Pith. sign in

REVIEW 4 major objections 4 minor 63 references

Unified Face Matching and Physical-Digital Spoofing Attack Detection

T0 review · 4 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 A single frozen face-recognition model can also detect physical and digital spoof attacks.

desk verdict A plausible unified architecture with a load-bearing evaluation flaw: the SOTA-parity claim rests on protocol-mismatched baselines and test-set model selection. read the letter →

arxiv 2501.09635 v1 pith:PSWRAA2T submitted 2025-01-16 cs.CV

Arun Kunwar , Ajita Rattani This is my paper · ORCID
classification cs.CV
keywords facerecognitionphysicalspoofattackdetectiondigitaldeepfakeSwinTransformerHiLoattentionunifiedmodelanti-spoofing
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 claims that one model can do two jobs usually handled by separate systems: recognizing faces and detecting both physical spoofs (printed photos, replay videos, masks) and digital spoofs (deepfakes, manipulated images). It builds a Swin Transformer backbone trained only for face recognition, then freezes it. A lightweight attack-detection head is attached to a middle layer of the backbone, using HiLo attention plus a small CNN, and trained on a dataset of physical and digital attacks with special augmentations. The claim is that this unified model matches the accuracy of separate state-of-the-art detectors on standard benchmarks and on unseen attack types, while leaving the face-recognition head untouched.

What carries the argument

The load-bearing component is the placement of a HiLo attention plus CNN head on an intermediate layer of a frozen Swin Transformer. HiLo attention separates features into high-frequency local details (via window attention) and low-frequency global structure (via average pooling), which the paper argues is exactly the mix needed to catch physical artifacts such as print noise and moiré patterns and digital artifacts such as face-swap boundaries and GAN texture. The recognition head stays on the final-layer embedding trained with ArcFace loss; the attack head is trained with binary cross-entropy. The choice of intermediate block is established empirically by a sweep over all 18 blocks of Stage 3, with block 5 performing best across datasets.

What would settle it

Train the same UAD head on features from the same Swin architecture with random, untrained weights, keeping all other settings identical. If the untrained-backbone UAD reaches similar accuracy on FaceForensics++ and SiW-Mv2, then the attribution of attack cues to face-recognition training is wrong. Alternatively, replace the HiLo attention module with a plain CNN of similar capacity on the same Block-5 features; if accuracy does not drop, the claim that HiLo is the key mechanism is unsupported.

Watch

Extended reading notes

Core claim

The central discovery is that intermediate-stage features of a face-recognition-trained Swin Transformer already encode the local texture and noise cues needed to separate live faces from both physical and digital forgeries. By appending the unified attack detection (UAD) module at the sixth block of Stage 3 (index 5) of a frozen backbone, the model reaches 97.2% accuracy on FaceForensics++ for digital attacks and 86.8% on SiW-Mv2 for physical attacks, while keeping face-matching accuracy on par with a standalone recognizer. This shows that face recognition and spoof detection can be unified without the feature-sharing mismatch that hurts earlier dual-head designs.

Load-bearing premise

The claim depends on the assumption that a face-recognition-trained frozen backbone keeps enough attack-specific low-level detail in its middle layers, so a lightweight head trained only on augmented UniAttack data can detect attacks from other datasets.

Editorial extensions

If this is right

  • A single frozen face recognizer with a small extra head could replace the usual three separate models (face verifier, physical anti-spoofing detector, deepfake detector), cutting computational overhead on resource-limited devices.
  • Because the backbone stays frozen after face-recognition training, adding attack detection does not degrade face-matching accuracy, so security upgrades would not require re-tuning identity verification.
  • The model detects unseen physical attacks (makeup, partial masks, silicone masks) and unseen digital manipulations (StarGAN, StyleGAN, PGGAN, FaceApp) with average accuracies above 80% and 72%, respectively, suggesting it generalizes beyond its training attack types.
  • The block-sweep result gives a practical design rule: probe intermediate stages of a frozen backbone to find the abstraction level where spoof cues survive, rather than defaulting to final-layer features.

