pith. sign in

arxiv: 2606.18510 · v1 · pith:CYKQ47B4new · submitted 2026-06-16 · 💻 cs.CV · cs.CR

Architectural Bias in Face Presentation Attack Detection: A Comparative Study of Vision Transformers and Convolutional Neural Networks

Ngela Landon Ntung , Floride Tuyisenge , Jema David Ndibwile This is my paper

Pith reviewed 2026-06-27 00:43 UTC · model grok-4.3

classification 💻 cs.CV cs.CR
keywords presentation attack detectionvision transformerconvolutional neural networkdemographic biasfairnesszero-shot generalizationCeFA dataset
0
0 comments X

The pith

Pretrained Vision Transformers reduce demographic performance gaps in face presentation attack detection compared to CNNs.

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

The paper tests whether Vision Transformer architectures can lessen systematic performance disparities in face presentation attack detection that currently affect darker skin tones more severely. It evaluates a pretrained DeiT-S, a scratch-trained Multimodal ViT-Tiny, and a ResNet18 CNN baseline on the CASIA-SURF CeFA dataset across African, East Asian, and zero-shot Central Asian groups. The pretrained transformer reaches 97.27 percent accuracy and 0.86 percent EER while shrinking the inter-ethnic ACER gap to 0.13 percent and cutting the zero-shot BPCER on unseen subjects to 2.89 percent versus 10.44 percent for the CNN. A sympathetic reader cares because biometric systems that perform unevenly across populations create exclusion in security-critical applications. The findings indicate that architecture itself can be adjusted to improve equity without separate fairness interventions.

Core claim

The paper establishes that a pretrained DeiT-S Vision Transformer fine-tuned on the CASIA-SURF CeFA dataset attains 97.27 percent accuracy and 0.86 percent EER, outperforming ResNet18 at 90.15 percent accuracy. It reduces the inter-ethnic ACER gap between African and East Asian subjects to 0.13 percent, an 83 percent reduction relative to an LBP baseline. On zero-shot Central Asian subjects the model records a BPCER of 2.89 percent against 10.44 percent for ResNet18, a 3.6 times generalization advantage. These outcomes indicate that pretrained Vision Transformers deliver higher PAD accuracy, smaller demographic gaps, and more equitable performance on unseen groups, showing that cross-demogra

What carries the argument

Side-by-side comparison of pretrained DeiT-S Vision Transformer against ResNet18 CNN on the multimodal CASIA-SURF CeFA dataset, using accuracy, EER, ACER ethnic gaps, and zero-shot BPCER as the evaluation measures.

If this is right

  • Pretrained Vision Transformers can reach 97.27 percent accuracy in PAD while limiting inter-ethnic ACER gaps to 0.13 percent.
  • An 83 percent reduction in ethnic performance gaps relative to LBP methods becomes achievable through architectural replacement.
  • Zero-shot generalization to unseen demographic groups improves by a factor of 3.6 in BPCER reduction.
  • Fairness across African, East Asian, and Central Asian groups can be advanced by preferring pretrained transformer backbones over standard CNNs.

Where Pith is reading between the lines

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

  • The fairness gains may trace more to large-scale pretraining than to the transformer structure itself, given that the scratch-trained ViT-Tiny receives less emphasis in the reported results.
  • The same comparative protocol could be applied to other PAD datasets to test whether the pattern of reduced gaps generalizes beyond CeFA.
  • System builders facing fairness requirements might default to pretrained transformer backbones for the PAD component to obtain equity as a byproduct of architecture.

Load-bearing premise

The measured differences in accuracy and demographic fairness between the Vision Transformer and CNN models arise from their architectures rather than from unequal training procedures, augmentations, or hyperparameter settings.

What would settle it

A re-run of both DeiT-S and ResNet18 on the CeFA dataset with identical training schedules, augmentations, and hyperparameters that still shows comparable demographic gaps would falsify the claim that architecture drives the fairness improvement.

Figures

Figures reproduced from arXiv: 2606.18510 by Floride Tuyisenge, Jema David Ndibwile, Ngela Landon Ntung.

