pith. sign in

arxiv: 1907.04048 · v1 · pith:DTIHDDQXnew · submitted 2019-07-09 · 💻 cs.CV

Domain Adaptation in Multi-Channel Autoencoder based Features for Robust Face Anti-Spoofing

Olegs Nikisins , Anjith George , Sebastien Marcel This is my paper

Pith reviewed 2026-05-25 00:41 UTC · model grok-4.3

classification 💻 cs.CV
keywords face anti-spoofingpresentation attack detectiondomain adaptationmulti-channelautoencoderfacial regionsRGBNIR
0
0 comments X

The pith

Domain adaptation from RGB transfers facial knowledge to multi-channel data, letting autoencoders detect spoofs better when features come from separate facial regions rather than the whole face.

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

The paper aims to strengthen face presentation attack detection by using multi-channel sensors and a domain-adapted autoencoder plus MLP pipeline. Instead of gathering large new labeled sets in the target domain, it adapts existing RGB face knowledge to the RGB plus depth plus NIR setting. A central result is that learning features region by region yields more discriminative representations than processing the entire face at once. The method is evaluated on a recent public multi-channel database that includes many different attack types.

Core claim

Domain adaptation allows an autoencoder-based system trained primarily on RGB face images to produce effective features in the multi-channel domain, and region-specific feature extraction outperforms whole-face processing for distinguishing live faces from presentation attacks.

What carries the argument

Domain-adapted autoencoder features learned separately on individual facial regions in multi-channel (RGB, Depth, NIR) images.

If this is right

  • Multi-channel input improves robustness to sophisticated presentation attacks relative to RGB-only systems.
  • Domain adaptation reduces reliance on large quantities of labeled multi-channel training data.
  • Region-wise feature learning produces more discriminative PAD cues than global face features.
  • The resulting system handles diverse attack instruments present in the test database.

Where Pith is reading between the lines

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

  • The same adaptation strategy could apply to other sensor combinations where labeled attack data remains scarce.
  • Emphasis on local regions suggests future detectors may benefit from prioritizing part-based rather than holistic analysis.
  • Devices already equipped with depth and NIR sensors could gain improved spoof resistance by applying this transfer approach.

Load-bearing premise

Domain adaptation can move useful facial appearance knowledge from the RGB domain to the multi-channel domain without needing large amounts of new labeled multi-channel data.

What would settle it

Training the same multi-channel autoencoder directly on the target database without any domain adaptation step and obtaining equal or superior detection rates on the held-out attacks would undermine the claimed benefit of the adaptation.

Figures

Figures reproduced from arXiv: 1907.04048 by Anjith George, Olegs Nikisins, Sebastien Marcel.

Figure 1
Figure 1. Figure 1: Schematic representation of the proposed MC face PAD approach: visualizing the preprocessing stage, internal structure of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of bona-fide data in 6 different sessions, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: DET curves for baselines. For the evaluation set of WMCA, grandtest-10 protocol. values are reported on the evaluation set given threshold. In addition to numerical performance, the DET curves are given for all experiments. 4.1. Results: MC face PAD baselines In this subsection a baseline RGB-NIR-D face PAD sys￾tem is discussed and evaluated, being composed of success￾ful, channel specific, hand-crafted fe… view at source ↗
Figure 4
Figure 4. Figure 4: DET curves for PAD system using AE for the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: DET curves for PAD system using 9 AE for MC fa [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: DET curves for PAD system using 16 AE for MC fa [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: DET curves for proposed RGB PAD system (using 9 AE [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
read the original abstract

While the performance of face recognition systems has improved significantly in the last decade, they are proved to be highly vulnerable to presentation attacks (spoofing). Most of the research in the field of face presentation attack detection (PAD), was focused on boosting the performance of the systems within a single database. Face PAD datasets are usually captured with RGB cameras, and have very limited number of both bona-fide samples and presentation attack instruments. Training face PAD systems on such data leads to poor performance, even in the closed-set scenario, especially when sophisticated attacks are involved. We explore two paths to boost the performance of the face PAD system against challenging attacks. First, by using multi-channel (RGB, Depth and NIR) data, which is still easily accessible in a number of mass production devices. Second, we develop a novel Autoencoders + MLP based face PAD algorithm. Moreover, instead of collecting more data for training of the proposed deep architecture, the domain adaptation technique is proposed, transferring the knowledge of facial appearance from RGB to multi-channel domain. We also demonstrate, that learning the features of individual facial regions, is more discriminative than the features learned from an entire face. The proposed system is tested on a very recent publicly available multi-channel PAD database with a wide variety of presentation attacks.

