Pith. sign in

REVIEW 5 major objections 6 minor 73 references

DOOMGAN:High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing

T0 review · 5 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 DOOMGAN shows that visible-spectrum ocular morphing attacks can defeat current verification and detection systems, reporting 99.85% attack success at 1% FMR on VISOB and over 20 percentage point gains over baselines.

desk verdict First GAN-based visible-spectrum ocular morphing, worth refereeing, but the headline attack-success claim is contradicted by the paper's own Table 1 at 0.01% FMR. read the letter →

arxiv 2507.17158 v1 pith:3OW65KOP submitted 2025-07-23 cs.CV

Bharath Krishnamurthy , Ajita Rattani This is my paper · ORCID
classification cs.CV
keywords ocularbiometricsmorphingattackgenerativeadversarialnetworkvisiblespectrumlandmark-guidedgenerationmorphdetectionperiocularrecognition
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

This paper tries to establish that visible-spectrum ocular biometrics, the eye and surrounding skin as captured by ordinary cameras, are vulnerable to a new class of morphing attack. It introduces DOOMGAN, a generative adversarial network that blends two people's ocular images so the resulting image matches both identities in verification systems. On the VISOB test set, reported attack success reaches 99.85% at a 1% false-match rate, more than 20 percentage points above existing GAN and landmark baselines, and the attack transfers across three other datasets. The paper also claims the generated eyes are anatomically better than prior methods: elliptical iris structure improves by roughly 20% and gaze consistency by 30%, while current morph-attack detectors catch them poorly. If these claims hold, any remote authentication pipeline relying on visible-light eye images needs a defense strategy designed for this modality.

What carries the argument

The load-bearing machinery is a landmark-driven GAN. A custom landmark generator predicts 66 ocular keypoints, which are converted into Gaussian heatmaps and a landmark-feature vector; the heatmaps are concatenated into the image encoder and discriminator, while the feature vector guides the generator through an attention module focused on iris and eyebrow regions. A spectrally normalized PatchGAN discriminator, which scores local image patches, enforces local realism, and a dynamic weighting scheme continuously rescales six losses, namely adversarial, MS-SSIM, perceptual, reconstruction, identity, and identity-difference, so that anatomical structure, gaze, and dual-identity matching converge together.

What would settle it

Evaluate DOOMGAN on a dataset with manual expert annotations of iris boundaries and gaze direction, or have trained raters judge whether generated irises are anatomically regular and gaze plausible; if Lg's landmark error rises sharply outside VISOB, or if independent ratings disagree with the near-perfect IR and Gaze scores, the claim of high-fidelity, gaze-consistent morphs fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that morphing attacks, previously demonstrated for face and near-infrared iris recognition, can be mounted successfully in the visible-spectrum ocular domain, where images are uncontrolled and landmarks are hard to localize. DOOMGAN drives this by encoding ocular landmarks into Gaussian heatmaps and landmark features, guiding an attention-based generator that preserves iris boundaries and periocular texture, and adaptively reweighting six losses during training. The result is a morph that scores high structural similarity, near-perfect iris-regularity and gaze-consistency metrics (IR 0.9380, Gaze 0.9775 on VISOB), and high attack success on both a ResNet-ArcFace verification model and a ViT verification model, including cross-dataset MMPMR above 95% at 1% FMR on UFPR, UBIPr, and MICHE. The paper also reports that existing morph-attack detectors, including DenseNet, leave a large fraction of these morphs undetected, with best D-EER 12.30% on VISOB and much worse cross-dataset.

Load-bearing premise

The pipeline's success depends on the landmark generator producing accurate ocular landmarks on unseen images; the paper reports only a single MSE of 0.54 on the VISOB test set and no independent landmark benchmark, so if that detector is biased, the morphs and the gaze-consistency metric inherit the bias.

Editorial extensions

If this is right

  • Visible-light ocular verification, including mobile and unconstrained capture, is susceptible to morphing attacks, not just controlled NIR iris or face systems.
  • Current morph-attack detectors, both handcrafted-feature and deep-learning, are inadequate against this attack class, especially across datasets, so ocular-specific detection methods are needed.
  • The released dataset of over 100,000 morphed ocular images provides a common ground for studying both morph generation and detection.
  • Landmark guidance plus dynamic loss weighting appears to be the combination that produces anatomically plausible morphs, as the ablation shows large IR and Gaze gains only when both are present.
  • High cross-dataset attack success means a morph made in one acquisition environment can threaten systems deployed elsewhere.

