Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that face anti-spoofing is better done as an interpretable visual question-answering task, and that the approach cuts average cross-domain error from 18.73% to 11.30% on a new benchmark.

desk verdict A genuinely useful FAS paper, but the headline cross-domain gains are clouded by a captioner that saw the target domains during fine-tuning. read the letter →

arxiv 2501.01720 v2 pith:IVDGNEHG submitted 2025-01-03 cs.CV

Guosheng Zhang , Keyao Wang , Haixiao Yue , Ajian Liu , Gang Zhang , Kun Yao , Errui Ding , Jingdong Wang This is my paper · ORCID
classification cs.CV
keywords faceanti-spoofingmultimodallargelanguagemodelvisualquestionansweringcross-domaingeneralizationspoof-awarecaptioninglopsidedlossgloballyawareconnectorinterpretability
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 sets out to show that face anti-spoofing (deciding whether a face image is genuine or an attack) is better done as an interpretable visual question-answering task than as a binary classification. The authors argue that if the model is trained not only to answer 'Is this a real person?' but also to generate natural-language captions that name the spoof cue (a screen, paper, mask, mannequin), it generalizes better to unseen environments and attack types. To produce those captions they filter generic captions by spoof keywords, fine-tune a captioner on the surviving subset, and then use a loss that weights the yes/no judgment more heavily than the caption. On a new One-to-Eleven benchmark (training on CelebA-Spoof only, testing on 11 other datasets) they report average HTER falling from 18.73% for the prior FLIP method to 11.30%.

What carries the argument

The central machinery is the interpretable VQA formulation with three components. SCF is a two-stage captioning pipeline: a general captioner (BLIP2) generates captions; those containing keywords from a spoof-type dictionary are kept; the captioner is fine-tuned on that filtered subset to become 'spoof-aware'; it then captions all fake images, while the general captioner captions real images. The L-LM loss decomposes the answer into the judgment token ('Yes'/'No') and the interpretation caption, computing separate cross-entropy terms with weight α on the judgment. The GAC connector computes multi-layer CLS tokens from the frozen ViT, projects them, concatenates them with learnable queries, and runs self- and cross-attention against local visual features to produce the visual tokens fed to the frozen OPT language model.

What would settle it

Train the full I-FAS pipeline but generate the training captions for CelebA-Spoof using a spoof-aware captioner that has never seen any target-domain images; if the average HTER on Protocol 2 rises well above the reported 11.30%, then the claimed domain generalization depends on target-domain information leaking through captions. Alternatively, have human annotators label whether each generated caption's spoof keyword matches the actual attack in the image; if a large fraction are wrong or miss the cue, the interpretation supervision is unreliable.

Watch

Extended reading notes

Core claim

The central claim is that reformulating face anti-spoofing as an interpretable VQA paradigm—where every image gets a 'Yes/No' judgment plus a spoof-aware caption—substantially improves cross-domain generalization relative to prior methods that treat FAS as binary classification or use hand-crafted text templates. The authors introduce three mechanisms that carry the result: Spoof-aware Captioning and Filtering (SCF) generates and selects captions that reference the attack medium; a Lopsided Language Model (L-LM) loss separates judgment and interpretation terms, prioritizing the binary judgment; and a Globally Aware Connector (GAC) feeds multi-level global visual features into the language model. The paper claims this combination outperforms state-of-the-art methods on both a standard leave-one-out protocol (Protocol 1, average HTER 1.33% vs. 2.98% for the best prior) and the newly proposed One-to-Eleven benchmark (Protocol 2, average HTER 11.30% vs. 18.73% for FLIP).

Load-bearing premise

The method presumes that the spoof-aware captioner, fine-tuned on a keyword-filtered subset of fake images, produces accurate captions that name the true spoof cue for every image — including images from target domains the captioner has seen during fine-tuning.

Editorial extensions

If this is right

  • If the model's judgment rests on a caption that names the spoof medium, the system can explain its decision in plain language rather than only outputting a confidence score.
  • The reported gains on Protocol 2 imply that a single-source training set (CelebA-Spoof) is sufficient to build a face-spoof detector that transfers to 11 unseen datasets, including 3D-mask and mannequin attacks absent from the source.
  • The approach should be applicable to other visual security tasks where binary labels are scarce and text descriptions can carry supervision, such as deepfake or document forgery detection.
  • The new One-to-Eleven protocol, with 12 public datasets and one source domain, gives the community a more demanding cross-domain test than the standard four-dataset leave-one-out protocol.