Reading between the lines

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

  • The same frozen-recognizer-plus-intermediate-head recipe may transfer to other domains where a model trained for one task inadvertently learns low-level cues useful for another, such as detecting adversarial perturbations or compression artifacts in images.
  • The paper does not ablate the SPSC and SDSC augmentations; if removing them drops performance substantially, the augmentation strategy—not the HiLo architecture—would be the main carrier of the reported gains.
  • Because the attack head outputs a single live/spoof probability, the model cannot tell physical from digital attacks; a three-way classification head could be added without changing the architecture if that distinction is needed.
  • The comparison baselines in Tables 5 and 6 are trained under different protocols, often on in-domain data, so the 'equivalent performance' claim should be read as accuracy-level parity rather than a strictly controlled benchmark comparison.
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

4 major / 4 minor

Summary. The paper proposes a unified model for face recognition and joint physical-digital spoofing attack detection. A frozen Swin Transformer backbone is fine-tuned on CASIA-WebFace for face representation and matching (FRM) using an ArcFace head at the final stage, while a lightweight HiLo-attention and CNN-based unified attack detection (UAD) head is appended to an intermediate block of Stage 3 and trained on the augmented UniAttack dataset. The authors evaluate FRM on FF++ and CelebDF, and UAD cross-dataset on FF++, SiW-Mv2, and MSU-MFSD, plus unknown physical attacks from SiW-Mv2 and unknown digital attacks from DFFD. They report accuracy/EER numbers and compare them with state-of-the-art face recognizers, deepfake detectors, and physical spoof detectors, concluding that the unified model achieves equivalent performance to SOTA across datasets and attack types.

Significance. If the empirical claims were established, the paper would make a useful contribution: a single frozen backbone with a lightweight auxiliary spoof-detection head could reduce computational overhead while preserving face-matching performance. The strengths are the clearly described architecture (backbone frozen after FRM training, UAD head with HiLo attention and a ~1.2M-parameter CNN), the systematic sweep over Stage 3 blocks (Table 4), and the inclusion of unknown-attack evaluations. However, the significance is contingent on the evaluation being protocol-matched and free of test-set-based model selection; as written, the headline SOTA-parity claims are not supported by the reported experiments.

major comments (4)
  1. [Section 5, Tables 5 and 6] The comparisons are not protocol-matched. The UAD head is trained on augmented UniAttack data and evaluated on self-selected test subsets (20,000 live/20,000 spoof for FF++; 5,000/5,000 for SiW-Mv2; 175/525 for MSU-MFSD), whereas the cited baselines are trained on the target datasets' official training partitions (e.g., deepfake detectors on FF++ train, CDCN on SiW-Mv2 train). Under these conditions a higher accuracy number does not demonstrate equivalence to SOTA; the paper must either include baselines trained under the identical cross-dataset protocol or retrain the proposed model under the baselines' protocols and report the corresponding metrics.
  2. [Section 5, Table 4] The choice of Stage 3 Block 5 is made by evaluating all 18 UAD heads on the same FF++, SiW-Mv2, and MSU-MFSD test sets that later produce the headline results (97.2% and 86.8% in Tables 5 and 6). Because the block index is a free parameter selected on the test data, the reported numbers are best-of-18 maxima and not the expected accuracy of a pre-specified unified model. A held-out validation split (or a nested selection procedure) must be used to select the block, and the selection rule must be stated before the reported test numbers are generated.
  3. [Section 5, Table 2] The FRM comparison is not protocol-matched either. The 99.43% accuracy on FF++ is compared with 99.53% from Deng et al. (2019) and 98.96% from Al-Refai et al. (2023), which are obtained on different test sets and with different pair-selection procedures; a fair comparison requires evaluating all models on the same verification pairs and the same threshold or metric.
  4. [Section 5, Tables 7 and 8] The unknown-attack results are reported without comparison to any cross-dataset baseline, so the statement that performance is 'at par with' [35, 36] is not supported by the data in the paper. The authors should include the corresponding baselines or explicitly state the comparison protocol used for those references.
minor comments (4)
  1. [Table 1] The row label 'FRAM' should be 'FRM'.
  2. [Section 5, Table 4] The text refers to 'the sixth block of Stage 3' and also 'Stage 3 Block 5'; because Table 4 indexes blocks from 0, the block selected should be denoted consistently (e.g., 'Block 5 (0-indexed)' or 'the sixth block').
  3. [Tables 5 and 6] Tables 5 and 6 contain reference errors: the entry 'Wang et al. (2019) Eff.B1 + LSTM' is cited as [5], which is Bittner et al., and 'Seferbekov' is cited as [33], which is the Swin Transformer paper; these citations need to be corrected or replaced.
  4. [Section 4.1] The paper does not report the number of identities in the CASIA-WebFace training set or the exact verification threshold used for cosine similarity; adding these details would improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

UAD parity claim is partially circular: the reported FF++/SiW-Mv2 accuracies are best-of-18 selections made on the same test sets used for the headline comparison.