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 presentation attack detection (PAD) system that leverages multi-channel (RGB + Depth + NIR) input and an autoencoder + MLP architecture. It introduces domain adaptation to transfer facial appearance knowledge from labeled RGB source data to the multi-channel target domain without requiring substantial new labeled multi-channel training data. The work also claims that features learned from individual facial regions are more discriminative than those from an entire face and evaluates the system on a recent public multi-channel PAD database containing a wide variety of presentation attacks.

Significance. If the domain adaptation step successfully aligns distributions and yields measurable gains on challenging attacks attributable to the adaptation rather than the multi-channel input alone, the approach would address a key practical bottleneck in PAD research: scarcity of labeled multi-channel data. The regional-feature finding, if robustly demonstrated, would also offer a concrete design principle for future multi-channel PAD systems. The use of publicly available multi-channel data is a positive step toward reproducibility.

major comments (2)
  1. [Abstract] Abstract and domain-adaptation description: the headline claim that domain adaptation transfers RGB knowledge to the multi-channel domain 'without requiring substantial new labeled multi-channel training data' is load-bearing for the performance improvement, yet the text supplies no quantitative results, ablation tables, or adaptation-loss values that isolate the contribution of the adaptation loss from the simple addition of Depth/NIR channels.
  2. [Domain adaptation description] Domain adaptation section (implied by the description of the proposed technique): the assumption that RGB and (RGB+Depth+NIR) distributions are sufficiently close for the adaptation loss to align them without target labels is not secured by any reported metric (e.g., feature-space distance before/after adaptation or cross-domain classification accuracy), leaving open the possibility that observed gains stem from multi-channel input or regional cropping rather than adaptation.
minor comments (1)
  1. [Abstract] The abstract states the evaluation plan but does not report any numerical performance figures, error bars, or dataset split details; these should be added to the abstract or a results table for immediate assessment.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We agree that the current version does not provide explicit quantitative isolation of the domain-adaptation contribution and will revise the manuscript to include the requested ablations and metrics.

read point-by-point responses

  1. Referee: [Abstract] Abstract and domain-adaptation description: the headline claim that domain adaptation transfers RGB knowledge to the multi-channel domain 'without requiring substantial new labeled multi-channel training data' is load-bearing for the performance improvement, yet the text supplies no quantitative results, ablation tables, or adaptation-loss values that isolate the contribution of the adaptation loss from the simple addition of Depth/NIR channels.

    Authors: We agree that the manuscript as submitted does not contain ablation tables or adaptation-loss curves that separate the effect of the domain-adaptation term from the simple use of Depth and NIR channels. In the revised version we will add (i) performance tables with and without the adaptation loss, (ii) plots of the adaptation loss during training, and (iii) a direct comparison against a multi-channel baseline trained without adaptation. These additions will make the contribution of domain adaptation explicit. revision: yes

  2. Referee: [Domain adaptation description] Domain adaptation section (implied by the description of the proposed technique): the assumption that RGB and (RGB+Depth+NIR) distributions are sufficiently close for the adaptation loss to align them without target labels is not secured by any reported metric (e.g., feature-space distance before/after adaptation or cross-domain classification accuracy), leaving open the possibility that observed gains stem from multi-channel input or regional cropping rather than adaptation.