Reading between the lines

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

  • Because the Gaze metric is computed with the same landmark generator that guides synthesis, a perfect Gaze score could reflect internal consistency rather than true anatomical gaze direction; an independent, human- or anatomy-validated gaze check would strengthen the fidelity claim.
  • The landmark generator was trained on VISOB, and the cross-dataset gains are reported on datasets with limited ethnic diversity, so behavior on more diverse ocular populations is an open question.
  • If the reported attack success generalizes, DOOMGAN is a natural red-team generator for testing ocular authentication, but the same capability is a dual-use attack tool, which raises the priority of ocular-specific morph-attack detection.
  • The Iris Irregularity metric relies on a U-Net trained on UBIPr segmentation masks, so near-perfect IR scores may partly reflect that the segmenter finds DOOMGAN's smooth iris boundaries easy to fit rather than that real anatomy is faithfully preserved.
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

5 major / 6 minor

Summary. The paper introduces DOOMGAN, a GAN-based framework for generating visible-spectrum ocular morphing attacks. The architecture combines a landmark generator, landmark-conditioned encoder, attention-guided generator, and spectrally normalized discriminator, trained with a dynamically weighted multi-loss objective that includes adversarial, MS-SSIM, perceptual, reconstruction, identity, and identity-difference terms. The authors report evaluations on VISOB, UFPR, UBIPr, and MICHE datasets, measuring image quality (SSIM), novel anatomical metrics (Iris Irregularity, Gaze consistency), attack success against two ocular verification systems (OVS-I, ResNet-50 ArcFace; OVS-II, ViT), and Morph Attack Detection evasion. They also release a large morphed ocular dataset of over 100,000 images and state that code is available. The central claims are high attack success relative to baselines, improved elliptical iris generation, and improved gaze consistency.

Significance. If the reported results hold, this is one of the first systematic studies of morphing attacks on visible-spectrum ocular biometrics, and the released dataset could be a valuable resource for the biometrics community. The paper proposes two novel evaluation metrics (Iris Irregularity and Gaze consistency) that address a real gap, since prior ocular morphing work concentrated on NIR iris or face morphing. The paper also includes a reasonable set of baselines (landmark-based and several GAN morphing models), a cross-dataset evaluation, and an ablation study of the loss components. However, the significance is currently undermined by several load-bearing evaluation issues: the headline abstract claim is contradicted by the paper's own Table 1, the OVS-I evaluation is potentially circular with the identity loss, and the Gaze metric is computed with the same landmark generator that guides the synthesis. These issues require careful revision before the contributions can be fully credited.

major comments (5)
  1. [Abstract; Table 1; Section 4.1] The abstract's claim that DOOMGAN achieves 'over 20% higher attack success rates than baseline methods under stringent thresholds' is not supported by Table 1 at the most stringent threshold. At 0.01% FMR on VISOB, DOOMGAN's MMPMR is 75.00 (OVS-I) and 86.15 (OVS-II), whereas the landmark baseline LM [45] achieves 92.50 and 96.09, respectively. Thus DOOMGAN is approximately 19% and 10% lower than LM on this primary metric at this threshold. The paper should either qualify the claim to a specific baseline class (e.g., GAN-based methods) and a specific metric (e.g., FMMPMR, or MMPMR at 1% FMR), or report the complete comparison matrix at all thresholds and operating points so that the abstract reflects the actual data.
  2. [Section 3.2, Eq. (10); Section 4, Implementation Details] The identity loss in Eq. (10) is computed with a pre-trained ocular recognition model based on ResNet-50 with ArcFace, and the OVS-I verification system used for attack-success evaluation in Section 4 is described as also based on ResNet-50 with ArcFace. The paper does not establish that these are different model instances, different training sets, or different checkpoints. If the same model is used for both training supervision and evaluation, the OVS-I results are circular and the reported attack success rates are inflated relative to a truly independent verification system. The authors should explicitly state the relation between the identity-loss network and the OVS-I evaluator, and ideally re-run the OVS-I vulnerability analysis with a verification model trained independently from the identity-loss model.
  3. [Section 3.1.1; Section 4, Gaze metric] The Gaze consistency metric is computed using landmark coordinates from the proposed Lg, the same landmark generator whose outputs guide the morph synthesis through the heatmaps and landmark features (Sections 3.1.2 and 3.1.4). The only reported accuracy of Lg is an MSE of 0.54 on the VISOB test set, with no independent benchmark and no cross-dataset landmark accuracy. Consequently, the Gaze metric measures consistency of the generated image with the model's own landmark estimates, not anatomical correctness. The authors should validate Lg against an independent landmark detector or manually annotated landmarks on multiple datasets, and preferably compute the Gaze metric with landmarks that are not part of the generation pipeline.
  4. [Table 2; Section 4.2] The cross-dataset evaluation in Table 2 reports only DOOMGAN's own scores; the text states that other GAN-based baselines 'failed to achieve reasonable FMMPMR scores' and are omitted, with only a few example numbers in prose. This makes it impossible to verify the claimed cross-dataset superiority of DOOMGAN, including whether it retains a 20% advantage over any baseline at any threshold. The authors should provide a full comparison table for all baselines on UFPR, UBIPr, and MICHE, at the same operating points, so that readers can assess the relative attack success and the effect of dataset shift.
  5. [Section 4, Metrics; Table 3] The Iris Irregularity metric relies on a U-Net iris-boundary segmenter trained on UBIPr segmentation masks, but the main evaluations are on VISOB and cross-dataset on UFPR and MICHE. The paper does not report the accuracy of this U-Net on the test datasets, nor does it report any calibration of the IR metric. Since IR is used to support the claim of '20% better elliptical iris structure generation,' the absence of a validation of the segmentation model is a load-bearing gap. A sensitivity analysis (e.g., comparison with another segmentation method or manual annotations) would help determine whether the IR gains reflect true anatomical irregularity or artifacts of the segmenter.
minor comments (6)
  1. [Table 2] In the UFPR row, the FMMPMR value at 0.1% FMR is printed as '51,18' (comma instead of decimal point); this should be corrected to 51.18.
  2. [Section 1, Contributions] The phrase 'over 1,00,000' uses an unconventional digit grouping; I suggest 'over 100,000' for consistency with the rest of the text.
  3. [Table 1] The inference time for MIPGAN is listed as '2764' without units or formatting; while the heading says milliseconds, this value appears to be a typo or should be formatted as '2764' consistently with the other entries.
  4. [Section 3.2, Eq. (11)] The identity difference loss in Eq. (11) is written asymmetrically between x2 and x1; the paper does not explain how the two source identities are ordered, which may affect the training signal. Please clarify the role of subject order.
  5. [Section 4, Datasets] The text states that 'all the GAN models and MAD defenses were trained on the VISOB dataset,' but the LM baseline uses landmarks from Lg trained on VISOB; please clarify whether the landmark generator was used to generate landmarks for the cross-dataset sets without any fine-tuning, and whether this is a possible source of domain shift.
  6. [References] The reference for [45] is listed with page numbers in the citation list; please ensure the reference format is consistent with the journal style.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: OVS-I attack success is measured with a ResNet-50/ArcFace model of the same description as the identity-loss embedder used to train DOOMGAN, and the Gaze metric is computed from the same landmark generator that conditions synthesis.