  1. fitted input called prediction [Section 5, UAD Evaluation; Table 4 and following paragraph; results reused in Tables 5-6.]
    "Among all intermediate blocks, the sixth block of Stage 3 (i.e., Stage 3 Block 5) demonstrated the best performance across the evaluated datasets, making it most suitable for UAD tasks. ... Therefore the results from the UAD module appended at the sixth block of Stage 3 (i.e., Stage 3 Block 5) are used as performance metrics for further analysis in this study."

    Table 4 evaluates 18 independent UAD classifiers on FF++, SiW-Mv2, and MSU-MFSD. The paper then selects Block 5 because it 'demonstrated the best performance across the evaluated datasets' and uses that block's results as the model's reported performance in Tables 5-6. Thus the same test sets are used both to choose the model and to produce the headline numbers; the reported 97.2% and 86.8% accuracies are maxima over 18 models evaluated on the target benchmarks, not the expected performance of a pre-specified architecture. The later claim of SOTA-equivalent performance is therefore partly an artifact of test-set model selection.

full rationale

The paper's derivation is empirical rather than analytic, so there is no equation-level self-definitional circularity. The frozen Swin + FRM head and the UAD head are trained on separate datasets; that part is not circular. The load-bearing UAD comparison is compromised by one evaluation-circular step: Table 4 sweeps 18 heads on the very test sets (FF++, SiW-Mv2, MSU-MFSD) and Section 5 chooses Block 5 by best accuracy, then Tables 5-6 report that same chosen accuracy as the model's performance. This is model selection on the test data, so the headline 97.2%/86.8% are best-of-18 selections rather than expected performance of a pre-specified architecture. I do not count the protocol mismatch against SOTA baselines (Tables 5-6) as circularity; it is a comparability/correctness risk. Self-citations [35,36,59] are used only as comparison baselines and are not load-bearing for the architecture. Overall, there is partial circularity in the UAD evaluation, but the recognition component and cross-dataset training remain independent, so the score is moderate.

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

The central claim rests on one free architectural parameter (the chosen block index) and several domain assumptions about feature transferability, augmentation realism, and benchmark comparability. No new theoretical entities are introduced.

free parameters (1)
  • UAD head block index = Stage 3, Block 5 (sixth block)
    The intermediate block for the UAD head was selected as the one with the best test performance among 18 independently trained heads. This is a free architectural parameter effectively tuned on the evaluation data (Section 5, Table 4).
assumptions (3)
  • domain assumption Features from an intermediate block of a frozen Swin Transformer, trained only for face recognition, are sufficient for joint physical and digital spoof detection.
    The backbone is frozen before UAD training, and UAD heads are attached only to Stage 3 and final features (Section 3.4).
  • domain assumption SPSC and SDSC augmentation applied to the UniAttack dataset produces training samples representative of real physical and digital attacks in other datasets.
    UAD is trained solely on augmented UniAttack and evaluated on FF++, SiW-Mv2, MSU-MFSD, and DFFD without domain adaptation or target-data fine-tuning (Sections 3.3 and 4.1).
  • ad hoc to paper The test datasets used for comparison are protocol-comparable to the datasets used by the cited state-of-the-art baselines.
    Tables 5 and 6 compare cross-dataset results to same-dataset baselines as if the protocols were equivalent, which is an implicit and unverified assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unified Face Matching and Physical-Digital Spoofing Attack Detection." pith.science (2026). https://pith.science/paper/PSWRAA2T

@misc{pith2026250109635,
  author       = {Pith},
  title        = {Pith review of: Unified Face Matching and Physical-Digital Spoofing Attack Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSWRAA2T}},
  note         = {Machine review of arXiv:2501.09635}
}
read the original abstract