    Authors: We acknowledge that the paper currently offers no distributional-alignment diagnostics. In the revision we will report (a) maximum mean discrepancy (MMD) between source and target feature distributions before and after adaptation and (b) a cross-domain classification experiment in which a classifier trained on adapted RGB features is evaluated on the multi-channel target. These metrics will directly address whether the adaptation loss produces measurable alignment. revision: yes

Circularity Check

0 steps flagged

No circularity; standard techniques applied to new data regime

full rationale

The paper describes a multi-channel autoencoder + MLP architecture with domain adaptation from RGB to RGB+Depth+NIR, plus regional feature learning, evaluated on an external multi-channel PAD database. No equations or claims in the provided text reduce a prediction to a fitted parameter by construction, nor does any load-bearing premise rest on a self-citation chain that itself lacks independent verification. The derivation chain is self-contained against external benchmarks and does not exhibit self-definitional, fitted-input, or uniqueness-imported circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available; therefore the ledger records only the high-level modeling assumptions stated or implied there. No explicit free parameters, new entities, or ad-hoc axioms are enumerated in the text.

axioms (2)
  • domain assumption Autoencoder features extracted from individual facial regions are more discriminative for PAD than whole-face features.
    Stated directly in the abstract as a demonstration claim.
  • domain assumption Domain adaptation can transfer useful facial appearance knowledge from RGB to Depth+NIR channels.
    Central premise of the proposed training strategy.

pith-pipeline@v0.9.0 · 5768 in / 1249 out tokens · 21910 ms · 2026-05-25T00:41:40.749297+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

27 extracted references · 27 canonical work pages · 2 internal anchors

  1. [1]

    Information technology - Biometric presentation attack de- tection - Part 3: Testing and reporting, 2017. 4

    work page 2017

  2. [2]

    S. R. Arashloo and J. Kittler. An anomaly detection approach to face spoofing detection: A new formulation and evaluation protocol. In 2017 IEEE International Joint Conference on Biometrics (IJCB), pages 80–89, Oct 2017. 1

    work page 2017

  3. [3]

    Bhattacharjee and S

    S. Bhattacharjee and S. Marcel. What you can’t see can help you - extended-range imaging for 3d-mask presentation at- tack detection. In 2017 International Conference of the Bio- metrics Special Interest Group, pages 1–7, Sept 2017. 1

    work page 2017

  4. [4]

    Bhattacharjee, A

    S. Bhattacharjee, A. Mohammadi, and S. Marcel. Spoofing deep face recognition with custom silicone masks. In Pro- ceedings of BTAS2018, Oct. 2018. 1

    work page 2018

  5. [5]

    Detecting Anomalous Faces with 'No Peeking' Autoencoders

    A. Bhattad, J. Rock, and D. A. Forsyth. Detecting anomalous faces with ’no peeking’ autoencoders. CoRR, abs/1802.05798, 2018. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    Boulkenafet, J

    Z. Boulkenafet, J. Komulainen, L. Li, X. Feng, and A. Hadid. OULU-NPU: A mobile face presentation attack database with real-world variations. May 2017. 1, 2

    work page 2017

  7. [7]

    Chingovska, A

    I. Chingovska, A. Anjos, and S. Marcel. On the effectiveness of local binary patterns in face anti-spoofing. 2012. 2

    work page 2012

  8. [8]

    Costa-Pazo, S

    A. Costa-Pazo, S. Bhattacharjee, E. Vazquez-Fernandez, and S. Marcel. The replay-mobile face presentation-attack database. In Proceedings of the International Conference on Biometrics Special Interests Group (BioSIG), Sept. 2016. 1

    work page 2016

  9. [9]

    de Freitas Pereira, A

    T. de Freitas Pereira, A. Anjos, and S. Marcel. Het- erogeneous face recognition using domain specific units. IEEE Transactions on Information Forensics and Security , page 13, Feb. 2019. 4

    work page 2019

  10. [10]

    G. B. de Souza, J. P. Papa, and A. N. Marana. On the learn- ing of deep local features for robust face spoofing detection. CoRR, abs/1806.07492, 2018. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  11. [11]

    S. N. Garg, R. Vig, and S. Gupta. A survey on different levels of fusion in multimodal biometrics. Indian Journal of Science and Technology, 10(44), 2017. 3

    work page 2017

  12. [12]