Figure 2
Figure 2. Figure 2: Fairness Improvement in ACER Across Ethnic Groups: ViT-Tiny vs [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: ROC Curve for DeiT-S on CASIA-SURF CeFA Dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: ResNet18 CNN Performance Across Seen and Zero-Shot Groups. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ResNet18 CNN Confusion Matrix by Ethnic Group Showing Zero [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

Face Presentation Attack Detection (PAD) systems constitute a critical security layer in biometric authentication; however, existing approaches exhibit systematic performance disparities across demographic groups, disproportionately affecting individuals with darker skin tones. This paper presents a comparative empirical investigation of whether Vision Transformer architectures reduce demographic bias in face PAD systems relative to convolutional baselines. Experiments are conducted on the CASIA-SURF Cross-Ethnicity Face Anti-Spoofing (CeFA) dataset. Three architectures are evaluated: a Multimodal ViT-Tiny trained from scratch, a ResNet18 CNN baseline, and a pretrained DeiT-S fine-tuned on CeFA across African, East Asian, and zero-shot Central Asian demographic groups. DeiT-S achieves the highest overall accuracy of 97.27% and the lowest EER of 0.86%, outperforming ResNet18 at 90.15% accuracy. In terms of fairness, DeiT-S reduces the inter-ethnic ACER gap between African and East Asian subjects to 0.13%, compared to 0.75% reported in an LBP-based work [6], representing an 83% reduction. Most notably, while ResNet18 records a BPCER of 10.44% on zero-shot Central Asian subjects, DeiT-S maintains 2.89% on the same unseen group, demonstrating a 3.6x generalization advantage. These results suggest that pretrained Vision Transformers achieve superior PAD accuracy, produce smaller demographic performance gaps, and generalize more equitably across unseen demographic groups, indicating that cross-demographic fairness in PAD may partly be influenced by architectural design.

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 / 2 minor

Summary. The paper conducts an empirical comparison of Vision Transformer (ViT) and CNN architectures for face Presentation Attack Detection (PAD) on the CASIA-SURF CeFA dataset. It evaluates a scratch-trained Multimodal ViT-Tiny, a ResNet18 baseline, and a pretrained DeiT-S, reporting that DeiT-S achieves 97.27% accuracy and 0.86% EER (vs. 90.15% for ResNet18), reduces the inter-ethnic ACER gap to 0.13% (83% smaller than prior LBP work), and shows a 3.6× lower BPCER (2.89% vs. 10.44%) on zero-shot Central Asian subjects, concluding that architectural design influences cross-demographic fairness.

Significance. If the fairness and generalization advantages can be isolated to the Vision Transformer inductive bias, the work would provide evidence that architecture choice can mitigate demographic disparities in PAD systems, which is relevant for equitable biometric security. The use of a cross-ethnicity dataset and concrete metrics (accuracy, EER, ACER, BPCER) across African, East Asian, and zero-shot groups offers a direct empirical test of the claim.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'cross-demographic fairness in PAD may partly be influenced by architectural design' is undermined because DeiT-S starts from large-scale pretraining while the ResNet18 baseline and ViT-Tiny do not; no details are given that the CNN received equivalent pretraining, data augmentation, or hyperparameter tuning, so the reported 3.6× BPCER reduction on zero-shot Central Asian subjects and 83% ACER gap reduction cannot be attributed to architecture rather than training procedure differences.
  2. [Abstract] Abstract: The reported performance differences (e.g., 97.27% vs. 90.15% accuracy, 0.13% vs. 0.75% ACER gap) lack error bars, statistical significance tests, or ablation on training procedure, making it impossible to assess whether the fairness gains are robust or load-bearing for the architectural-bias conclusion.
minor comments (2)
  1. [Abstract] Abstract: The comparison to the LBP-based work [6] for the 83% gap reduction should specify whether the same dataset splits, protocols, and demographic groupings were used.
  2. [Abstract] Abstract: Full training details (optimizer, schedule, augmentation, hyperparameters) for all three models are needed to allow reproduction and isolation of the architecture effect.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments correctly identify limitations in how the experimental comparisons isolate architectural effects and in the statistical robustness of the reported results. We address each point below and commit to revisions that clarify these issues without overstating the current evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'cross-demographic fairness in PAD may partly be influenced by architectural design' is undermined because DeiT-S starts from large-scale pretraining while the ResNet18 baseline and ViT-Tiny do not; no details are given that the CNN received equivalent pretraining, data augmentation, or hyperparameter tuning, so the reported 3.6× BPCER reduction on zero-shot Central Asian subjects and 83% ACER gap reduction cannot be attributed to architecture rather than training procedure differences.

    Authors: We agree that the pretraining regime for DeiT-S constitutes a confounding factor that prevents strong attribution of the fairness and generalization gains solely to architectural inductive bias. The manuscript already distinguishes the training setups (pretrained DeiT-S versus scratch-trained ViT-Tiny and ResNet18), but the abstract and discussion overstate the architectural claim. We will revise the abstract, introduction, and conclusion to state that the results show advantages for a pretrained Vision Transformer over standard CNN and scratch ViT baselines under typical training protocols. We will also expand the methods section with full hyperparameter, augmentation, and optimization details for all three models and add an explicit limitations paragraph noting that matched-pretraining ablations are required to isolate architecture. revision: yes

  2. Referee: [Abstract] Abstract: The reported performance differences (e.g., 97.27% vs. 90.15% accuracy, 0.13% vs. 0.75% ACER gap) lack error bars, statistical significance tests, or ablation on training procedure, making it impossible to assess whether the fairness gains are robust or load-bearing for the architectural-bias conclusion.

