pith. sign in

arxiv: 2605.18238 · v1 · pith:BPOECJASnew · submitted 2026-05-18 · 💻 cs.CV

Non-Colliding Biometric Identities for Digital Entities: Geometry, Capacity, and Million-Scale Virtual Identity Provisioning

Yuyang Ji , Yixuan Shen , Anil Jain , Xiaoming Liu , Feng Liu This is my paper

Pith reviewed 2026-05-20 10:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords biometric identity provisioningvirtual face identitiesembedding manifold gapsnon-colliding identitiesface image generationdigital entitiesrepulsion allocation
0
0 comments X

The pith

Real face identities leave unclaimed gaps in embedding space that can host at least 10 million non-colliding virtual identities realizable as photorealistic images.

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

The paper establishes that real human faces occupy only a low-dimensional subspace on the embedding hypersphere, so the manifold itself contains usable gaps for placing virtual identity points. These points are allocated through repulsion to stay separated from all real identities and from each other, with the allocation remaining valid even after additional real identities are enrolled later. The resulting virtual embeddings are turned into images by a generator trained to synthesize outside the real-face distribution, producing a million-scale set of usable face images. This geometry-based provisioning directly supports biometric-style identity for digital entities that must coexist with human recognition systems.

Core claim

Because real face identities occupy a low-dimensional subspace of the embedding hypersphere, the remaining gaps within the real face manifold can be used to allocate virtual identity embeddings that never collide with any enrolled real identity and maintain inter-class separability. These embeddings are obtained by repulsion-based placement that is not limited by any fixed capacity bound, and they are realized as high-fidelity images by GapGen, a generator trained with a curriculum that progressively moves synthesis into the identified gaps.

What carries the argument

Repulsion-based allocation of points placed in the unclaimed gaps inside the real face manifold on the embedding hypersphere.

If this is right

  • Virtual identities can be provisioned at scales exceeding any practical enrollment of real identities without an upper bound.
  • Once allocated, the virtual set stays non-colliding with all future real identities that are added.
  • The same embeddings support multiple protocols including virtual verification, cross-reality matching, and real-versus-virtual detection.
  • A virtual counterpart dataset can be built that mirrors standard face benchmarks while adding cross-reality tasks.

Where Pith is reading between the lines

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

  • Existing face recognition pipelines could treat virtual identities as an additional class without retraining the core embedding model.
  • The gap-packing approach might extend to other biometric modalities such as voice or gait where similar manifold structure is observed.
  • If the generator continues to work at larger scales, digital entities could receive unique biometric credentials that integrate directly with human-centric security systems.

Load-bearing premise

A curriculum-trained generator can synthesize high-fidelity face images from points in those gaps without creating collisions or losing separability when the real identity gallery grows.

What would settle it

Add new real identities to the original gallery and verify whether any of the previously provisioned virtual embeddings now fall inside the decision boundary of a real identity or produce matching scores above the verification threshold.

Figures

Figures reproduced from arXiv: 2605.18238 by Anil Jain, Feng Liu, Xiaoming Liu, Yixuan Shen, Yuyang Ji.