    George, Z

    A. George, Z. Mostaani, D. Geissenbuhler, O. Nikisins, A. Anjos, and S. Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security (under review), 2019. 4

    work page 2019

  13. [13]

    L. Li, X. Feng, Z. Boulkenafet, Z. Xia, M. Li, and A. Ha- did. An original face anti-spoofing approach using partial convolutional neural network. In 2016 Sixth International Conference on Image Processing Theory, Tools and Applica- tions (IPTA), pages 1–6, Dec 2016. 2

    work page 2016

  14. [14]

    Y . Liu*, A. Jourabloo*, and X. Liu. Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In In Proceeding of IEEE Computer Vision and Pattern Recogni- tion, Salt Lake City, UT, June 2018. 1, 2

    work page 2018

  15. [15]

    Y . Liu, A. Jourabloo, and X. Liu. Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In 2018 IEEE Conference on Computer Vision and Pattern Recogni- tion, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 389–398, 2018. 2

    work page 2018

  16. [16]

    Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. In Proceedings of International Con- ference on Computer Vision (ICCV), 2015. 3

    work page 2015

  17. [17]

    Mohammadi, S

    A. Mohammadi, S. Bhattacharjee, and S. Marcel. Deeply vulnerable – a study of the robustness of face recognition to presentation attacks. IET (The Institution of Engineering and Technology) – Biometrics, pages 1–13, 2017. Accepted on 29-Sept-2017. 1

    work page 2017

  18. [18]

    Nikisins, A

    O. Nikisins, A. Mohammadi, A. Anjos, and S. Marcel. On effectiveness of anomaly detection approaches against un- seen presentation attacks in face anti-spoofing. In 2018 In- ternational Conference on Biometrics (ICB), 2018. 1, 5

    work page 2018

  19. [19]

    O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep face recognition. In British Machine Vision Conference. 2

    work page

  20. [20]

    Raghavendra, K

    R. Raghavendra, K. B. Raja, S. Venkatesh, and C. Busch. Ex- tended multispectral face presentation attack detection: An approach based on fusing information from individual spec- tral bands. In 2017 20th International Conference on Infor- mation Fusion (Fusion), pages 1–6, July 2017. 2

    work page 2017

  21. [21]

    Ramachandra and C

    R. Ramachandra and C. Busch. Presentation attack detec- tion methods for face recognition systems: A comprehensive survey. ACM Comput. Surv., 50(1):8:1–8:37, Mar. 2017. 2

    work page 2017

  22. [22]

    M. O. Simn, C. Corneanu, K. Nasrollahi, O. Nikisins, S. Es- calera, Y . Sun, H. Li, Z. Sun, T. B. Moeslund, and M. Grei- tans. Improved rgb-d-t based face recognition. IET Biomet- rics, 5(4):297–303, 2016. 1

    work page 2016

  23. [23]

    Steiner, S

    H. Steiner, S. Sporrer, A. Kolb, and N. Jung. Design of an active multispectral SWIR camera system for skin de- tection and face verification. J. Sensors, 2016:9682453:1– 9682453:16, 2016. 1

    work page 2016

  24. [24]

    D. Wen, H. Han, and A. Jain. Face Spoof Detection with Im- age Distortion Analysis. IEEE Trans. Information Forensic and Security, 10(4):746–761, April 2015. 1

    work page 2015

  25. [25]

    Xiong and W

    F. Xiong and W. Abdalmageed. Unknown presentation at- tack detection with face rgb images. In IEEE International Conference on Biometrics: Theory, Applications, and Sys- tems, 2018. 2, 5

    work page 2018

  26. [26]

    Zhang, Z

    K. Zhang, Z. Zhang, Z. Li, and Y . Qiao. Joint face detec- tion and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10), 2016. 3

    work page 2016

  27. [27]

    Zhang, J

    Z. Zhang, J. Yan, S. Liu, Z. Lei, D. Yi, and S. Z. Li. A face antispoofing database with diverse attacks. In2012 5th IAPR International Conference on Biometrics (ICB), pages 26–31, March 2012. 2

    work page 2012