Face recognition technology has dramatically transformed the landscape of security, surveillance, and authentication systems, offering a user-friendly and non-invasive biometric solution. However, despite its significant advantages, face recognition systems face increasing threats from physical and digital spoofing attacks. Current research typically treats face recognition and attack detection as distinct classification challenges. This approach necessitates the implementation of separate models for each task, leading to considerable computational complexity, particularly on devices with limited resources. Such inefficiencies can stifle scalability and hinder performance. In response to these challenges, this paper introduces an innovative unified model designed for face recognition and detection of physical and digital attacks. By leveraging the advanced Swin Transformer backbone and incorporating HiLo attention in a convolutional neural network framework, we address unified face recognition and spoof attack detection more effectively. Moreover, we introduce augmentation techniques that replicate the traits of physical and digital spoofing cues, significantly enhancing our model robustness. Through comprehensive experimental evaluation across various datasets, we showcase the effectiveness of our model in unified face recognition and spoof detection. Additionally, we confirm its resilience against unseen physical and digital spoofing attacks, underscoring its potential for real-world applications.

Figures

Figures reproduced from arXiv: 2501.09635 by the authors.

Figure 1
Figure 1. Unified architecture illustrating the integration of the FRM head at the output of the Swin Transformer backbone for face [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the data augmentation process for a live [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. UAD module leveraging Stage 3 intermediate features along with HiLo attention and a cascaded CNN for joint physical-digital attack detection. High-frequency and low-frequency blocks use an attention module defined as a scaled dot-product function, where query, key, and value matrices are denoted as Q, K, and V, respectively. Given Nid training samples and C unique identities in the dataset, the loss function is defi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 50 canonical work pages

  1. [1]

    Uncertainty-aware face embed- ding with contrastive learning for open-set evaluation

    Kyeongjin Ahn, Seungeon Lee, Sungwon Han, Cheng Yaw Low, and Meeyoung Cha. Uncertainty-aware face embed- ding with contrastive learning for open-set evaluation. IEEE Transactions on Information Forensics and Security, 2024. 3

    work page 2024

  2. [2]

    A unified model for face matching and presentation attack detection using an ensemble of vision transformer features

    Rouqaiah Al-Refai and Karthik Nandakumar. A unified model for face matching and presentation attack detection using an ensemble of vision transformer features. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 662–671, 2023. 2, 3, 6, 7, 8

    work page 2023

  3. [3]

    A robust framework for spoofing detection in faces using deep learn- ing

    Shefali Arora, MPS Bhatia, and Vipul Mittal. A robust framework for spoofing detection in faces using deep learn- ing. The Visual Computer, 38(7):2461–2472, 2022. 8

    work page 2022

  4. [4]

    Belhumeur, Joao P Hespanha, and David J

    Peter N. Belhumeur, Joao P Hespanha, and David J. Krieg- man. Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on pattern analysis and machine intelligence, 19(7):711–720, 1997. 3

    work page 1997

  5. [5]

    Late or earlier information fusion from depth and spectral data? large-scale digital surface model refinement by hybrid-cgan

    Ksenia Bittner, Peter Reinartz, and Marco Korner. Late or earlier information fusion from depth and spectral data? large-scale digital surface model refinement by hybrid-cgan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , June

    work page

  6. [6]

    Face recognition with learning-based descriptor

    Zhimin Cao, Qi Yin, Xiaoou Tang, and Jian Sun. Face recognition with learning-based descriptor. In 2010 IEEE Computer society conference on computer vision and pattern recognition, pages 2707–2714. IEEE, 2010. 3

    work page 2010

  7. [7]

    Adversarial examples are not easily detected: Bypassing ten detection methods

    Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM workshop on artificial intelli- gence and security, pages 3–14, 2017. 2

    work page 2017

  8. [8]

    Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation

    Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789–8797,

Show all 63 references
  1. [9]

    Xception: Deep learning with depthwise separable convolutions

    Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 1251–1258, 2017. 8

    2017

  2. [10]

    Id-reveal: Identity-aware deepfake video detection

    Davide Cozzolino, Andreas R ¨ossler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision , pages 15108– 15117, 2021. 8

    2021

  3. [11]

    On the detection of digital face manipulation

    Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K Jain. On the detection of digital face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5781–5790, 2020. 6, 8

    2020

  4. [12]

    Look locally infer globally: A generalizable face anti-spoofing approach

    Debayan Deb and Anil K Jain. Look locally infer globally: A generalizable face anti-spoofing approach. IEEE Transac- tions on Information Forensics and Security, 16:1143–1157,

  5. [13]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4690–4699, 2019. 5, 6, 7

    2019

  6. [14]

