Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Group-wise Scaling and Orthogonal Decomposition for Domain-Invariant Feature Extraction in Face Anti-Spoofing

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that aligning classifier biases as well as weights across domains lets face-anti-spoofing models generalize to unseen domains, reporting a 1.83 percent average half total error rate.

desk verdict Useful combination of two existing ideas with strong reported numbers, but the paper's own ECE ablation undercuts the central bias-alignment claim and the evaluation protocol inflates results. read the letter →

arxiv 2507.04006 v1 pith:WF6K3EZ4 submitted 2025-07-05 cs.CV

Seungjin Jung , Kanghee Lee , Yonghyun Jeong , Haeun Noh , Jungmin Lee , Jongwon Choi This is my paper · ORCID
classification cs.CV
keywords faceanti-spoofingdomaingeneralizationdomain-invariantfeaturesorthogonaldecompositionGram-Schmidtprocessgroup-wisescalingcalibrationerrortext-guidedlearning
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 tries to establish that cross-domain face anti-spoofing fails for two separate reasons: the directions of local decision boundaries differ across domains, and the bias terms do too. It proposes GD-FAS, which addresses both at once by decomposing image features into orthogonal domain-invariant and domain-specific subspaces and by reweighting per-group losses so no domain dominates training. If the claim is right, a model trained on a few known domains can draw a single, consistently calibrated decision boundary on unseen domains, with a reported average HTER of 1.83 percent and AUC of 99.46 percent on the standard leave-one-domain-out protocol. The paper also introduces Expected Calibration Error as a quantitative measure of whether bias alignment is actually achieved.

What carries the argument

The load-bearing object is the orthogonal decomposition of image features relative to a text-derived invariant direction. From the CLIP text embeddings for live and spoof prompts, Gram-Schmidt yields a basis $b_I$ for the invariant subspace; invariant features are $f_I = \operatorname{proj}_{b_I}(w_I)$, and domain-specific features are $f_S = w_I - f_I$, so $f_I \cdot f_S = 0$ by construction. Group-wise Scaling Risk Minimization reweights losses per $(y,e)$ group with a normalized-loss scaler $\sigma(\bar{L}_g; \alpha, \beta)$, and the full objective adds the image-text similarity loss, the FOD contrastive loss, an image-image InfoNCE loss, and a class loss. The text direction is reused as the classifier direction, which is what makes $f_S \cdot w_T = 0$ the mechanism that removes bias terms.

What would settle it

Train the model on the public benchmark domains, then evaluate on a target domain whose spoof type is absent from the text prompts. If a linear classifier trained only on the residual features separates live from spoof in that domain with meaningful accuracy, the residual contains invariant information and the orthogonality premise is false.

Watch

Extended reading notes

Core claim

The central claim is that the residual domain gap in face anti-spoofing can be split cleanly into a weight misalignment and a bias misalignment, and that correcting both together gives a universal decision boundary. Feature Orthogonal Decomposition projects each image embedding onto an invariant direction built from the text-embedding span of live and spoof prompts via Gram-Schmidt; the leftover component is treated as domain-specific. A contrastive loss on that residual enforces orthogonality between invariant and specific parts, so the specific part has zero inner product with the text direction and therefore does not shift the classification threshold. Group-wise Scaling Risk Minimization balances losses separately for each (label, domain) group, aligning bias terms by preventing any one domain from dominating. The authors report that this jointly aligned boundary reaches state-of-the-art accuracy and lower Expected Calibration Error than prior approaches on unseen target domains.

Load-bearing premise

The method assumes that the span of the live/spoof text embeddings points exactly along the true domain-invariant classification direction, so everything orthogonal to it is purely domain-specific; if that direction is even slightly off, the residual keeps part of the invariant signal and the bias-alignment conclusion no longer follows.

