Alceu S. Britto Jr This is my paper

Pith reviewed 2026-05-24 20:39 UTC · model grok-4.3

classification 💻 cs.CV
keywords face liveness detectionanti-spoofingstyle transferuser-centered modelsconvolutional neural networkSiW databasespoof image generation
0
0 comments X

The pith

Style transfer generates spoof images to train user-centered face liveness models without real fraudulent samples from each user.

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

The paper proposes a face anti-spoofing user-centered model (FAS-UCM) that addresses the challenge of collecting spoof images from every user for training. It divides the method into three parts: using style transfer to generate new spoof images from representation models, training a CNN for liveness detection, and evaluating live versus spoof images per subject. The CNN shows promising generalization for style transfer tasks. On the SiW database, this yields an average classification error rate of 0.22 for distinguishing live from spoof images. The approach enables per-user models by substituting generated spoofs for hard-to-obtain real ones.

Core claim

The proposed FAS-UCM uses style transfer to create spoof images, trains a CNN on them for liveness detection, and evaluates per subject to distinguish live and spoof images, achieving an average classification error rate of 0.22 on the SiW database while overcoming the need for real fraudulent images from all users.

What carries the argument

The FAS-UCM pipeline, which applies style transfer and spoof image representation models to generate training data for per-user CNN liveness detectors.

If this is right

  • Per-user liveness detectors become feasible without collecting real spoof samples from each individual.
  • The CNN generalizes sufficiently to perform style transfer with promising qualitative results.
  • The three-part structure separates generation of spoofs, CNN training, and per-subject evaluation.
  • An average error rate of 0.22 is achieved when distinguishing live and spoof images on the SiW database.

Where Pith is reading between the lines

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

  • The technique could extend to other biometric systems where per-user attack examples are scarce.
  • Efficiency of the style transfer step would determine suitability for real-time applications like mobile authentication.
  • Validation against physical and digital spoof variants beyond style transfer would test broader robustness.

Load-bearing premise

Spoof images generated by style transfer are representative enough of real-world spoof attacks to train effective per-user liveness detectors.

What would settle it

Testing the trained per-user CNN models on independently collected real-world spoof attacks (not produced by style transfer) and measuring whether the average classification error rate remains near 0.22.

Figures

Figures reproduced from arXiv: 1907.07270 by Alceu S. Britto Jr, Alessandro L. Koerich, Israel A. Laurensi R., Luciana T. Menon, Manoel Camillo O. Penna N..

Figure 1
Figure 1. Figure 1: Face samples from the SiW database: (a)–(d) live samples; (e)–(h) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spoof representations of each type of presentation attack and image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Spoof image generation protocol pipeline. The VGG19 [27] architecture was used to obtain the style of the images. For each spoof representation, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training process of a CNN for each subject. Two architectures were [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spoof images generated. On the left the live images and on the right [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results in terms of accuracy for each subject grouped in a boxplot [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

This paper proposes a face anti-spoofing user-centered model (FAS-UCM). The major difficulty, in this case, is obtaining fraudulent images from all users to train the models. To overcome this problem, the proposed method is divided in three main parts: generation of new spoof images, based on style transfer and spoof image representation models; training of a Convolutional Neural Network (CNN) for liveness detection; evaluation of the live and spoof testing images for each subject. The generalization of the CNN to perform style transfer has shown promising qualitative results. Preliminary results have shown that the proposed method is capable of distinguishing between live and spoof images on the SiW database, with an average classification error rate of 0.22.

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

2 major / 1 minor

Summary. The paper proposes a face anti-spoofing user-centered model (FAS-UCM) that generates synthetic spoof images via style transfer to train per-user CNN liveness detectors, addressing the difficulty of collecting real fraudulent samples from every subject. The approach consists of three stages: style-transfer-based spoof generation, CNN training for live/spoof classification, and per-subject evaluation. Preliminary results on the SiW database report an average classification error rate of 0.22.

Significance. If validated, the approach would allow user-specific anti-spoofing models without requiring real attack samples from each enrolled user, which could improve generalization in face recognition deployments. The work highlights a practical data-acquisition bottleneck but currently supplies insufficient experimental detail to assess whether the style-transfer proxy reproduces the visual statistics of genuine spoofs.

major comments (2)
  1. [Abstract] Abstract: the central claim of an average classification error rate of 0.22 on SiW is presented without any description of the number of subjects, train/test protocol, cross-validation procedure, baseline comparisons, or per-user error breakdown. These omissions make the quantitative result impossible to interpret or reproduce.
  2. [Abstract] Abstract (method description): the claim that style-transferred spoofs overcome the need for real fraudulent images rests on the untested assumption that the generated images reproduce the texture, moiré, and replay artifacts present in real attacks. No distribution comparison, ablation (real vs. synthetic negatives), or failure-case analysis is supplied to support this proxy.
minor comments (1)
  1. [Abstract] Abstract: 'divided in three main parts' should read 'divided into three main parts'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our preliminary work. We agree that the abstract requires more detail for interpretability and that the style-transfer proxy needs explicit discussion of its assumptions. We will revise the manuscript accordingly while preserving its focus on the user-centered approach.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of an average classification error rate of 0.22 on SiW is presented without any description of the number of subjects, train/test protocol, cross-validation procedure, baseline comparisons, or per-user error breakdown. These omissions make the quantitative result impossible to interpret or reproduce.