Reading between the lines

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

  • The reported cross-domain gains may be partly inflated by the fact that the spoof-aware captioner is fine-tuned on captions from all 12 datasets (including the 11 target domains) before the main model is trained on CelebA-Spoof; a stricter protocol would fine-tune the captioner only on the source-domain images, or test captions on held-out domains.
  • If caption quality is the ceiling, improving the captioner (e.g., with visual grounding or larger MLLMs) should directly raise FAS accuracy; this suggests a testable scaling hypothesis.
  • The lopsided loss design—decoupling a hard judgment from auxiliary text generation—is a generally applicable recipe for noisy-language supervision and could be adapted to other VQA tasks where explanations are unreliable.
  • One could stress-test the claim by ablating the SCF captioner with a random subset of fake captions that still mention keywords; if gains vanish, the specific caption content rather than mere language diversity is doing the work.
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 / 4 minor

Summary. The paper proposes Interpretable Face Anti-Spoofing (I-FAS), a multimodal large language model framework that reframes face anti-spoofing as a visual question answering task. The method has three components: Spoof-aware Captioning and Filtering (SCF) generates natural-language captions emphasizing spoof cues; a Lopsided Language Model (L-LM) loss separates the loss for the yes/no judgment from the loss for the interpretation; and a Globally Aware Connector (GAC) injects multi-layer global visual features into the language model. The authors evaluate on two protocols: Protocol 1 is the standard leave-one-domain-out benchmark on MSU-MFSD, CASIA-FASD, Replay-Attack, and OULU-NPU, achieving 1.33% average HTER; Protocol 2 is a new 'One to Eleven' benchmark with CelebA-Spoof as the only source domain and 11 other datasets as targets, achieving 11.30% average HTER versus 18.73% for FLIP. The paper also reports ablations and qualitative examples showing that the model produces interpretable rationales along with decisions.

Significance. If the reported results are valid, the work is significant: it is among the first to treat face anti-spoofing as an interpretable VQA problem with an MLLM, and the large cross-domain gains, particularly on datasets with no spoof-type overlap with the source, would be an important advance. The proposal is technically concrete (a connector, a loss function, and a captioning pipeline) and the experiments cover 12 public datasets. However, the validity of the central claim is seriously undermined by the SCF leakage issue described below, which makes the Protocol 2 numbers not a clean measure of domain generalization. The reported inconsistencies (e.g., the FLIP AUC discrepancy between Tables 2 and .8) and the lack of variance reporting prevent the reader from independently judging the significance of the numerical advantages. The idea of language-supervised FAS is promising and worth pursuing, but the current evaluation protocol does not support the strong claim of 'significantly outperforming state-of-the-art' in a single-source, multi-target setting.