Editorial extensions

If this is right

  • On the leave-one-domain-out benchmark, the method reports an average HTER of 1.83 percent and AUC of 99.46 percent, best among the compared methods.
  • Because bias alignment is measured by Expected Calibration Error, lower ECE in most target settings is taken as evidence that thresholds, not just directions, now transfer to unseen domains.
  • Training curves stay stable across epochs, so the method does not depend on early stopping the way the closest text-guided baseline does.
  • With an additional source dataset, the model reaches 0.00 percent HTER on one target split, a perfect separation of live and spoof in that setting.
  • Under leave-two- and leave-three-domain-out protocols and on a different three-dataset benchmark, the same framework keeps the best or most balanced HTER and AUC.

Reading between the lines

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

  • If the text-prompt direction is not the true invariant boundary, the residual $f_S$ will still contain useful invariant signal, and the reported gains could come largely from the auxiliary losses rather than from genuine bias alignment.
  • One direct test is to freeze the text direction and probe the residual with a fresh linear classifier; if that classifier can separate live from spoof above chance inside a held-out domain, the orthogonal decomposition is not removing all invariant information.
  • The ECE argument assumes calibration error is dominated by threshold or bias mismatch; comparing per-domain decision thresholds directly would separate calibration effects from confidence miscalibration unrelated to domain shift.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GD-FAS, a domain-generalizable face anti-spoofing framework with two components: Group-wise Scaling Risk Minimization (GS-RM), which reweights group-wise losses to balance learning across domains, and Feature Orthogonal Decomposition (FOD), which projects image features onto the span of CLIP text embeddings and treats the residual as domain-specific, enforcing orthogonality via a contrastive loss. The authors claim that this jointly aligns classifier weights and biases, and introduce Expected Calibration Error (ECE) as a metric for bias alignment. Experiments on OULU-NPU, CASIA, Replay-Attack, MSU-MFSD and additional protocols report state-of-the-art HTER/AUC numbers. The paper includes a public code release, ablation studies, stability plots, and an extensive supplementary appendix.

Significance. If the claims were fully supported, the paper would make a useful contribution to domain-generalizable FAS: explicit decomposition into invariant and specific subspaces is a principled idea, and using text embeddings as a fixed invariant basis is computationally lightweight. The paper is also transparent about code release and includes a broad set of experiments (leave-one-out, leave-two-out, leave-three-out, cross-dataset). However, as it stands, the empirical evidence does not substantiate the central bias-alignment claim, and the evaluation protocol has methodological flaws that make the reported state-of-the-art numbers unverifiable as a fair generalization test. The strengths are the clarity of the geometric motivation and the breadth of experiments; the weaknesses are the oracle-style selection rule and the contradiction between the proposed ECE metric and the ablation results.