    Surveillance face anti-spoofing

    Hao Fang, Ajian Liu, Jun Wan, Sergio Escalera, Chenxu Zhao, Xu Zhang, Stan Z Li, and Zhen Lei. Surveillance face anti-spoofing. IEEE Transactions on Information Forensics and Security, 2023. 1

    2023

  7. [15]

    Unified physical-digital face attack detection

    Hao Fang, Ajian Liu, Haocheng Yuan, Junze Zheng, Dingheng Zeng, Yanhong Liu, Jiankang Deng, Sergio Escalera, Xiaoming Liu, Jun Wan, et al. Unified physical-digital face attack detection. arXiv preprint arXiv:2401.17699, 2024. 2, 5, 6

    2024 arXiv

  8. [16]

    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. 5

    2022

  9. [17]

    Joint physical-digital facial attack detection via simulating spoofing clues

    Xianhua He, Dashuang Liang, Song Yang, Zhanlong Hao, Hui Ma, Binjie Mao, Xi Li, Yao Wang, Pengfei Yan, and Ajian Liu. Joint physical-digital facial attack detection via simulating spoofing clues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2024

  10. [18]

    Attgan: Facial attribute editing by only changing what you want

    Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only changing what you want. IEEE transactions on image pro- cessing, 28(11):5464–5478, 2019. 1

    2019

  11. [19]

    Pose adapted shape learning for large-pose face reenactment

    Gee-Sern Jison Hsu, Jie-Ying Zhang, Huang Yu Hsiang, and Wei-Jie Hong. Pose adapted shape learning for large-pose face reenactment. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 7413–7422, 2024. 3

    2024

  12. [20]

    Adversarial attacks on neural network policies

    Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284, 2017. 2

    2017 arXiv

  13. [21]

    Handbook of face recognition , volume 1

    Anil K Jain and Stan Z Li. Handbook of face recognition , volume 1. Springer, 2011. 1, 2

    2011

  14. [22]

    A survey on 3d mask presentation attack detection and countermeasures

    Shan Jia, Guodong Guo, and Zhengquan Xu. A survey on 3d mask presentation attack detection and countermeasures. Pattern recognition, 98:107032, 2020. 1

    2020

  15. [23]

    Liveness detection in computer vision: Transformer-based self-supervised learn- ing for face anti-spoofing

    Arman Keresh and Pakizar Shamoi. Liveness detection in computer vision: Transformer-based self-supervised learn- ing for face anti-spoofing. arXiv preprint arXiv:2406.13860,

    arXiv

  16. [24]

    Celeb-df: A large-scale challenging dataset for deep- fake forensics

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deep- fake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 3207– 3216, 2020. 5

    2020

  17. [25]

    Robust clip-based detector for exposing diffusion model-generated images

    Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2

    2024 arXiv

  18. [26]

    Cross-ethnicity face anti- spoofing recognition challenge: A review

    Ajian Liu, Xuan Li, Jun Wan, Yanyan Liang, Sergio Es- calera, Hugo Jair Escalante, Meysam Madadi, Yi Jin, Zhuoyuan Wu, Xiaogang Yu, et al. Cross-ethnicity face anti- spoofing recognition challenge: A review. IET Biometrics, 10(1):24–43, 2021. 1, 2

    2021

  19. [27]

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

    Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z Li. Casia-surf cefa: A benchmark for multi- modal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 1179–1187, 2021. 1

    2021

  20. [28]

    Multi-modal face anti-spoofing attack detection challenge at cvpr2019

    Ajian Liu, Jun Wan, Sergio Escalera, Hugo Jair Escalante, Zichang Tan, Qi Yuan, Kai Wang, Chi Lin, Guodong Guo, Isabelle Guyon, et al. Multi-modal face anti-spoofing attack detection challenge at cvpr2019. In Proceedings of the IEEE CVPRW, pages 0–0, 2019. 1, 2

    2019

  21. [29]

    3d high-fidelity mask face presentation attack detection challenge

    Ajian Liu, Chenxu Zhao, Zitong Yu, Anyang Su, Xing Liu, Zijian Kong, Jun Wan, Sergio Escalera, Hugo Jair Escalante, Zhen Lei, et al. 3d high-fidelity mask face presentation attack detection challenge. In Proceedings of the IEEE/CVF in- ternational conference on computer vision...