major comments (3)
  1. [Algorithm 1, Appendix 'Spoof-aware Captioning and Filtering', 'Captioner'] The SCF pipeline fine-tunes the spoof-aware captioner C_S on the filtered fake-image subset D_S, which is aggregated from all 12 datasets (Algorithm 1, lines 1-10). In Protocol 2, the source is CelebA-Spoof and the 11 other datasets are declared targets. Because C_S is optimized on fake images from those target datasets, the captions T_S used as supervision for the source-domain CelebA-Spoof training images are generated by a model that has already seen target-domain spoof types and image statistics. This breaks the claimed 'training solely on the CelebA-Spoof dataset' (Table 2 heading) and the domain-generalization assumption: target-domain information can flow into the FAS model through the caption text, even though the visual encoder and LLM never see target pixels. The very large gain on Rose-Youtu (HTER from 80.73 to 5.52) is consistent with this leak, since Rose-Youtu fake images were part of the captioner fine-tuning set. A source-only control is required: either fine-tune C_S only on CelebA-Spoof filtered captions, or use an off-the-shelf captioner without any spoof-aware fine-tuning, and re-run Protocol 2. Without such a control, the reported cross-domain superiority cannot be attributed to the I-FAS formulation itself, and the One-to-Eleven benchmark does not measure pure source-domain generalization.
  2. [Table .8 versus Table 2] There is an internal numerical inconsistency: Table 2 reports FLIP's average AUC in Protocol 2 as 87.90, but Table .8 in the appendix reports the same quantity as 97.90. Since Table .8 is used to argue that Ours achieves a better HTER at comparable trainable parameters (Ours: 104M, FLIP: 170M), the discrepancy is not purely cosmetic. The reader cannot tell which value is correct or whether the parameter-efficiency comparison is based on the right numbers. The authors should unify the reported FLIP results and re-check all other numbers in the appendix against the main text.
  3. [Implementation Details and Tables 1-4] The paper states that all results are the mean of three independent runs with different seeds, but no standard deviations, confidence intervals, or per-run values are reported anywhere in the main text or appendix. For the central cross-domain claims, especially the Protocol 2 improvements over FLIP (e.g., average HTER 11.30% vs 18.73%) and the small gaps in Protocol 1 between Ours (1.33%) and the second-best method (2.98%), variance matters. The absence of any measure of statistical spread makes it impossible to know whether the differences are meaningful or could overlap with the baselines across seeds. The authors should provide standard deviations (or at least per-seed results) for all reported averaged numbers.
minor comments (4)
  1. [Abstract / Protocol 2 description] There is a typo: 'One to Eleve' in the abstract and in the Protocol 2 setup should be 'One to Eleven'.
  2. [Table .5 and Table .7 references] In the appendix, the text says 'Table .7 presents the details of keyword dictionary K' but the keyword dictionary is in Table .5; Table .7 is the detailed ablation results. The cross-references are off by one and should be corrected.
  3. [Figure 5 and appendix Figure .7] The main text refers to 'Figure .7 (Left)' for the alpha ablation and 'Figure .7 (Right)' for the loss-convergence plot, but the appendix Figure .7 is a bar chart of keyword filtering counts. The numbering and locations of the two panels should be disambiguated so that the reader can actually find the referenced plots.
  4. [Interpretability claim] The paper frames the captions as 'interpretations' that explain the model's decisions. Because the captions are generated from labels and keyword filtering and then used as training targets, they are post-hoc rationales rather than independently derived causal explanations. The qualitative examples in Figure 4 are consistent with this limitation; the authors should temper the interpretability language or provide additional evidence that the generated text reflects model-internal reasoning rather than memorized caption patterns.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the FAS judgment is trained from labels with a standard classification loss, and the caption supervision is auxiliary rather than input-equivalent.

full rationale

The paper's core claim is an empirical performance improvement. The I-FAS judgment is the probability of the word 'Yes' under a standard autoregressive loss; the Lopsided LM loss separates judgment from interpretation, so the caption supervision does not by construction determine the classification output. The SCF captions are generated by a captioner fine-tuned on label-filtered outputs, but this is a pseudo-labeling/self-training procedure, not a reduction of the prediction to the input. No uniqueness theorem or load-bearing self-citation is invoked. The main non-circularity concern is experimental: the spoof-aware captioner is fine-tuned on a filtered subset of all 12 datasets, including the 11 Protocol 2 target domains, so target-domain statistics can leak into the source-domain captions. This threatens the validity of the claimed cross-domain generalization but is a benchmark-contamination/correctness issue, not a circular derivation. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim depends mainly on the caption generation pipeline and the keyword dictionary. The only free parameter is the loss weight α. The method does not introduce new physical entities.

free parameters (1)
  • alpha (α) in L-LM loss = not stated (ablation in Figure .7)
    Balancing weight between judgment and interpretation losses; chosen via ablation.
assumptions (3)
  • domain assumption The spoof-aware captioner, fine-tuned on keyword-filtered captions, produces captions that correctly identify spoof cues in unseen domains.
    Central to SCF; if captions are wrong, the interpretation supervision degrades. See Algorithm 1 and Figure .8.
  • domain assumption The keyword dictionary K covers the spoof types in all target datasets.
    If a spoof type lacks a keyword, samples will be filtered out and the captioner will not learn to describe it. See Table .5.
  • domain assumption Multi-level CLS tokens from a frozen CLIP ViT-L/14 encode both low-level texture and high-level semantic features needed for spoof detection.
    GAC relies on this to improve perception of moiré patterns and blur. See Eq. 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/IVDGNEHG