major comments (3)
  1. [Table 1 (caption), Section 4.2, Appendix H] The reported state-of-the-art results are not based on a selection rule that would transfer to unseen domains. Table 1's caption ('Best results considered across epochs') means the final numbers are the best over the training trajectory, which requires access to the target domain to pick the epoch. Appendix H further states that hyperparameters λ1 and β were selected based on HTER, i.e., performance on the held-out target domain. Together these make the comparison an oracle evaluation: the model is effectively tuned on the test domains. To substantiate the claimed generalization, the authors should report final-epoch results under a fixed training schedule and select hyperparameters using only source-domain validation (or a fixed rule). If best-epoch numbers are kept, they should be reported as an upper bound, not as state-of-the-art.
  2. [Section 4.3, Eq. (16), Section 4.5, Table 3] The central claim that GS-RM aligns bias terms is contradicted by the paper's own evidence. Section 4.3 states that a lower ECE indicates better bias-term alignment, and Eq. (16) defines ECE. Yet in Table 3, adding GS-RM to FOD increases ECE in all four leave-one-out settings (OCI→M: 1.60 to 3.39; OMI→C: 1.00 to 1.82; OCM→I: 1.48 to 2.37; ICM→O: 2.38 to 2.86). The text in Section 4.5 acknowledges a 'trade-off' and says GS-RM mainly improves generalization, but this is inconsistent with the claim that FOD 'does not negatively impact bias alignment' and that the framework 'jointly aligns weights and biases.' Moreover, ECE as defined in Eq. (16) is a calibration metric (confidence vs. accuracy) computed on the target distribution; it does not measure the spread of classification thresholds across domains. The paper therefore either provides no direct evidence for the bias-alignment mechanism, or its own metric shows the full model aligns biases worse than FOD alone. The authors should either revise the claim, or measure bias alignment directly, e.g., by reporting per-domain threshold statistics or the variance of decision boundaries across source domains.
  3. [Section 3.3, Eq. (11); Table F] The orthogonal decomposition's validity rests on the assumption that the span of CLIP text embeddings for live/spoof prompts coincides with the optimal domain-invariant classification direction w*. If the text subspace is not aligned with w*, then fS = wI - proj_{bI}(wI) in Eq. (11) contains useful invariant signal, and the FOD loss in Eq. (12) may discard that signal. The paper does not test this assumption. The sensitivity shown in Table F — where replacing text templates with a single template or random vectors changes HTER dramatically (e.g., OMI→C: 0.93 to 12.04 for random vectors) — is consistent with the concern that the 'invariant' subspace is template-dependent rather than causally invariant. The authors should verify the assumption, for example by measuring the alignment between the learned classifier direction and the text basis, or by comparing FOD performance against decomposition using a learned invariant direction.
minor comments (6)
  1. [Table 1, SAFAS row] The reported average HTER of 3.60 and AUC of 99.00 are inconsistent with the four per-domain values (5.95, 8.78, 6.58, 10.00 for HTER; 96.55, 95.37, 97.54, 96.23 for AUC), which average to 7.83 and 96.42. Please correct or clarify.
  2. [Section 4.1, Metrics] 'Halt Total Error Rate' is a typo for 'Half Total Error Rate.'
  3. [Section 4.3] The sentence 'A lower ECE score indicates greater confidence in the model's predictions' misstates what ECE measures; a lower ECE indicates better calibration between confidence and accuracy, not greater confidence.
  4. [Figure 4] The x-axis label '100/epochs' is ambiguous; it should be 'epochs.'
  5. [Section 4.5] The text says 'GS-RM primarily focuses on improving generalization' but Section 3.4 claims GS-RM is designed for bias alignment; the two roles should be clarified.
  6. [Appendix B, Fig. D] The reliability diagrams in Fig. D are discussed qualitatively, but the ECE values in the figure captions (e.g., 20.22 for SAFAS uncalibrated) are not reconciled with Table 2, which reports SAFAS ECE values in a different range; please specify the protocol used for the figure.

Circularity Check

3 steps flagged · score 5.0 of 10

Bias alignment is definitional and its ECE evidence is contradictory; reported HTER is partly closed-loop via target-domain hyperparameter/epoch selection.

  1. self definitional [Sec. 3.3 'Geometrical Motivation' and Sec. 3.5 'Feature Orthogonal Decomposition', Eqs. 10-11]
    "We assume that a feature f is always combination of both domain-invariant fI and domain-specific fS components ... Under these conditions, fS · wT stays zero due to the orthogonality of the inner product, ensuring that b is consistently aligned as zero across domains. ... Then, we derive the domain-specific feature fS by removing the domain-invariant feature fI from the image embedding features WI. As a result, the domain-invariant fI and domain-specific features fS become fully orthogonalized, ensuring their inner product is always zero."

    fS is defined as the orthogonal residual after projecting wI onto the text basis bI (Eq. 11), so the claimed key property fS · wT = 0 holds by construction, not by empirical discovery. The paper presents this as the geometrical reason bias terms can be aligned, but it is a restatement of the decomposition definition and depends on the untested premise that the text-embedding span is the true invariant direction w*. The 'prediction' is therefore the definition.

  2. fitted input called prediction [Appendix H 'Sensitivity of Hyperparameter' and Table 1 caption]
    "We selected the final hyperparameter values based on HTER, as it directly reflects the model's domain generalization ability. ... Best results considered across epochs."