    Authors: We agree that the abstract is too concise and omits critical experimental context. The full manuscript references the SiW database (165 subjects) and describes subject-specific training, but we will expand the abstract to state the number of subjects, the per-user train/test protocol, and the preliminary nature of the 0.22 average error rate. In the revised results section we will add the cross-validation procedure, baseline comparisons (e.g., non-user-specific CNN), and per-user error breakdown to enable reproduction and assessment. revision: yes

  2. Referee: [Abstract] Abstract (method description): the claim that style-transferred spoofs overcome the need for real fraudulent images rests on the untested assumption that the generated images reproduce the texture, moiré, and replay artifacts present in real attacks. No distribution comparison, ablation (real vs. synthetic negatives), or failure-case analysis is supplied to support this proxy.

    Authors: The method is explicitly positioned as a practical proxy to avoid collecting real attack samples per user. We acknowledge that the current version provides no quantitative distribution comparison or ablation study. In revision we will add qualitative side-by-side examples of style-transferred versus real spoofs, discuss the assumption's limitations, and include failure-case analysis. A full real-vs-synthetic ablation would require new data collection outside the present scope, but we will clarify this boundary. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical pipeline with external test set

full rationale

The paper describes a three-stage empirical pipeline: (1) generate synthetic spoof images via style transfer, (2) train per-user CNNs on live images plus the generated spoofs, (3) evaluate classification error on the held-out SiW test set. No equations, fitted parameters, or derivations appear. The reported 0.22 average error is a direct empirical measurement on real test images, not a quantity obtained by renaming or re-using any training input. No self-citations are invoked as load-bearing premises, and the style-transfer step is presented as an external tool rather than a self-referential definition. The central claim therefore remains independent of its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review is based solely on the abstract; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.0 · 5676 in / 1077 out tokens · 17681 ms · 2026-05-24T20:39:13.951051+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Domain-generalizable Face Anti-Spoofing with Patch-based Multi-tasking and Artifact Pattern Conversion

    cs.CV 2026-04 unverdicted novelty 5.0

    PCGAN disentangles latent vectors for spoof artifacts and facial features to generate diverse spoof images, paired with patch-based multi-task learning to boost domain generalization and partial-attack detection in fa...

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    FaceNet: A unified embed- ding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “FaceNet: A unified embed- ding for face recognition and clustering,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, Jun. 2015

    work page 2015

  2. [2]

    DeepFace: Closing the gap to human-level performance in face verification,

    Y . Taigman, M. Yang, M. Ranzato, and L. Wolf, “DeepFace: Closing the gap to human-level performance in face verification,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition . IEEE, Jun. 2014

    work page 2014

  3. [3]

    Deep learning face repre- sentation by joint identification-verification,

    Y . Sun, Y . Chen, X. Wang, and X. Tang, “Deep learning face repre- sentation by joint identification-verification,” inProceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, ser. NIPS’14. Cambridge, MA, USA: MIT Press, 2014, pp. 1988–1996

    work page 2014

  4. [4]

    Face liveness detection from a single image with sparse low rank bilinear discriminative model,

    X. Tan, Y . Li, J. Liu, and L. Jiang, “Face liveness detection from a single image with sparse low rank bilinear discriminative model,” in Computer Vision – ECCV 2010, K. Daniilidis, P. Maragos, and N. Paragios, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2010, pp. 504–517

    work page 2010

  5. [5]

    A face anti- spoofing database with diverse attacks,

    Z. Zhang, J. Yan, S. Liu, Z. Lei, D. Yi, and S. Z. Li, “A face anti- spoofing database with diverse attacks,” in 2012 5th IAPR International Conference on Biometrics (ICB) . IEEE, Mar. 2012

    work page 2012

  6. [6]

    On the effectiveness of local binary patterns in face anti-spoofing,

    I. Chingovska, A. Anjos, and S. Marcel, “On the effectiveness of local binary patterns in face anti-spoofing,” inIEEE BIOSIG 2012, September 2012

    work page 2012

  7. [7]

    Learning Deep Models for Face Anti-Spoofing: Binary or Auxiliary Supervision

    Y . Liu, A. Jourabloo, and X. Liu, “Learning deep models for face anti- spoofing: Binary or auxiliary supervision,” CoRR, vol. abs/1803.11097, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  8. [8]

    How far did we get in face spoofing detection?