@misc{pith2026250101720,
  author       = {Pith},
  title        = {Pith review of: Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IVDGNEHG}},
  note         = {Machine review of arXiv:2501.01720}
}
read the original abstract

Face Anti-Spoofing (FAS) is essential for ensuring the security and reliability of facial recognition systems. Most existing FAS methods are formulated as binary classification tasks, providing confidence scores without interpretation. They exhibit limited generalization in out-of-domain scenarios, such as new environments or unseen spoofing types. In this work, we introduce a multimodal large language model (MLLM) framework for FAS, termed Interpretable Face Anti-Spoofing (I-FAS), which transforms the FAS task into an interpretable visual question answering (VQA) paradigm. Specifically, we propose a Spoof-aware Captioning and Filtering (SCF) strategy to generate high-quality captions for FAS images, enriching the model's supervision with natural language interpretations. To mitigate the impact of noisy captions during training, we develop a Lopsided Language Model (L-LM) loss function that separates loss calculations for judgment and interpretation, prioritizing the optimization of the former. Furthermore, to enhance the model's perception of global visual features, we design a Globally Aware Connector (GAC) to align multi-level visual representations with the language model. Extensive experiments on standard and newly devised One to Eleven cross-domain benchmarks, comprising 12 public datasets, demonstrate that our method significantly outperforms state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2501.01720 by the authors.

Figure 1
Figure 1. Comparison of performance under a challenging [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed Interpretable Face Anti-Spoofing (I-FAS) framework. section illustrates the process of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. For real samples, we again employ CG to generate captions for all real samples IR. Finally, by incorporating differential captions corresponding to the original category labels, we achieve a more comprehensive dataset that not only categorizes but also elucidates the rationale underpin￾ning each judgment. Interpretative Instruction Tuning Revisiting MLLMs: MLLMs are designed to address so￾phisticated tasks by genera… view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Illustration of some image-caption pair from the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: The output responses of I-FAS on some images from SIW-M-V2 and Rose-Youtu dataset under the unified question: [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Left: Ablation analysis of hyperparameters [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Exploring Task-Solving Paradigm for Generalized Cross-Domain Face Anti-Spoofing via Reinforcement Fine-Tuning

    cs.CV 2025-06 conditional novelty 6.0 of 10

    Reinforcement fine-tuning of a 7B vision-language model with GRPO and task-specific rewards achieves state-of-the-art cross-domain face anti-spoofing and interpretable reasoning.

Reference graph

Works this paper leans on

62 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    Alayrac, J.-B.; Donahue, J.; Luc, P.; Miech, A.; Barr, I.; Hasson, Y.; Lenc, K.; Mensch, A.; Millican, K.; Reynolds, M.; et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 23716--23736

    2022

  2. [2]

    Boulkenafet, Z.; Komulainen, J.; Li, L.; Feng, X.; and Hadid, A. 2017. OULU-NPU: A mobile face presentation attack database with real-world variations. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), 612--618. IEEE

    work page 2017

  3. [3]

    Cai, R.; Cui, Y.; Li, Z.; Yu, Z.; Li, H.; Hu, Y.; and Kot, A. 2023. Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In CVPR, 8037--8048

    work page 2023

  4. [4]

    Cai, R.; Li, Z.; Wan, R.; Li, H.; Hu, Y.; and Kot, A. C. 2022. Learning meta pattern for face anti-spoofing. TIFS, 17: 1201--1213

    work page 2022

  5. [5]

    Cai, R.; Soh, C.; Yu, Z.; Li, H.; Yang, W.; and Kot, A. C. 2024. Towards Data-Centric Face Anti-spoofing: Improving Cross-Domain Generalization via Physics-Based Data Synthesis. International Journal of Computer Vision, 1--22

    2024

  6. [6]

    Chen, Z.; Yao, T.; Sheng, K.; Ding, S.; Tai, Y.; Li, J.; Huang, F.; and Jin, X. 2021. Generalizable representation learning for mixture domain face anti-spoofing. In Proceedings of the AAAI conference on artificial intelligence, volume 35, 1132--1139

    work page 2021

  7. [7]

    Chingovska, I.; Anjos, A.; and Marcel, S. 2012. On the effectiveness of local binary patterns in face anti-spoofing. In BIOSIG, 1--7. IEEE

    work page 2012

  8. [8]