    The FOD/GS-RM hyperparameters (λ1, β) are chosen by looking at HTER on the same leave-one-out target domains that are then reported as the method's predictions. Table 1 additionally selects the best-performing epoch on the target. Thus the headline SOTA numbers are not independent out-of-sample evaluations; the reported HTER/AUC gains are partly a function of fitting the model-selection rule to the test set, so the prediction is statistically forced rather than fully predicted.

1 more flagged steps
  1. other [Sec. 4.3 'Evaluate Reliability and Bias Term Alignment' and Table 3 ablation]
    "The confidence score is directly influenced by misaligned bias terms across different domains, as they affect the classification threshold. Therefore, a lower Expected Calibration Error (ECE) score indicates better alignment ... using only FOD can significantly reduce ECE, but our full model achieves a more balanced trade-off"

    ECE (Eq. 16) is a standard confidence-calibration metric, not a direct measurement of cross-domain threshold/bias alignment; the paper's assertion that lower ECE 'indicates better [bias] alignment' already assumes the conclusion it is offered to support. Moreover, Table 3 shows the full FOD+GS-RM model has worse ECE than FOD-only in all four settings (e.g., 3.39 vs 1.60 for OCI→M), contradicting the abstract's claim that FOD 'does not negatively impact bias alignment.' The only quantitative evidence for bias alignment is either a renamed calibration metric or contradicted by the paper's own ablation.

full rationale

The paper is not fully circular: the benchmark comparisons are against external methods, the architecture is concrete, and the self-citation [16] is used only as inspiration for the scaling estimator rather than as load-bearing proof. However, the central bias-alignment claim is built into the definition of FOD: fS is constructed as the orthogonal residual to the text basis, so the Sec. 3.3 conclusion fS · wT = 0 is a tautology once the unverified text-subspace assumption is granted. The ECE-based 'quantitative validation' of bias alignment is either a renaming of a calibration metric or is contradicted by Table 3, where adding GS-RM worsens ECE in all four leave-one-out settings. Finally, the state-of-the-art claim is partially closed-loop because hyperparameters are selected on target HTER (Appendix H) and the best epoch is selected on the target (Table 1). These issues make the derivation and evaluation partially circular, but the external benchmark content and reproducible pipeline keep the score moderate rather than maximal.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a geometric assumption about CLIP text embeddings defining the invariant subspace, an empirical loss-balancing rule, and the treatment of ECE as a bias-alignment measure. The method introduces no new physical or architectural entities, but it does rely on several parameters tuned on target-domain performance.

free parameters (4)
  • beta (GS estimator scale) = 1.5
    Equation 6; the authors state alpha and beta were empirically determined through experimentation.
  • alpha (GS estimator slope) = log(|g|)/2
    Equation 6; the slope constant was chosen empirically and depends on the number of groups.
  • lambda_1 (FOD loss weight) = 0.8
    Total loss in Eq 15; Appendix H says final hyperparameters were selected based on HTER on the held-out target domains.
  • lambda_2 (image-image similarity loss weight) = 0.1
    Total loss in Eq 15; fixed in all experiments without an ablation of this parameter.
assumptions (5)
  • standard math Gram-Schmidt orthogonalization and vector projection produce an orthogonal decomposition of image features (Eqs 10-11).
    Standard linear algebra invoked in Sec 3.5 to construct invariant and specific features.
  • domain assumption A feature f is exactly the sum of a domain-invariant component fI in the text-embedding span and a domain-specific component fS, and fS dot wT = 0 forces classifier bias to be aligned (Sec 3.3).
    Load-bearing geometric motivation; not proven for CLIP features or the actual network.
  • domain assumption Frozen CLIP text embeddings for live/spoof templates are a reliable invariant classifier direction for face anti-spoofing across unseen domains.
    Adopted from FLIP [28]; both FOD and the image-text similarity loss depend on this direction.
  • ad hoc to paper Balancing normalized group-wise losses with the sigmoid scaler aligns classifier bias terms across domains.
    GS-RM claim; the scaler form in Eq 6 is chosen empirically, and no derivation links it to bias terms.
  • domain assumption Expected Calibration Error is an appropriate quantitative proxy for bias-term alignment across domains.
    Sec 4.3 asserts lower ECE means better bias alignment, but ECE measures confidence calibration, not classifier bias directly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Group-wise Scaling and Orthogonal Decomposition for Domain-Invariant Feature Extraction in Face Anti-Spoofing." pith.science (2026). https://pith.science/paper/WF6K3EZ4