  22. [30]

    Contrastive context-aware learning for 3d high-fidelity mask face presentation attack detection

    Ajian Liu, Chenxu Zhao, Zitong Yu, Jun Wan, Anyang Su, Xing Liu, Zichang Tan, Sergio Escalera, Junliang Xing, Yanyan Liang, et al. Contrastive context-aware learning for 3d high-fidelity mask face presentation attack detection. IEEE Transactions on Information Forensics and Se...

    2022

  23. [31]

    Gabor feature based classification using the enhanced fisher linear discriminant model for face recognition

    Chengjun Liu and Harry Wechsler. Gabor feature based classification using the enhanced fisher linear discriminant model for face recognition. IEEE Transactions on Image processing, 11(4):467–476, 2002. 3

    2002

  24. [32]

    An identity authentication method combining liveness detection and face recognition

    Shuhua Liu, Yu Song, Mengyu Zhang, Jianwei Zhao, Shi- hao Yang, and Kun Hou. An identity authentication method combining liveness detection and face recognition. Sensors, 19(21):4733, 2019. 3

    2019

  25. [33]

    Proceedings of the ieee/cvf international conference on computer vision

    Zhaoyang Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Proceedings of the ieee/cvf international conference on computer vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021. 8

    2021

  26. [34]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 1, 2, 3, 6

    2021

  27. [35]

    Proactive deep- fake detection using gan-based visible watermarking

    Aakash Varma Nadimpalli and Ajita Rattani. Proactive deep- fake detection using gan-based visible watermarking. ACM Transactions on Multimedia Computing, Communications and Applications, 20(11):1–27, 2024. 8

    2024

  28. [36]

    Social media authentication and combating deepfakes using semi-fragile invisible image watermarking

    Aakash Varma Nadimpalli and Ajita Rattani. Social media authentication and combating deepfakes using semi-fragile invisible image watermarking. Digital Threats, 5(4), Dec

  29. [37]

    Fast vision transformers with hilo attention

    Zizheng Pan, Jianfei Cai, and Bohan Zhuang. Fast vision transformers with hilo attention. Advances in Neural Infor- mation Processing Systems, 35:14541–14554, 2022. 2, 4

    2022

  30. [38]

    Deep face recognition

    Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015 . British Machine Vision Association, 2015. 3

    2015

  31. [39]

    Face recognition methods & applications

    Divyarajsinh N Parmar and Brijesh B Mehta. Face recognition methods & applications. arXiv preprint arXiv:1403.0485, 2014. 1

    2014 arXiv

  32. [40]

    Deepfake generation and detection: A benchmark and survey

    Gan Pei, Jiangning Zhang, Menghan Hu, Zhenyu Zhang, Chengjie Wang, Yunsheng Wu, Guangtao Zhai, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. arXiv preprint arXiv:2403.17881, 2024. 2

    2024

  33. [41]

    A review of transformer-based models for computer vision tasks: Captur- ing global context and spatial relationships

    Gracile Astlin Pereira and Muhammad Hussain. A review of transformer-based models for computer vision tasks: Captur- ing global context and spatial relationships. arXiv preprint arXiv:2408.15178, 2024. 1

    2024 arXiv

  34. [42]

    Swinface: a multi-task trans- former for face recognition, expression recognition, age esti- mation and attribute estimation

    Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task trans- former for face recognition, expression recognition, age esti- mation and attribute estimation. IEEE Transactions on Cir- cuits and Systems for Video Technology, 2023. 3

    2023

  35. [43]

    Deep learning for un- derstanding faces: Machines may be just as good, or better, than humans

    Rajeev Ranjan, Swami Sankaranarayanan, Ankan Bansal, Navaneeth Bodla, Jun-Cheng Chen, Vishal M Patel, Car- los D Castillo, and Rama Chellappa. Deep learning for un- derstanding faces: Machines may be just as good, or better, than humans. IEEE Signal Processing Magazine, 35(1):...