    L. Souza, L. Oliveira, M. Pamplona, and J. Papa, “How far did we get in face spoofing detection?” Eng. Appl. Artif. Intell. , vol. 72, no. C, pp. 368–381, Jun. 2018

    work page 2018

  9. [9]

    A neural algorithm of artistic style,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “A neural algorithm of artistic style,” arXiv preprint, 2015

    work page 2015

  10. [10]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing Systems 27 . Curran Associates, Inc., 2014, pp. 2672–2680

    work page 2014

  11. [11]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2019, pp. 4401–4410

    work page 2019

  12. [12]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2223–2232

    work page 2017

  13. [13]

    Unsupervised representation learning with deep convolutional generative adversarial networks,

    A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv preprint, 2015

    work page 2015

  14. [14]

    From attributes to faces: a conditional generative network for face generation,

    Y . Wang, A. Dantcheva, and F. Bremond, “From attributes to faces: a conditional generative network for face generation,” in 2018 Interna- tional Conference of the Biometrics Special Interest Group (BIOSIG) . IEEE, 2018, pp. 1–5

    work page 2018

  15. [15]

    Face generation for low-shot learning using generative adversarial networks,

    J. Choe, S. Park, K. Kim, J. Hyun Park, D. Kim, and H. Shim, “Face generation for low-shot learning using generative adversarial networks,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 1940–1948

    work page 2017

  16. [16]

    3d-aided dual- agent gans for unconstrained face recognition,

    J. Zhao, L. Xiong, J. Li, J. Xing, S. Yan, and J. Feng, “3d-aided dual- agent gans for unconstrained face recognition,” IEEE transactions on pattern analysis and machine intelligence , 2018

    work page 2018

  17. [17]

    Face de-spoofing: Anti-spoofing via noise modeling,

    A. Jourabloo, Y . Liu, and X. Liu, “Face de-spoofing: Anti-spoofing via noise modeling,” in The European Conference on Computer Vision (ECCV), September 2018

    work page 2018

  18. [18]

    OULU- NPU: A mobile face presentation attack database with real-world variations,

    Z. Boulkenafet, J. Komulainen, L. Li, X. Feng, and A. Hadid, “OULU- NPU: A mobile face presentation attack database with real-world variations,” in IEEE International Conference on Automatic Face and Gesture Recognition, May 2017

    work page 2017

  19. [19]

    Face liveness detection using convolutional-features fusion of real and deep network generated face images,

    Y . A. U. Rehman, L.-M. Po, M. Liu, Z. Zou, W. Ou, and Y . Zhao, “Face liveness detection using convolutional-features fusion of real and deep network generated face images,” Journal of Visual Communication and Image Representation, vol. 59, pp. 574–582, Feb. 2019

    work page 2019

  20. [20]

    Exploiting temporal and depth information for multi-frame face anti-spoofing

    Z. Wang, C. Zhao, Y . Qin, Q. Zhou, and Z. Lei, “Exploiting temporal and depth information for multi-frame face anti-spoofing,” CoRR, vol. abs/1811.05118, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  21. [21]

    A Dataset and Benchmark for Large-scale Multi-modal Face Anti-spoofing

    S. Zhang, X. Wang, A. Liu, C. Zhao, J. Wan, S. Escalera, H. Shi, Z. Wang, and S. Z. Li, “CASIA-SURF: A dataset and benchmark for large-scale multi-modal face anti-spoofing,”CoRR, vol. abs/1812.00408, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    Meta anti-spoofing: Learning to learn in face anti-spoofing,

    C. Zhao, Y . Qin, Z. Wang, T. Fu, and H. Shi, “Meta anti-spoofing: Learning to learn in face anti-spoofing,” CoRR, vol. abs/1904.12490, 2019

    work page arXiv 1904

  23. [23]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” inEuropean conference on computer vision. Springer, 2016, pp. 694–711

    work page 2016

  24. [24]

    Instance normalization: The missing ingredient for fast stylization,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Instance normalization: The missing ingredient for fast stylization,” arXiv preprint, 2016

    work page 2016

  25. [25]

    Dlib-ml: A machine learning toolkit,

    D. E. King, “Dlib-ml: A machine learning toolkit,” Journal of Machine Learning Research, vol. 10, pp. 1755–1758, 2009

    work page 2009

  26. [26]

    Fast style transfer,

    L. Engstrom, “Fast style transfer,” https://github.com/lengstrom/ fast-style-transfer/, 2016

    work page 2016

  27. [27]

    Very deep convolutional neural network based image classification using small training sample size,

    S. Liu and W. Deng, “Very deep convolutional neural network based image classification using small training sample size,” in 2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR) . IEEE, Nov. 2015

    work page 2015

  28. [28]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 4510–4520

    work page 2018

  29. [29]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition . IEEE, Jun. 2009

    work page 2009

  30. [30]

    Information technology Biometric presentation attack detection Part 3: Testing and reporting,

    “Information technology Biometric presentation attack detection Part 3: Testing and reporting,” Tech. Rep., Mar. 2017

    work page 2017