    George, A.; Mostaani, Z.; Geissenbuhler, D.; Nikisins, O.; Anjos, A.; and Marcel, S. 2019. Biometric face presentation attack detection with multi-channel convolutional neural network. TIFS, 15: 42--55

    work page 2019

Show all 62 references
  1. [9]

    Guo, J.; Liu, H.; Luo, Y.; Hu, X.; Zou, H.; Zhang, Y.; Liu, H.; and Zhao, B. 2024 a . Style-conditional Prompt Token Learning for Generalizable Face Anti-spoofing. In ACM MM, 994--1003

    2024

  2. [10]

    Guo, X.; Liu, X.; Masi, I.; and Liu, X. 2024 b . Language-guided Hierarchical Fine-grained Image Forgery Detection and Localization. International Journal of Computer Vision

    2024

  3. [11]

    Guo, X.; Liu, X.; Ren, Z.; Grosz, S.; Masi, I.; and Liu, X. 2023. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3155--3165

    2023

  4. [12]

    Guo, X.; Liu, Y.; Jain, A.; and Liu, X. 2022. Multi-domain learning for updating face anti-spoofing models. In European Conference on Computer Vision, 230--249. Springer

    2022

  5. [13]

    Hu, C.; Zhang, K.-Y.; Yao, T.; Liu, S.; Ding, S.; Tan, X.; and Ma, L. 2024. Domain-Hallucinated Updating for Multi-Domain Face Anti-spoofing. In AAAI, volume 38, 2193--2201

    2024

  6. [14]

    Huang, H.-P.; Sun, D.; Liu, Y.; Chu, W.-S.; Xiao, T.; Yuan, J.; Adam, H.; and Yang, M.-H. 2022. Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In European conference on computer vision, 37--54. Springer

    2022

  7. [15]

    Huang, W.; Zheng, X.; Ma, X.; Qin, H.; Lv, C.; Chen, H.; Luo, J.; Qi, X.; Liu, X.; and Magno, M. 2024. An Empirical Study of LLaMA3 Quantization: From LLMs to MLLMs. arXiv:2404.14047

    2024 arXiv

  8. [16]

    Jia, Y.; Zhang, J.; Shan, S.; and Chen, X. 2020. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8484--8493

    2020

  9. [17]

    Jiang, D.; Liu, Y.; Liu, S.; Zhao, J.; Zhang, H.; Gao, Z.; Zhang, X.; Li, J.; and Xiong, H. 2023. From clip to dino: Visual encoders shout in multi-modal large language models. arXiv preprint arXiv:2310.08825

    2023 arXiv

  10. [18]

    Jiang, Y.; Yan, X.; Ji, G.-P.; Fu, K.; Sun, M.; Xiong, H.; Fan, D.-P.; and Khan, F. S. 2024. Effectiveness assessment of recent large vision-language models. Visual Intelligence, 2(1): 17

    2024

  11. [19]

    M.; and Woo, S

    Le, B. M.; and Woo, S. S. 2024. Gradient alignment for cross-domain face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 188--199

    2024

  12. [20]

    Li, H.; Li, W.; Cao, H.; Wang, S.; Huang, F.; and Kot, A. C. 2018. Unsupervised domain adaptation for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 13(7): 1794--1809

    2018

  13. [21]

    Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, 19730--19742. PMLR

    2023

  14. [22]

    Li, J.; Li, D.; Xiong, C.; and Hoi, S. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, 12888--12900. PMLR

    2022

  15. [23]

    Liao, C.-H.; Chen, W.-C.; Liu, H.-T.; Yeh, Y.-R.; Hu, M.-C.; and Chen, C.-S. 2023. Domain invariant vision transformer learning for face anti-spoofing. In WACV, 6098--6107

    2023

  16. [24]

    Liu, A. 2024. CA-MoEiT: Generalizable Face Anti-spoofing via Dual Cross-Attention and Semi-fixed Mixture-of-Expert. International Journal of Computer Vision, 1--14

    2024

  17. [25]

    Liu, A.; Ma, H.; Zheng, J.; Yuan, H.; Yu, X.; Liang, Y.; Escalera, S.; Wan, J.; and Lei, Z. 2024 a . FM-CLIP: Flexible Modal CLIP for Face Anti-Spoofing. In ACM MM, 8228–8237