@misc{pith2026250704006,
  author       = {Pith},
  title        = {Pith review of: Group-wise Scaling and Orthogonal Decomposition for Domain-Invariant Feature Extraction in Face Anti-Spoofing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WF6K3EZ4}},
  note         = {Machine review of arXiv:2507.04006}
}
read the original abstract

Domain Generalizable Face Anti-Spoofing (DGFAS) methods effectively capture domain-invariant features by aligning the directions (weights) of local decision boundaries across domains. However, the bias terms associated with these boundaries remain misaligned, leading to inconsistent classification thresholds and degraded performance on unseen target domains. To address this issue, we propose a novel DGFAS framework that jointly aligns weights and biases through Feature Orthogonal Decomposition (FOD) and Group-wise Scaling Risk Minimization (GS-RM). Specifically, GS-RM facilitates bias alignment by balancing group-wise losses across multiple domains. FOD employs the Gram-Schmidt orthogonalization process to decompose the feature space explicitly into domain-invariant and domain-specific subspaces. By enforcing orthogonality between domain-specific and domain-invariant features during training using domain labels, FOD ensures effective weight alignment across domains without negatively impacting bias alignment. Additionally, we introduce Expected Calibration Error (ECE) as a novel evaluation metric for quantitatively assessing the effectiveness of our method in aligning bias terms across domains. Extensive experiments on benchmark datasets demonstrate that our approach achieves state-of-the-art performance, consistently improving accuracy, reducing bias misalignment, and enhancing generalization stability on unseen target domains.

Figures

Figures reproduced from arXiv: 2507.04006 by the authors.

Figure 1
Figure 1. Bias term and weight term across domains: Decision boundaries are defined by the classifier’s weights and biases. Do￾main invariance maintains consistent distinguishability across do￾mains, typically by sharing a common weight term. In contrast, domain specificity represents domain gaps arising from variations in bias terms. (a) Domain-wise contrastive loss enhances domain￾specific separability with bias alignment b… view at source ↗
Figure 2
Figure 2. Framework overview: Our framework, based on CLIP [28], achieves domain generality by decomposing representations into invariant and specific components within the Text-Image embedding space. (1) GS-RM balances group-wise loss across domains to en￾hance domain invariance. (2) FOD leverages the Gram-Schmidt process to decompose domain-invariant and domain-specific components, effectively mitigating major spurious corr… view at source ↗
Figure 3
Figure 3. Insight for GS-RM and FOD: The center panel illustrates text features and the features of each domain represented within a hy￾persphere. The left panel demonstrates GS, which penalizes low-loss (high-similarity) groups and boosts high-loss (low-similarity) groups to achieve balanced learning quantities across groups. The right panel utilizes orthogonal bases and orthogonal projections to decompose domain-specific co… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Learning stability: Changes in performance measurements according to epoch settings for each domain. Method OCI→M OMI→C OCM→I ICM→O SSDG-R [14] 8.40 11.09 8.94 9.47 SSAN-R [36] 9.58 5.92 12.03 11.10 PatchNet [31] 19.98 20.98 24.31 34.91 SAFAS [29] 18.12 17.77 7.77 13.8…
Figure 5
Figure 5. Figure 5: Ablation study on each component for GD-FAS: 2D PCA Visualization for OCI→M according to each module 4.5. Ablation Study We conducted an ablation study to evaluate the effectiveness of GSRM and FOD on domain generalization and model reliability, as summarized in Tab. 3…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages

  1. [1]

    Invariance principle meets in- formation bottleneck for out-of-distribution generalization

    Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean- Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets in- formation bottleneck for out-of-distribution generalization. Advances in Neural Information Processing Systems , 34: 3438–3450, 2021. 3

    work page 2021

  2. [2]

    Invariant risk minimization

    Martin Arjovsky, L ´eon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 3, 4

    arXiv 1907

  3. [3]

    Boulkenafet, J

    Z. Boulkenafet, J. Komulainen, Lei. Li, X. Feng, and A. Hadid. OULU-NPU: A mobile face presentation attack database with real-world variations. 2017. 6

    work page 2017

  4. [4]

    Linear algebra: Theory and applications

    Ward Cheney and David Kincaid. Linear algebra: Theory and applications. The Australian Mathematical Society, 110: 544–550, 2009. 5, 11

    work page 2009

  5. [5]

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

    Ivana Chingovska, Andr ´e Anjos, and S ´ebastien Marcel. On the effectiveness of local binary patterns in face anti- spoofing. 2012. 6

    work page 2012

  6. [6]

    Probable domain generalization via quantile risk minimization

    Cian Eastwood, Alexander Robey, Shashank Singh, Julius V on K¨ugelgen, Hamed Hassani, George J Pappas, and Bern- hard Sch¨olkopf. Probable domain generalization via quantile risk minimization. Advances in Neural Information Process- ing Systems, 2022. 3

    work page 2022

  7. [7]

    Biometric face presentation attack detection with multi-channel convolutional neural network

    George et al. Biometric face presentation attack detection with multi-channel convolutional neural network. TIFS,

    work page

  8. [8]

    Casia-surf cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing

    Liu et al. Casia-surf cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing. In WACV, 2021. 15

    work page 2021

Show all 45 references
  1. [9]

    Casia-surf: A large-scale multi-modal bench- mark for face anti-spoofing

    Zhang et al. Casia-surf: A large-scale multi-modal bench- mark for face anti-spoofing. TBBIS, 2020. 15

    2020

  2. [10]

    Dlow: Domain flow for adaptation and generalization

    Rui Gong, Wen Li, Yuhua Chen, and Luc Van Gool. Dlow: Domain flow for adaptation and generalization. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2477–2486, 2019. 3

    2019

  3. [11]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321–1330. PMLR,

  4. [12]

    Multi- domain learning for updating face anti-spoofing models

    Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multi- domain learning for updating face anti-spoofing models. In European Conference on Computer Vision, pages 230–249. Springer, 2022. 6

    2022

  5. [13]

    Domain-generalized face anti- spoofing with unknown attacks

    Zong-Wei Hong, Yu-Chen Lin, Hsuan-Tung Liu, Yi-Ren Yeh, and Chu-Song Chen. Domain-generalized face anti- spoofing with unknown attacks. In 2023 IEEE International Conference on Image Processing (ICIP) , pages 820–824. IEEE, 2023. 15

    2023

  6. [14]

    Single-side domain generalization for face anti-spoofing

    Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 8484–8493, 2020. 2, 6, 7, 15

    2020

  7. [15]

    Fea- ture alignment and restoration for domain generalization and adaptation

    Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Fea- ture alignment and restoration for domain generalization and adaptation. arXiv preprint arXiv:2006.12009, 2020. 3

    2006 arXiv

  8. [16]

    Scaling of class-wise training losses for post-hoc calibration

    Seungjin Jung, Seungmo Seo, Yonghyun Jeong, and Jong- won Choi. Scaling of class-wise training losses for post-hoc calibration. In Proceedings of the 40th International Con- ference on Machine Learning, pages 15421–15434. PMLR,

  9. [17]

    Does invariant risk minimization capture in- variance? In International Conference on Artificial Intelli- gence and Statistics, pages 4069–4077