    Authors: We concur that the absence of error bars, significance testing, and training-procedure ablations weakens the evidential support for the conclusions. In the revised version we will (1) report means and standard deviations over at least three independent training runs for all key metrics, (2) include paired statistical tests (e.g., McNemar or t-tests) on the accuracy, EER, and ACER differences, and (3) add an ablation that fine-tunes a pretrained ResNet18 on the same CeFA splits to better separate pretraining effects from architecture. These additions will be placed in the experimental results and discussion sections. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical comparison without derivations or self-referential fitting

full rationale

The paper conducts an experimental comparison of three architectures (ViT-Tiny trained from scratch, ResNet18, pretrained DeiT-S) on the fixed CASIA-SURF CeFA dataset, reporting accuracy, EER, ACER gaps, and BPCER on demographic subgroups. No equations, parameter fitting to subsets followed by 'prediction' of related quantities, or self-citation chains are present. All claims reduce directly to measured performance numbers on held-out test splits rather than to any input by construction. Minor self-citation risk is absent; the single cited work [6] is external and used only for baseline comparison. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the CeFA dataset splits and evaluation protocol isolate architectural effects from training and data factors; no invented entities or fitted constants beyond standard ML training are introduced.

axioms (1)
  • domain assumption The CASIA-SURF CeFA dataset provides representative samples across African, East Asian, and Central Asian demographic groups for measuring bias.
    Invoked when reporting inter-ethnic ACER gaps and zero-shot results on Central Asian subjects.

pith-pipeline@v0.9.1-grok · 5835 in / 1293 out tokens · 35711 ms · 2026-06-27T00:43:09.275955+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

23 extracted references · 13 canonical work pages · 1 internal anchor

  1. [1]

    Introduc- tion to Presentation Attack Detection in Face Biometrics and Recent Advances,

    J. Hernandez-Ortega, J. Fierrez, A. Morales, and J. Galbally, “Introduc- tion to Presentation Attack Detection in Face Biometrics and Recent Advances,” arXiv:2111.11794, Nov. 2021

    work page arXiv 2021

  2. [2]

    Presentation Attack Detection Methods for Face Recognition Systems: A Comprehensive Survey,

    R. Ramachandra and C. Busch, “Presentation Attack Detection Methods for Face Recognition Systems: A Comprehensive Survey,”ACM Comput. Surv., vol. 50, no. 1, pp. 1–37, Jan. 2018

    2018

  3. [3]

    Review of Demographic Fairness in Face Recognition,

    K. Kotwal and S. Marcel, “Review of Demographic Fairness in Face Recognition,” arXiv:2502.02309, 2025

    work page arXiv 2025

  4. [4]

    Issues Related to Face Recognition Accuracy Varying Based on Race and Skin Tone,

    K. S. Krishnapriya, V . Albiero, K. Vangara, M. C. King, and K. W. Bowyer, “Issues Related to Face Recognition Accuracy Varying Based on Race and Skin Tone,”IEEE Trans. Technol. Soc., vol. 1, no. 1, pp. 8–20, Mar. 2020

    2020

  5. [5]

    Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,

    J. Buolamwini and T. Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” inProc. Conf. Fair- ness, Accountability and Transparency (FAT*), 2018

    2018

  6. [6]

    Fairness-Aware Face Presentation Attack Detection Using Local Binary Patterns: Bridging Skin Tone Bias in Biometric Systems,

    J. D. Ndibwile, N. N. Landon, and F. Tuyisenge, “Fairness-Aware Face Presentation Attack Detection Using Local Binary Patterns: Bridging Skin Tone Bias in Biometric Systems,”J. Cybersecurity Priv., vol. 6, no. 1, p. 12, Jan. 2026

    2026

  7. [7]

    CASIA-SURF CeFA: A Benchmark for Multi-modal Cross- ethnicity Face Anti-spoofing,

    A. Liet al., “CASIA-SURF CeFA: A Benchmark for Multi-modal Cross- ethnicity Face Anti-spoofing,” arXiv:2003.05136, Mar. 2020

    work page arXiv 2003

  8. [8]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    A. Dosovitskiyet al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv:2010.11929, Jun. 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  9. [9]

    Training Data-Efficient Image Transformers & Distillation through Attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J ´egou, “Training Data-Efficient Image Transformers & Distillation through Attention,” arXiv:2012.12877, 2020

    work page arXiv 2012

  10. [10]