    2024

  18. [26]

    Liu, A.; Wan, J.; Jiang, N.; Wang, H.; and Liang, Y. 2022 a . Disentangling Facial Pose and Appearance Information for Face Anti-spoofing. In 2022 26th International Conference on Pattern Recognition (ICPR), 4537--4543. IEEE

    2022

  19. [27]

    Liu, A.; Xue, S.; Gan, J.; Wan, J.; Liang, Y.; Deng, J.; Escalera, S.; and Lei, Z. 2024 b . CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing. In CVPR, 222--232

    2024

  20. [28]

    Liu, A.; Zhao, C.; Yu, Z.; Wan, J.; Su, A.; Liu, X.; Tan, Z.; Escalera, S.; Xing, J.; Liang, Y.; et al. 2022 b . Contrastive context-aware learning for 3d high-fidelity mask face presentation attack detection. TIFS, 17: 2497--2507

    2022

  21. [29]

    Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2024 c . Visual instruction tuning. Advances in neural information processing systems, 36

    2024

  22. [30]

    Liu, S.; Lu, S.; Xu, H.; Yang, J.; Ding, S.; and Ma, L. 2022 c . Feature generation and hypothesis verification for reliable face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 1782--1791

    2022

  23. [31]

    Liu, S.-Q.; Lan, X.; and Yuen, P. C. 2018. Remote photoplethysmography correspondence feature for 3D mask face presentation attack detection. In Proceedings of the European Conference on Computer Vision (ECCV), 558--573

    2018

  24. [32]

    Liu, Y.; Chen, Y.; Dai, W.; Gou, M.; Huang, C.-T.; and Xiong, H. 2022 d . Source-free domain adaptation with contrastive domain alignment and self-supervised exploration for face anti-spoofing. In European Conference on Computer Vision, 511--528. Springer

    2022

  25. [33]

    Liu, Y.; Chen, Y.; Dai, W.; Gou, M.; Huang, C.-T.; and Xiong, H. 2024 d . Source-Free Domain Adaptation With Domain Generalized Pretraining for Face Anti-Spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  26. [34]

    Liu, Y.; Chen, Y.; Gou, M.; Huang, C.-T.; Wang, Y.; Dai, W.; and Xiong, H. 2023. Towards unsupervised domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 20654--20664

    2023

  27. [35]

    Liu, Y.; Jourabloo, A.; and Liu, X. 2018. Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, 389--398

    2018

  28. [36]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

    2021

  29. [37]

    Shao, R.; Lan, X.; Li, J.; and Yuen, P. C. 2019. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10023--10031

    2019

  30. [38]

    Shi, Y.; Gao, Y.; Lai, Y.; Wang, H.; Feng, J.; He, L.; Wan, J.; Chen, C.; Yu, Z.; and Cao, X. 2024. Shield: An evaluation benchmark for face spoofing and forgery detection with multimodal large language models. arXiv preprint arXiv:2402.04178

    2024 arXiv

  31. [39]

    Srivatsan, K.; Naseer, M.; and Nandakumar, K. 2023. Flip: Cross-domain face anti-spoofing with language guidance. In CVPR, 19685--19696

    2023

  32. [40]

    Sun, Y.; Liu, Y.; Liu, X.; Li, Y.; and Chu, W.-S. 2023. Rethinking domain generalization for face anti-spoofing: Separability and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 24563--24574

    2023

  33. [41]

    Wang, C.-Y.; Lu, Y.-D.; Yang, S.-T.; and Lai, S.-H. 2022 a . Patchnet: A simple face anti-spoofing framework via fine-grained patch recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20281--20290

    2022

  34. [42]

    Wang, J.; Zhang, J.; Bian, Y.; Cai, Y.; Wang, C.; and Pu, S. 2021. Self-domain adaptation for face anti-spoofing. In Proceedings of the AAAI conference on artificial intelligence, volume 35, 2746--2754

    2021

  35. [43]

    Wang, K.; Zhang, G.; Yue, H.; Liang, Y.; Huang, M.; Zhang, G.; Han, J.; Ding, E.; and Wang, J. 2024 a . CSDG-FAS: Closed-Space Domain Generalization for Face Anti-spoofing. International Journal of Computer Vision, 1--14

    2024

  36. [44]