  1. fitted input called prediction [Section 3.2 (Identity Loss, Eq. 10) and Section 4 (Implementation Details, OVS models)]
    "Identity Loss: To maintain identity-related features, we utilize a pre-trained ocular recognition model based on the ResNet-50 backbone and the ArcFace loss function. The identity loss is computed using the cosine similarity of the embedding of the morphed image to the embeddings from the pair of images from the contributing subjects ... Pre-trained ocular verification (OVS) models were used to access the vulnerability of generated morphs based on the ResNet-50 backbone using the Arcface loss function [42, 41] (OVS-I)."

    LId in Eq. (10) minimizes 1 minus the average cosine similarity between f(morph) and f(x1), f(x2), where f is a ResNet-50/ArcFace ocular embedder. OVS-I, the system used to compute MMPMR in Table 1, is described with the identical architecture and loss and cites the authors' own prior OVS models [42, 41]; the paper never states that the weights differ. MMPMR at threshold t is the fraction of morphs whose cosine similarity to either subject embedding under OVS-I exceeds t. If f is OVS-I (the natural reading of the text), the reported 'attack success' on OVS-I is the very quantity maximized during training, so the high MMPMR values are partly a fitted optimization outcome rather than an independent vulnerability measurement.

  2. self definitional [Section 4 (Gaze Direction metric) with Sections 3.1.2-3.1.4 (landmark conditioning of generator)]
    "Gaze Direction (Gaze): The proposed Gaze metric measures the geometric consistency of the gaze direction in the morphed image. It ensures the generated iris position lies plausibly between the iris coordinates of the two contributing subjects ... The spatial coordinates of the two irises are obtained using our Lg, and evaluated using Euclidean distance."

    The same Lg that outputs the landmarks used to build the heatmaps H(l) and landmark features fl = Le(l) that condition the generator (Eqs. 2-4) is also used to obtain the iris coordinates for the Gaze consistency metric. The metric therefore measures how closely the generator follows the anatomical signal it was given, not whether the gaze direction is anatomically correct relative to the true subject irises. The abstract's '30% improved gaze consistency' is a self-consistency score with the model's own conditioning signal rather than an independent evaluation.