    Pritish Kamath, Akilesh Tangella, Danica Sutherland, and Nathan Srebro. Does invariant risk minimization capture in- variance? In International Conference on Artificial Intelli- gence and Statistics, pages 4069–4077. PMLR, 2021. 3, 4

    2021

  10. [18]

    Domain generalization with adversarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2018. 6

    2018

  11. [19]

    Rethinking feature- based knowledge distillation for face recognition

    Jingzhi Li, Zidong Guo, Hui Li, Seungju Han, Ji-won Baek, Min Yang, Ran Yang, and Sungjoo Suh. Rethinking feature- based knowledge distillation for face recognition. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20156–2016...

    2023

  12. [20]

    Cfpl- fas: Class free prompt learning for generalizable face anti- spoofing

    Ajian Liu, Shuai Xue, Jianwen Gan, Jun Wan, Yanyan Liang, Jiankang Deng, Sergio Escalera, and Zhen Lei. Cfpl- fas: Class free prompt learning for generalizable face anti- spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 222–232,

  13. [21]

    Adaptive normalized representation learning for generaliz- able face anti-spoofing

    Shubao Liu, Ke-Yue Zhang, Taiping Yao, Mingwei Bi, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Adaptive normalized representation learning for generaliz- able face anti-spoofing. In Proceedings of the 29th ACM international conference on multimedia , pages 1469–1477,

  14. [22]

    Dual reweighting domain generalization for face pre- sentation attack detection

    Shubao Liu, Ke-Yue Zhang, Taiping Yao, Kekai Sheng, Shouhong Ding, Ying Tai, Jilin Li, Yuan Xie, and Lizhuang Ma. Dual reweighting domain generalization for face pre- sentation attack detection. arXiv preprint arXiv:2106.16128,

    arXiv

  15. [23]

    Bottom-up do- main prompt tuning for generalized face anti-spoofing

    Si-Qi Liu, Qirui Wang, and Pong C Yuen. Bottom-up do- main prompt tuning for generalized face anti-spoofing. In European Conference on Computer Vision . Springer, 2025. 2, 3, 5, 6, 14, 15

    2025

  16. [24]

    Learning deep models for face anti-spoofing: Binary or auxiliary su- pervision

    Yaojie Liu, Amin Jourabloo, and Xiaoming Liu. Learning deep models for face anti-spoofing: Binary or auxiliary su- pervision. In Proceedings of the IEEE conference on com- puter vision and pattern recognition , pages 389–398, 2018. 2

    2018

  17. [25]

    Obtaining well calibrated probabilities using 9 bayesian binning

    Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using 9 bayesian binning. In Twenty-Ninth AAAI Conference on Ar- tificial Intelligence, 2015. 6

    2015

  18. [26]

    The risks of invariant risk minimization

    Elan Rosenfeld, Pradeep Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. In International Conference on Learning Representations, 2021. 3, 4

    2021

  19. [27]

    Multi-adversarial discriminative deep domain generalization for face presentation attack detection

    Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023–10031, 2019. 3, 6

    2019

  20. [28]

    Flip: Cross-domain face anti-spoofing with lan- guage guidance

    Koushik Srivatsan, Muzammal Naseer, and Karthik Nan- dakumar. Flip: Cross-domain face anti-spoofing with lan- guage guidance. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) , pages 19685–19696, 2023. 1, 2, 3, 5, 6, 7, 14, 15

    2023

  21. [29]

    Rethinking domain generalization for face anti- spoofing: Separability and alignment

    Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, and Wen- Sheng Chu. Rethinking domain generalization for face anti- spoofing: Separability and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24563–24574, 2023. 1, 2, 3, 4, 6, 7

    2023

  22. [30]

    V . Vapnik. Principles of risk minimization for learning the- ory. In Advances in Neural Information Processing Systems. Morgan-Kaufmann, 1991. 3

    1991

  23. [31]

    Patchnet: A simple face anti-spoofing frame- work via fine-grained patch recognition