    2018

  36. [44]

    Facedancer: Pose-and occlusion-aware high fidelity face swapping

    Felix Rosberg, Eren Erdal Aksoy, Fernando Alonso- Fernandez, and Cristofer Englund. Facedancer: Pose-and occlusion-aware high fidelity face swapping. In Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 3454–3463, 2023. 1

    2023

  37. [45]

    Faceforen- sics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019. 5

    2019

  38. [46]

    Facenet: A unified embedding for face recognition and clus- tering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815–823, 2015. 3

    2015

  39. [47]

    Deep learning model for deep fake face recognition and detection

    ST Suganthi, Mohamed Uvaze Ahamed Ayoobkhan, Nebo- jsa Bacanin, K Venkatachalam, Hub ´alovsk`y ˇStˇep´an, Tro- jovsk`y Pavel, et al. Deep learning model for deep fake face recognition and detection. PeerJ Computer Science, 8:e881,

  40. [48]

    Part-based face recognition with vision transformers

    Zhonglin Sun and Georgios Tzimiropoulos. Part-based face recognition with vision transformers. arXiv preprint arXiv:2212.00057, 2022. 3

    2022 arXiv

  41. [49]

    Deepface: Closing the gap to human-level perfor- mance in face verification

    Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level perfor- mance in face verification. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 1701–1708, 2014. 3

    2014

  42. [50]

    Catface: Cross-attribute-guided transformer with self-attention distillation for low-quality face recognition

    Niloufar Alipour Talemi, Hossein Kashiani, and Nasser M Nasrabadi. Catface: Cross-attribute-guided transformer with self-attention distillation for low-quality face recognition. IEEE Transactions on Biometrics, Behavior, and Identity Science, 2024. 3

    2024

  43. [51]

    Face liveness detection from a single image with sparse low rank bilin- ear discriminative model

    Xiaoyang Tan, Yi Li, Jun Liu, and Lin Jiang. Face liveness detection from a single image with sparse low rank bilin- ear discriminative model. In Computer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings...

    2010

  44. [52]

    Presen- tation attack detection: an analysis of spoofing in the wild (siw) dataset using deep learning models

    Niraj Thapa, Meenal Chaudhari, and Kaushik Roy. Presen- tation attack detection: an analysis of spoofing in the wild (siw) dataset using deep learning models. Discover Artificial Intelligence, 3(1):30, 2023. 8

    2023

  45. [53]

    Deep face recognition: A survey

    Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215–244, 2021. 1, 2

    2021

  46. [54]

    Spoofing at- tack detection in face recognition system using vision trans- former with patch-wise data augmentation

    Kota Watanabe, Koichi Ito, and Takafumi Aoki. Spoofing at- tack detection in face recognition system using vision trans- former with patch-wise data augmentation. In 2022 Asia- Pacific Signal and Information Processing Association An- nual Summit and Conference (APSIPA ASC) , ...

    2022

  47. [55]

    Face spoof detection with image distortion analysis

    Di Wen, Hu Han, and Anil K Jain. Face spoof detection with image distortion analysis. IEEE Transactions on Information Forensics and Security, 10(4):746–761, 2015. 6

    2015

  48. [56]

    Quality network based on enhanced confidence- aware and adaptive margin for facial recognition

    Yihang Wu. Quality network based on enhanced confidence- aware and adaptive margin for facial recognition. In2024 4th International Conference on Machine Learning and Intelli- gent Systems Engineering (MLISE) , pages 352–355. IEEE,

  49. [57]

    Learn- ing face representation from scratch

    Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learn- ing face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 5

    2014 arXiv

  50. [58]

    Searching 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. Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5295– 5...

    2020

  51. [59]

    Codeit: Contrastive data-efficient transform- ers for deepfake detection

    John Zakkam, Umarani Jayaraman, Subin Sahayam, and Ajita Rattani. Codeit: Contrastive data-efficient transform- ers for deepfake detection. In International Conference on Pattern Recognition, pages 62–77. Springer, 2025. 2, 8

    2025

  52. [60]

    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. 2, 6

    2016

  53. [61]

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

    Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z Li. Casia-surf: A large-scale multi-modal benchmark for face anti-spoofing. IEEE Transactions on Biometrics, Behavior, and Identity Science, 2(2):182–193, 2020. 1, 2

    2020

  54. [62]

    A dataset and benchmark for large-scale multi-modal face anti-spoofing

    Shifeng Zhang, Xiaobo Wang, Ajian Liu, Chenxu Zhao, Jun Wan, Sergio Escalera, Hailin Shi, Zezheng Wang, and Stan Z Li. A dataset and benchmark for large-scale multi-modal face anti-spoofing. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition...

    2019

  55. [63]

    Face transformer for recognition

    Yaoyao Zhong and Weihong Deng. Face transformer for recognition. arXiv preprint arXiv:2103.14803, 2021. 2, 3

    2021 arXiv

Pith tools

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