full rationale

Two localized circularities account for the score. First, the identity loss (Eq. 10) is computed with a pre-trained ResNet-50/ArcFace ocular embedder, and the OVS-I verification system used to report MMPMR is described identically and cites the authors' own prior ocular models [42, 41]; absent any statement that different weights are used, the Table 1 OVS-I numbers are at least partly the training objective rather than an independent attack measurement. Second, the proposed Gaze consistency metric uses Lg to obtain the iris coordinates while Lg's landmarks are the conditioning signal fed to the generator, making the reported gaze improvement self-referential. The IR metric, by contrast, uses a U-Net trained on UBIPr segmentation masks and is not circular. The abstract's 'over 20% higher attack success' claim, if contradicted by Table 1 at the 0.01% FMR threshold, is a correctness/consistency issue, not a circularity, and is therefore not scored here. The remainder of the pipeline (adversarial, MS-SSIM, perceptual, and reconstruction losses; cross-dataset UFPR/UBIPr/MICHE results; MAD evaluation) has independent content. The self-citations [42, 41] become load-bearing only through the identity-loss/OVS-I overlap; if the authors disclose distinct embedder weights for training and evaluation, the circularity would reduce to the Gaze metric alone.

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

The central empirical claims rest on several unverified modeling choices: the sufficiency of the landmark representation, the identity embedder used in both loss and evaluation, the landmark ground-truth generation procedure, and the U-Net used for the IR metric. Dynamic weighting and gradient penalty are standard but add hand-set constants. No invented entities are postulated.

free parameters (4)
  • Dynamic weighting adjustment rate r = 0.05
    Set empirically in Section 3.3; controls convergence speed of loss weights and affects final morph quality.
  • Initial loss weights for dynamic weighting = Equal initialization; final: 0.0964, 0.0967, 0.0965, 0.5177, 0.097, 0.097
    The weighting mechanism in Section 3.3 adaptively fits these during training. Reconstruction dominates the final weights, shaping the generated morphs.
  • Gradient penalty weight = 10
    Chosen for WGAN-GP training stability in Section 3.1.5, not derived from data.
  • Landmark and heatmap counts = 66 landmarks, 19 heatmaps
    Hand-specified in Section 3.1.2; the landmark configuration determines anatomical guidance and is also used by the Gaze metric.
assumptions (5)
  • domain assumption Sixty-six 2D landmarks (and 19 heatmaps) sufficiently encode visible-spectrum ocular anatomy for morph synthesis.
    Invoked in Section 3.1.2 and 3.1.4; the landmark-guided pipeline and the Gaze metric depend on this representational sufficiency.
  • domain assumption The pretrained ResNet-50 ArcFace ocular embedder is a valid identity measure for visible-spectrum ocular morphs.
    Used for identity losses in Eqs. 10-11 and for the OVS-I verification model in Section 4; the paper does not state whether these are distinct weights.
  • ad hoc to paper MediaPipe-Iris outputs, after manual refinement, are valid ground truth for training Lg.
    Supplementary Section 2 describes building the landmark dataset this way; no independent landmark benchmark is provided.
  • domain assumption The U-Net trained on UBIPr segmentation masks generalizes to VISOB, UFPR, and MICHE for computing Iris Irregularity.
    Section 4 uses this U-Net for the IR metric on all datasets without reporting cross-dataset segmentation accuracy.
  • standard math WGAN-GP with spectral normalization and a multi-scale patch discriminator trains stably and produces realistic images.
    Standard GAN methodology used in Sections 3.1.5 and 3.2; accepted as background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DOOMGAN:High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing." pith.science (2026). https://pith.science/paper/3OW65KOP

@misc{pith2026250717158,
  author       = {Pith},
  title        = {Pith review of: DOOMGAN:High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3OW65KOP}},
  note         = {Machine review of arXiv:2507.17158}
}
read the original abstract

Ocular biometrics in the visible spectrum have emerged as a prominent modality due to their high accuracy, resistance to spoofing, and non-invasive nature. However, morphing attacks, synthetic biometric traits created by blending features from multiple individuals, threaten biometric system integrity. While extensively studied for near-infrared iris and face biometrics, morphing in visible-spectrum ocular data remains underexplored. Simulating such attacks demands advanced generation models that handle uncontrolled conditions while preserving detailed ocular features like iris boundaries and periocular textures. To address this gap, we introduce DOOMGAN, that encompasses landmark-driven encoding of visible ocular anatomy, attention-guided generation for realistic morph synthesis, and dynamic weighting of multi-faceted losses for optimized convergence. DOOMGAN achieves over 20% higher attack success rates than baseline methods under stringent thresholds, along with 20% better elliptical iris structure generation and 30% improved gaze consistency. We also release the first comprehensive ocular morphing dataset to support further research in this domain.

Figures

Figures reproduced from arXiv: 2507.17158 by the authors.