    Chien-Yi Wang, Yu-Ding Lu, Shang-Ta Yang, and Shang- Hong Lai. Patchnet: A simple face anti-spoofing frame- work via fine-grained patch recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20281–20290, 2022. 1, 2, 6, 7

    2022

  24. [32]

    Cross-domain face presentation attack detection via multi- domain disentangled representation learning

    Guoqing Wang, Hu Han, Shiguang Shan, and Xilin Chen. Cross-domain face presentation attack detection via multi- domain disentangled representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6678–6687, 2020. 2

    2020

  25. [33]

    Self-domain adaptation for face anti-spoofing

    Jingjing Wang, Jingyi Zhang, Ying Bian, Youyi Cai, Chun- mao Wang, and Shiliang Pu. Self-domain adaptation for face anti-spoofing. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, 2021. 6

    2021

  26. [34]

    Domainmix: Learning generalizable person re- identification without human annotations

    Wenhao Wang, Shengcai Liao, Fang Zhao, Cuicui Kang, and Ling Shao. Domainmix: Learning generalizable person re- identification without human annotations. arXiv preprint arXiv:2011.11953, 2020. 3

    2011 arXiv

  27. [35]

    Tf-fas: twofold-element fine-grained semantic guidance for general- izable face anti-spoofing

    Xudong Wang, Ke-Yue Zhang, Taiping Yao, Qianyu Zhou, Shouhong Ding, Pingyang Dai, and Rongrong Ji. Tf-fas: twofold-element fine-grained semantic guidance for general- izable face anti-spoofing. In European Conference on Com- puter Vision, pages 148–168. Springer, 2025. 2, 3, 5, 6, 14

    2025

  28. [36]

    Domain gen- eralization via shuffled style assembly for face anti-spoofing

    Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jia- hong Li, Tingting Gao, and Zhongyuan Wang. Domain gen- eralization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4123–4133, ...

    2022

  29. [37]

    K Jain, and Hu Han

    Di Wen, Anil. K Jain, and Hu Han. Face Spoof Detection with Image Distortion Analysis. IEEE Trans. Information Forensic and Security, 2015 (To Appear). 6, 14

    2015

  30. [38]

    Towards effective adversarial textured 3d meshes on physical face recognition

    Xiao Yang, Chang Liu, Longlong Xu, Yikai Wang, Yinpeng Dong, Ning Chen, Hang Su, and Jun Zhu. Towards effective adversarial textured 3d meshes on physical face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 4119–4128,

  31. [39]

    Multi-modal face anti- spoofing based on central difference networks

    Zitong Yu, Yunxiao Qin, Xiaobai Li, Zezheng Wang, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Multi-modal face anti- spoofing based on central difference networks. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 650–651, 2020. 2

    2020

  32. [40]

    Search- ing central difference convolutional networks for face anti- spoofing

    Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Search- ing central difference convolutional networks for face anti- spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5295...

    2020

  33. [41]

    Cycli- cally disentangled feature translation for face anti-spoofing

    Haixiao Yue, Keyao Wang, Guosheng Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Cycli- cally disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 3358–3366, 2023. 2

    2023

  34. [42]

    Joint face detection and alignment using multitask cascaded convolutional networks

    Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters , 23 (10):1499–1503, 2016. 6

    2016

  35. [43]

    A face antispoofing database with diverse attacks

    Zhiwei Zhang, Junjie Yan, Sifei Liu, Zhen Lei, Dong Yi, and Stan Z Li. A face antispoofing database with diverse attacks. In 2012 5th IAPR international conference on Bio- metrics (ICB), pages 26–31. IEEE, 2012. 6

    2012

  36. [44]

    Generative do- main adaptation for face anti-spoofing

    Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative do- main adaptation for face anti-spoofing. In European Con- ference on Computer Vision, pages 335–356. Springer, 2022. 2

    2022

  37. [45]

    Instance-aware do- main generalization for face anti-spoofing

    Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware do- main generalization for face anti-spoofing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20453–20463, 2023. 2 10 Group-w...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.