    ImageNet-trained CNNs are biased towards texture; increasing shape bias improves robustness and human relevance

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel, “ImageNet-Trained CNNs Are Biased Towards Tex- ture; Increasing Shape Bias Improves Accuracy and Robustness,” arXiv:1811.12231, Nov. 2022

    work page arXiv 2022

  11. [11]

    Intriguing Properties of Vision Transformers,

    M. Naseer, K. Ranasinghe, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Intriguing Properties of Vision Transformers,” arXiv:2105.10497, Nov. 2021

    work page arXiv 2021

  12. [12]

    Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition,

    S. Dooley, R. S. Sukthanker, J. P. Dickerson, C. White, F. Hutter, and M. Goldblum, “Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition,” arXiv:2210.09943, Dec. 2023

    work page arXiv 2023

  13. [13]

    Faces of Fairness: Examining Bias in Facial Expression Recognition Datasets and Models,

    M. M. Hosseini, A. P. Fard, and M. H. Mahoor, “Faces of Fairness: Examining Bias in Facial Expression Recognition Datasets and Models,” arXiv:2502.11049, Oct. 2025

    work page arXiv 2025

  14. [14]

    Face Spoofing Detection from Single Images Using Micro-Texture Analysis,

    J. Maatta, A. Hadid, and M. Pietikainen, “Face Spoofing Detection from Single Images Using Micro-Texture Analysis,” inProc. Int. Joint Conf. Biometrics (IJCB), Washington, DC, USA, Oct. 2011, pp. 1–7

    2011

  15. [15]

    Face Anti-Spoofing Detection with Multi- Modal CNN Enhanced by ResNet,

    H. Shaker and S. Al-Darraji, “Face Anti-Spoofing Detection with Multi- Modal CNN Enhanced by ResNet,”Basrah Res. Sci., vol. 50, no. 1, p. 12, Jun. 2024

    2024

  16. [16]

    Can Your Face Detector Do Anti-Spoofing? Face Presentation Attack Detection with a Multi-Channel Face Detec- tor,

    A. George and S. Marcel, “Can Your Face Detector Do Anti-Spoofing? Face Presentation Attack Detection with a Multi-Channel Face Detec- tor,” arXiv:2006.16836, Jul. 2020

    work page arXiv 2006

  17. [17]

    Using Infrared to Improve Face Recognition of Individuals with Highly Pigmented Skin,

    A. G. Muthua, R. P. Theart, and M. J. Booysen, “Using Infrared to Improve Face Recognition of Individuals with Highly Pigmented Skin,” iScience, vol. 26, no. 7, p. 107039, 2023

    2023

  18. [18]

    Fairness in Face Presentation Attack Detection,

    M. Fang, W. Yang, A. Kuijper, V . Struc, and N. Damer, “Fairness in Face Presentation Attack Detection,” arXiv:2209.09035, 2022

    work page arXiv 2022

  19. [19]

    On the Effectiveness of Vision Transformers for Zero-Shot Face Anti-Spoofing,

    A. George and S. Marcel, “On the Effectiveness of Vision Transformers for Zero-Shot Face Anti-Spoofing,” inProc. IEEE Int. Joint Conf. Biometrics (IJCB), Shenzhen, China, Aug. 2021, pp. 1–8

    2021

  20. [20]

    S-Adapter: Generalizing Vision Transformer for Face Anti-Spoofing with Statistical Tokens,

    R. Caiet al., “S-Adapter: Generalizing Vision Transformer for Face Anti-Spoofing with Statistical Tokens,” arXiv:2309.04038, Jun. 2024

    work page arXiv 2024

  21. [21]

    FM-ViT: Flexible Modal Vision Transformers for Face Anti-Spoofing,

    A. Liuet al., “FM-ViT: Flexible Modal Vision Transformers for Face Anti-Spoofing,”IEEE Trans. Inf. Forensics Secur., vol. 18, pp. 4775– 4786, 2023

    2023

  22. [22]

    Robust Face Anti-Spoofing Framework with Convolutional Vision Transformer,

    Y . Lee, Y . Kwak, and J. Shin, “Robust Face Anti-Spoofing Framework with Convolutional Vision Transformer,” arXiv:2307.12459, Jul. 2023

    work page arXiv 2023

  23. [23]

    The Performance Analysis of Facial Expression Recognition System Using Local Regions and Features,

    Y . Yang, B. Vuksanovic, and H. Ma, “The Performance Analysis of Facial Expression Recognition System Using Local Regions and Features,”Journal of Image and Graphics, vol. 11, no. 2, pp. 104–114, Jun. 2023

    2023