Figure 1
Figure 1. Illustration of morphed ocular images generated by our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed DOOMGAN for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the Landmark-based and best performing [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative Ablation Study across Models. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 1
Figure 1. Figure 1: Encoder, Generator, and Discriminator Architecture of the proposed DOOMGAN. [PITH_FULL_IMAGE:figures/full_fig_p014_1.png]
Figure 2
Figure 2. Figure 2: Top: Scatter plot of similarity scores on the VISOB [PITH_FULL_IMAGE:figures/full_fig_p016_2.png]
Figure 3
Figure 3. Figure 3: Disentangled Morphs for contributing subjects across di [PITH_FULL_IMAGE:figures/full_fig_p017_3.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 69 canonical work pages

  1. [45]

    Sharma and A

    R. Sharma and A. Ross. Image-level iris morph attack. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3013–3017. IEEE, 2021. 2, 3, 6, 7, 1

    work page 2021

  2. [1]

    Real-time Pupil Tracking from Monocular Video for Digital Puppetry

    A. Ablavatski, A. Vakunov, I. Grishchenko, K. Raveen- dran, and M. Zhdanovich. Real-time pupil tracking from monocular video for digital puppetry. arXiv preprint arXiv:2006.11341, 2020. 2, 3, 1

    work page Pith review arXiv 2006

  3. [2]

    Ahonen, E

    T. Ahonen, E. Rahtu, V . Ojansivu, and J. Heikkila. Recogni- tion of blurred faces using local phase quantization. In 2008 19th international conference on pattern recognition , pages 1–4. IEEE, 2008. 3, 6, 7

    work page 2008

  4. [3]

    Alonso-Fernandez, J

    F. Alonso-Fernandez, J. Bigun, J. Fierrez, N. Damer, H. Proenc ¸a, and A. Ross. Periocular biometrics: A modality for unconstrained scenarios. Computer, 57(6):40–49, 2024. 2

    work page 2024

  5. [4]

    Arjovsky, S

    M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017. 5

    arXiv 2017

  6. [5]

    K. W. Bowyer, K. Hollingsworth, and P. J. Flynn. Image understanding for iris biometrics: A survey.Computer vision and image understanding, 110(2):281–307, 2008. 2

    work page 2008

  7. [6]

    K. I. Chang, K. W. Bowyer, and P. J. Flynn. An evaluation of multimodal 2d+ 3d face biometrics. IEEE transactions on pattern analysis and machine intelligence , 27(4):619–624,

    work page

  8. [7]

    Chavarro and M

    D. Chavarro and M. Karakaya. Blink detection for o ff-angle iris recognition using deep learning. In SoutheastCon 2024, pages 1256–1261. IEEE, 2024. 3

    work page 2024

Show all 73 references
  1. [8]

    Chen and J.-c

    L. Chen and J.-c. Xu. Optimal delaunay triangulations. Jour- nal of Computational Mathematics, pages 299–308, 2004. 1

    2004

  2. [9]

    Dalal and B

    N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05), vol- ume 1, pages 886–893. Ieee, 2005. 3, 6, 7

    2005

  3. [10]

    Damer, V

    N. Damer, V . Boller, Y . Wainakh, F. Boutros, P. Terh ¨orst, A. Braun, and A. Kuijper. Realistic face morphing attacks based on 3d modeling. In 2018 IEEE 4th International Con- ference on Identity, Security, and Behavior Analysis (ISBA), pages 1–8. IEEE, 2018. 5

    2018

  4. [11]

    Damer, M

    N. Damer, M. Fang, P. Siebke, J. N. Kolf, M. Huber, and F. Boutros. Mordi ff: Recognition vulnerability and attack detectability of face morphing attacks created by di ffusion autoencoders. In 2023 11th International Workshop on Bio- metrics and Forensics (IWBF), pages 1–6. IE...

    2023

  5. [12]

    Damer, A

    N. Damer, A. M. Saladie, A. Braun, and A. Kuijper. Mor- gan: Recognition vulnerability and attack detectability of face morphing attacks created by generative adversarial net- work. In 2018 IEEE 9th international conference on biomet- rics theory, applications and systems (BTA...

    2018

  6. [13]

    De Marsico, M

    M. De Marsico, M. Nappi, D. Riccio, and H. Wechsler. Mo- bile iris challenge evaluation (miche)-i, biometric iris dataset and protocols. Pattern Recognition Letters, 57:17–23, 2015. 5

    2015

  7. [14]

    Dosovitskiy

    A. Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6

    2010 arXiv

  8. [15]

    Ferrara, A

    M. Ferrara, A. Franco, D. Maltoni, and C. Busch. Morphing attack potential. In 2022 International workshop on biomet- rics and forensics (IWBF), pages 1–6. IEEE, 2022. 2

    2022

  9. [16]

    Grimmer and C

    M. Grimmer and C. Busch. Ladimo: Face morph generation through biometric template inversion with latent di ffusion. In 2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–7. IEEE, 2024. 3