    Wang, K.; Zhang, G.; Yue, H.; Liu, A.; Zhang, G.; Feng, H.; Han, J.; Ding, E.; and Wang, J. 2024 b . Multi-domain incremental learning for face presentation attack detection. In AAAI, volume 38, 5499--5507

    2024

  37. [45]

    Wang, Z.; Wang, Z.; Yu, Z.; Deng, W.; Li, J.; Gao, T.; and Wang, Z. 2022 b . Domain generalization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4123--4133

    2022

  38. [46]

    Wen, D.; Han, H.; and Jain, A. K. 2015. Face spoof detection with image distortion analysis. IEEE Transactions on Information Forensics and Security, 10(4): 746--761

    2015

  39. [47]

    Z.; and Zhao, G

    Yu, Z.; Wan, J.; Qin, Y.; Li, X.; Li, S. Z.; and Zhao, G. 2020 a . NAS-FAS: Static-dynamic central difference network search for face anti-spoofing. IEEE transactions on pattern analysis and machine intelligence, 43(9): 3005--3023

    2020

  40. [48]

    Yu, Z.; Zhao, C.; Wang, Z.; Qin, Y.; Su, Z.; Li, X.; Zhou, F.; and Zhao, G. 2020 b . Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5295--5305

    2020

  41. [49]

    Yuan, Y.; Li, W.; Liu, J.; Tang, D.; Luo, X.; Qin, C.; Zhang, L.; and Zhu, J. 2024. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 28202--28211

    2024

  42. [50]

    Yue, H.; Wang, K.; Zhang, G.; Feng, H.; Han, J.; Ding, E.; and Wang, J. 2023. Cyclically disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 3358--3366

    2023

  43. [51]

    Zanella, L.; Menapace, W.; Mancini, M.; Wang, Y.; and Ricci, E. 2024. Harnessing Large Language Models for Training-free Video Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18527--18536

    2024

  44. [52]

    Zhang, K.-Y.; Yao, T.; Zhang, J.; Liu, S.; Yin, B.; Ding, S.; and Li, J. 2021. Structure destruction and content combination for face anti-spoofing. In 2021 IEEE International Joint Conference on Biometrics (IJCB), 1--6. IEEE

    2021

  45. [53]

    Zhang, K.-Y.; Yao, T.; Zhang, J.; Tai, Y.; Ding, S.; Li, J.; Huang, F.; Song, H.; and Ma, L. 2020 a . Face anti-spoofing via disentangled representation learning. In ECCV, 641--657. Springer

    2020

  46. [54]

    V.; et al

    Zhang, S.; Roller, S.; Goyal, N.; Artetxe, M.; Chen, M.; Chen, S.; Dewan, C.; Diab, M.; Li, X.; Lin, X. V.; et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  47. [55]

    Zhang, Y.; Colman, B.; Guo, X.; Shahriyari, A.; and Bharaj, G. 2025. Common Sense Reasoning for Deepfake Detection. In European Conference on Computer Vision

    2025

  48. [56]

    Zhang, Y.; Yin, Z.; Li, Y.; Yin, G.; Yan, J.; Shao, J.; and Liu, Z. 2020 b . Celeba-spoof: Large-scale face anti-spoofing dataset with rich annotations. In ECCV, 70--85

    2020

  49. [57]

    Zhang, Z.; Yan, J.; Liu, S.; Lei, Z.; Yi, D.; and Li, S. Z. 2012. A face antispoofing database with diverse attacks. In ICB, 26--31. IEEE

    2012

  50. [58]

    Zhou, Q.; Zhang, K.-Y.; Yao, T.; Lu, X.; Ding, S.; and Ma, L. 2024. Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 175--187

    2024

  51. [59]

    Zhou, Q.; Zhang, K.-Y.; Yao, T.; Lu, X.; Yi, R.; Ding, S.; and Ma, L. 2023. Instance-aware domain generalization for face anti-spoofing. In CVPR, 20453--20463

    2023

  52. [60]

    Zhou, Q.; Zhang, K.-Y.; Yao, T.; Yi, R.; Sheng, K.; Ding, S.; and Ma, L. 2022. Generative domain adaptation for face anti-spoofing. In European Conference on Computer Vision, 335--356. Springer

    2022

  53. [61]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  54. [62]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.