    2024

  10. [17]

    Gulrajani, F

    I. Gulrajani, F. Ahmed, M. Arjovsky, V . Dumoulin, and A. C. Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 2

    2017

  11. [18]

    H. Guo, S. Hu, X. Wang, M.-C. Chang, and S. Lyu. Eyes tell all: Irregular pupil shapes reveal gan-generated faces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2904–2908. IEEE, 2022. 2

    2022

  12. [19]

    A. K. Jain, A. A. Ross, K. Nandakumar, and T. Swearingen. Multibiometrics. In Introduction to Biometrics, pages 289–

  13. [20]

    Jiang, Z

    S. Jiang, Z. Tao, and Y . Fu. Geometrically editable face im- age translation with adversarial networks. IEEE Transac- tions on Image Processing, 30:2771–2783, 2021. 3

    2021

  14. [21]

    Johnson, A

    J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision , pages 694–711. Springer,

  15. [22]

    Kannala and E

    J. Kannala and E. Rahtu. Bsif: Binarized statistical image features. In Proceedings of the 21st international confer- ence on pattern recognition (ICPR2012), pages 1363–1366. IEEE, 2012. 6, 7

    2012

  16. [23]

    Krishnamurthy and A

    B. Krishnamurthy and A. Rattani. Abrobocular: Adversarial benchmarking and robustness analysis of datasets and tools for ocular-based user recognition. Neurocomputing, page 130352, 2025. 1

    2025

  17. [24]

    A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther. Autoencoding beyond pixels using a learned similarity metric. In International conference on machine learning, pages 1558–1566. PMLR, 2016. 3, 6

    2016

  18. [25]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE /CVF international conference on computer vision , pages 10012– 10022, 2021. 6

    2021

  19. [26]

    Meden, P

    B. Meden, P. Rot, P. Terh ¨orst, N. Damer, A. Kuijper, W. J. Scheirer, A. Ross, P. Peer, and V .ˇStruc. Privacy–enhancing face biometrics: A comprehensive survey. IEEE Transac- tions on Information Forensics and Security, 16:4147–4183,

  20. [27]

    Nguyen, H

    K. Nguyen, H. Proenc ¸a, and F. Alonso-Fernandez. Deep learning for iris recognition: A survey.ACM Computing Sur- veys, 56(9):1–35, 2024. 2

    2024

  21. [28]

    C. N. Padole and H. Proenca. Periocular recognition: Analy- sis of performance degradation factors. In 2012 5th IAPR in- ternational conference on biometrics (ICB), pages 439–445. IEEE, 2012. 5

    2012

  22. [29]

    P ´erez, M

    P. P ´erez, M. Gangnet, and A. Blake. Poisson image edit- ing. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 577–582. 2023. 1

    2023

  23. [30]

    Pietik ¨ainen

    M. Pietik ¨ainen. Local binary patterns. Scholarpedia, 5(3):9775, 2010. 3

    2010

  24. [31]

    A. R. PN, R. Ramachandra, S. Venkatesh, K. S. Rao, P. Mitra, and R. Krishna. Morcode: Face morphing at- tack generation using generative codebooks. arXiv preprint arXiv:2410.07625, 2024. 2, 3, 6

    2024 arXiv

  25. [32]

    Proenc ¸a

    H. Proenc ¸a. Unconstrained iris recognition in visible wave- lengths. Handbook of iris recognition, pages 321–358, 2016. 2

    2016

  26. [33]

    Raghavendra, K

    R. Raghavendra, K. Raja, S. Venkatesh, and C. Busch. Face morphing versus face averaging: Vulnerability and detection. In 2017 IEEE International Joint Conference on Biometrics (IJCB), pages 555–563. IEEE, 2017. 2

    2017

  27. [34]

    K. Raja, G. Gupta, S. Venkatesh, R. Ramachandra, and C. Busch. Towards generalized morphing attack detec- tion by learning residuals. Image and Vision Computing , 126:104535, 2022. 1

    2022

  28. [35]

    Rathgeb and C

    C. Rathgeb and C. Busch. On the feasibility of creating mor- phed iris-codes. In 2017 IEEE International Joint Confer- ence on Biometrics (IJCB) , pages 152–157. IEEE, 2017. 2, 3

    2017

  29. [36]

    Rattani and R

    A. Rattani and R. Derakhshani. Ocular biometrics in the visi- ble spectrum: A survey. Image and Vision Computing, 59:1– 16, 2017. 1, 2

    2017

  30. [37]

    Rattani and R

    A. Rattani and R. Derakhshani. A survey of mobile face biometrics. Computers & Electrical Engineering, 72:39–52,

  31. [38]

    Rattani, R

    A. Rattani, R. Derakhshani, S. K. Saripalle, and V . Gotte- mukkula. Icip 2016 competition on mobile ocular biometric recognition. In 2016 IEEE international conference on im- age processing (ICIP), pages 320–324. IEEE, 2016. 3, 5

    2016

  32. [39]

    Rattani, N

    A. Rattani, N. Reddy, and R. Derakhshani. Convolutional neural networks for gender prediction from smartphone- based ocular images. Iet Biometrics, 7(5):423–430, 2018. 1

    2018

  33. [40]

    Reddy, A

    N. Reddy, A. Rattani, and R. Derakhshani. Ocularnet: deep patch-based ocular biometric recognition. In 2018 IEEE in- ternational symposium on technologies for homeland secu- rity (HST), pages 1–6. IEEE, 2018. 1

    2018

  34. [41]

    Reddy, A

    N. Reddy, A. Rattani, and R. Derakhshani. Robust subject- invariant feature learning for ocular biometrics in visible spectrum. In 2019 IEEE 10th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–9. IEEE, 2019. 1, 6

    2019

  35. [42]

    Reddy, A

    N. Reddy, A. Rattani, and R. Derakhshani. Generalizable deep features for ocular biometrics. Image and Vision Com- puting, 103:103996, 2020. 1, 6

    2020

  36. [43]

    Scherhag, A

    U. Scherhag, A. Nautsch, C. Rathgeb, M. Gomez-Barrero, R. N. Veldhuis, L. Spreeuwers, M. Schils, D. Maltoni, P. Grother, S. Marcel, et al. Biometric systems under mor- phing attacks: Assessment of morphing techniques and vul- nerability reporting. In 2017 International Confere...

    2017

  37. [44]

    Scherhag, C

    U. Scherhag, C. Rathgeb, J. Merkle, R. Breithaupt, and C. Busch. Deep face representations for di fferential mor- phing attack detection. IEEE Transactions on Information Forensics and Security, 15:3625–3639, 2020. 2

    2020

  38. [46]

    Srivastava, L

    A. Srivastava, L. Valkov, C. Russell, M. U. Gutmann, and C. Sutton. Veegan: Reducing mode collapse in gans using implicit variational learning. Advances in neural information processing systems, 30, 2017. 2

    2017

  39. [47]

    Thanh-Tung and T

    H. Thanh-Tung and T. Tran. Catastrophic forgetting and mode collapse in gans. In 2020 international joint confer- ence on neural networks (ijcnn) , pages 1–10. IEEE, 2020. 2

    2020

  40. [48]

    Venkatesh, K

    S. Venkatesh, K. Raja, R. Ramachandra, and C. Busch. On the influence of ageing on face morph attacks: Vulnerability and detection. In 2020 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2020. 6

    2020

  41. [49]

    Venkatesh, R

    S. Venkatesh, R. Ramachandra, K. Raja, and C. Busch. Face morphing attack generation and detection: A comprehen- sive survey. IEEE transactions on technology and society , 2(3):128–145, 2021. 2, 3

    2021

  42. [50]

    Venkatesh, H

    S. Venkatesh, H. Zhang, R. Ramachandra, K. Raja, N. Damer, and C. Busch. Can gan generated morphs threaten face recognition systems equally as landmark based morphs?-vulnerability and detection. In 2020 8th Interna- tional Workshop on Biometrics and Forensics (IWBF), pages 1–6...

    2020

  43. [51]

    C. Wang, J. Muhammad, Y . Wang, Z. He, and Z. Sun. To- wards complete and accurate iris segmentation using deep multi-task attention network for non-cooperative iris recog- nition. IEEE Transactions on information forensics and se- curity, 15:2944–2959, 2020. 2

    2020

  44. [52]

    Y . Wang, T. Tan, and A. K. Jain. Combining face and iris bio- metrics for identity verification. In Audio-and Video-Based Biometric Person Authentication: 4th International Confer- ence, AVBPA 2003 Guildford, UK, June 9–11, 2003 Proceed- ings 4, pages 805–813. Springer, 2003. 2

    2003

  45. [53]

    Z. Wang, E. P. Simoncelli, and A. C. Bovik. Multiscale structural similarity for image quality assessment. Confer- ence Record of the Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, 2:1398–1402, 2003. 5

    2003

  46. [54]

    Contact lenses in iris recognition

    West Virginia University. Contact lenses in iris recognition. https://researchrepository.wvu.edu/etd/4965/ ,

  47. [55]

    Yadav, C

    S. Yadav, C. Chen, and A. Ross. Synthesizing iris images us- ing rasgan with application in presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition workshops , pages 0–0, 2019. 2

    2019

  48. [56]

    Yadav and A

    S. Yadav and A. Ross. Synthesizing iris images using gener- ative adversarial networks: Survey and comparative analysis. arXiv preprint arXiv:2404.17105, 2024. 2

    2024 arXiv

  49. [57]

    L. A. Zanlorensi, R. Laroca, D. R. Lucio, L. R. Santos, A. S. Britto Jr, and D. Menotti. A new periocular dataset collected by mobile devices in unconstrained scenarios. Scientific Re- ports, 12(1):17989, 2022. 5

    2022

  50. [58]

    Zhang, I

    H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena. Self- attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR,

  51. [59]

    Zhang, R

    H. Zhang, R. Ramachandra, K. Raja, and C. Busch. Gener- alized single-image-based morphing attack detection using deep representations from vision transformer. In Proceed- ings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pages 1510–1518, 2024. 3

    2024

  52. [60]

    Zhang, S

    H. Zhang, S. Venkatesh, R. Ramachandra, K. Raja, N. Damer, and C. Busch. Mipgan—generating strong and high quality morphing attacks using identity prior driven gan. IEEE Transactions on Biometrics, Behavior, and Iden- tity Science, 3(3):365–383, 2021. 3, 4, 5, 6

    2021

  53. [61]

    Zhu and S

    Y . Zhu and S. Newsam. Densenet for dense flow. In 2017 IEEE international conference on image processing (ICIP) , pages 790–794. IEEE, 2017. 6, 7 Supplementary: DOOMGAN: High-Fidelity Dynamic Identity Obfuscation Ocular Gener- ative Morphing

    2017

  54. [63]

    As can be seen, all the existing studies on iris morphing are conducted in the near-infrared spec- trum

    Existing Methods on Iris Morphing at NIR spectrum Table 1 presents existing studies on iris morphing at the NIR spectrum. As can be seen, all the existing studies on iris morphing are conducted in the near-infrared spec- trum. At the NIR spectrum, the images are acquired in an...

  55. [64]

    To address this limitation, we propose the ocular landmark generator, Lg, specifically designed to extract the landmarks from the periocular region

    Proposed Ocular Landmark Generator The MediaPipe-Iris-based ocular landmark detector re- quires complete facial images and precise image padding ratios to function effectively, limiting its application in sce- narios where only ocular images are available. To address this limi...

  56. [65]

    The landmark-based image morphing process consists of five key steps, established in [45], discussed as follows:

    Landmark-based Ocular Image Morphing Building on our Ocular Landmark Generator’s capa- bilities, we implemented landmark-based ocular morph- ing with 38 points (19 pairs) for the LM approach. The landmark-based image morphing process consists of five key steps, established in ...

  57. [66]

    Landmark Extraction (Correspondence) : Precise landmark extraction from source and target images us- ing our Ocular Landmark Generator, with the number of landmarks serving as a tunable parameter

  58. [67]

    Delaunay Triangulation: Implementation of Delau- nay triangulation on landmark points, creating a com- prehensive mesh for proper alignment [8]

  59. [68]

    Triangle Warping: Computation of affine transforma- tions for each corresponding triangle pair between the images of two contributing subjects

  60. [69]

    Texture Blending: Smooth interpolation of pixel in- tensities within the corresponding triangles

  61. [70]

    Seamless Cloning: Application of advanced blending techniques to achieve natural transitions in the perioc- ular area [29]. Reference Domain Contribution Results Rathgeb & Busch [35] Iris (Normal- ized) Feature-level morphing using stability-based bit substitution for iris-cod...

  62. [71]

    The notations associated with the Table are represented as shown below

    DOOM-GAN Architecture and Training Al- gorithm Figure 1 shows the architecture of the Encoder and Gen- erator of the DOOMGAN model, and Table 3 shows the network flow of each layer. The notations associated with the Table are represented as shown below. • ngf: Size of feature ...

  63. [72]

    The UFPR dataset features im- ages of 1,122 subjects, acquired across 196 different mobile devices, representing a broad range of acquisition hardware

    Dataset Details The VISOB dataset comprises images from 585 subjects, captured using a variety of mobile phone cameras under di- verse lighting conditions. The UFPR dataset features im- ages of 1,122 subjects, acquired across 196 different mobile devices, representing a broad ...

  64. [73]

    Top: Scatter plot of similarity scores on the VISOB dataset for (a) OVS-I and (b) ViT OVS-II

    Analysis of Similarity Score Distributions and Image Quality Metrics (a) (b) (c) (d) Figure 2. Top: Scatter plot of similarity scores on the VISOB dataset for (a) OVS-I and (b) ViT OVS-II. Bottom: Spread of SSIM scores for (c) VISOB and (d) UFPR, for both contributing subjects...

  65. [2013]

    Graduate Theses, Dissertations, and Problem Reports, No. 4965. 2, 3

Pith tools

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