Figure 1
Figure 1. Figure 1: Should a digital entity have a distinctive face? [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: BIP pipeline. Left: Repulsion-based allocation provisions V={v1, . . . , vN } satisfying cos(vj , ci)<τ and cos(vj , vj ′ )<τ , scaling to |V|=10M with zero observed collision. Middle: GapGen renders each s ∈ V into a 1024×1024 face image x˜=G(s), producing 1M virtual identity images and the v-LFW benchmark. Right: v-LFW supports four protocols spanning virtual face verification, cross-reality matching, re… view at source ↗
Figure 3
Figure 3. Figure 3: Repulsion-based virtual identity allocation. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with synthetic face generators. Four randomly sampled identities per method. GapGen produces photorealistic 1024×1024 portrait faces with natural texture and coherent lighting, free from blurring, distortion, and uncontrolled backgrounds of the baselines. is not capacity-limited at any scale tested. At stricter thresholds (smaller τ ), Inter-Sep degrades more noticeably at 10M (e.g.,… view at source ↗
Figure 5
Figure 5. Figure 5: Open-world non-collision. At α≥3.0, collision is near zero in all scales. Open-World Non-Collision as Real Galleries Grow. The hard checks guarantee non-collision against R by construc￾tion, but provide no guarantee against real identities enrolled after provisioning. To stress-test this, we use 180K held-out identities from WebFace4M [42] (disjoint from R) and fix |V| at one million virtual identities, me… view at source ↗
Figure 6
Figure 6. Figure 6: ROC curve for ArcFace ResNet-100 on IJB-B 1:1 verification (8.01M pairs; 10,270 genuine / 8.00M impostor; AUC = 99.49%). Red circles mark the six operating points used as BIP threshold candidates; τ denotes the cosine similarity threshold achieving the stated FAR. We use τ = 0.391 as the primary operating point in all main experiments, corresponding to FAR ≈ 2 × 10−5 and TAR = 94.30% on IJB-B. At this oper… view at source ↗
Figure 7
Figure 7. Figure 7: PCA of Glint360K identity centroids (N=360,232, ArcFace ResNet-100, d=512). Left (eigenvalue spectrum): λk decays sharply after k≈300, with λ512 = 2.4×10−5 five orders of magnitude below λ1 = 3.55; effective rank = 238.4. Right (cumulative explained variance): 50% of variance is captured within k=93 components, 90% within k=238, 95% within k=269, and 99% within k=306. C.2 Spherical Geometry Preliminaries S… view at source ↗
Figure 8
Figure 8. Figure 8: GapGen training pipeline. Real and virtual steps are interleaved at mixing ratio β during fine-tuning. Real step (left): real face x ∈ X is encoded by the frozen face encoder ϕ to produce target e ∗ = ϕ(x); GapGen renders x˜ = G(e ∗ ), supervised by Ldenoise + λidLRT + λpercLperc. Virtual step (right): a provisioned embedding s ∈ V is fed directly as condition with no paired ground-truth image; supervision… view at source ↗
Figure 9
Figure 9. Figure 9: v-LFW visual examples. Each row shows three images rendered from a single virtual identity s ∈ V at α=4, varying pose, lighting, and style while preserving identity. E.2 IAPCT: Identity-Anchored Patch Consistency Transformer Design. IAPCT augments the frozen ArcFace backbone with a transformer head that interrogates whether local patch statistics at each intermediate layer are consistent with the global id… view at source ↗
Figure 10
Figure 10. Figure 10: t-SNE of real and virtual identities. Real centroids R are shown as a gray density background. Each virtual identity in V is coloured by maxcj∈R cos(s, cj ) (red: close to a real cluster; blue: located in a low-density gap). Increasing α shifts virtual identities from regions dense with real centroids (α=2, mean max-cos 0.344) toward low-density gaps (α=5, mean max-cos 0.288). No internal collapse is obse… view at source ↗
Figure 11
Figure 11. Figure 11: Sample virtual identities from BIP. Forty randomly sampled virtual identities rendered by GapGen at 1024×1024, spanning diverse demographics, age, and appearance. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
read the original abstract

Digital entities such as AI agents and humanoid robots increasingly operate alongside real humans, yet their identity infrastructure is based on credentials rather than embodied biometric identity. We introduce Biometric Identity Provisioning (BIP), a new problem and solution framework that addresses: given an enrollment gallery of real human identities, provision virtual identities that are non-colliding with every enrolled identity, maintain sufficient inter-class separability, and are realizable as high-fidelity face images. The key geometric insight is that real face identities occupy a low-dimensional subspace of the embedding hypersphere, leaving no residual subspace for virtual identities. Hence, virtual identities must instead be allocated as unclaimed gaps within the real face manifold itself. BIP is therefore a constrained packing problem: available gaps vastly exceed any foreseeable enrollment scale, and provisioned identities remain non-colliding even as new real identities are subsequently enrolled. Grounded in this geometry, our repulsion-based allocation is not bounded by any fixed provisioning count; we demonstrate 10M non-colliding virtual identity embeddings against a gallery of 360K real identities. Realizing these embeddings as face images requires a generator that operates outside the training distribution of real face images; we introduce GapGen, a gap-aware generator trained with a curriculum that progressively extends synthesis into non-colliding regions, validated at 1M photorealistic virtual face images. We further construct v-LFW, a virtual counterpart to LFW face dataset, with protocols for virtual face verification, cross-reality matching, real-vs-virtual detection, and unified recognition and detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces Biometric Identity Provisioning (BIP), a framework for allocating non-colliding virtual face identities for digital entities given a gallery of real identities. It claims that real face embeddings occupy a low-dimensional subspace of the embedding hypersphere, leaving gaps that can be filled via repulsion-based allocation to produce at least 10M non-colliding virtual embeddings against 360K real identities. These are realized as 1M photorealistic images using a curriculum-trained GapGen generator that extends synthesis outside the real distribution. The work also presents the v-LFW dataset with protocols for virtual verification, cross-reality matching, and detection.

Significance. If the geometric allocation and GapGen synthesis are rigorously validated with quantitative metrics, this could provide a scalable approach to biometric identity management for AI agents and robots, addressing a growing need for embodied virtual identities distinct from real humans. The low-dimensional subspace insight and large-scale provisioning numbers would represent a notable contribution to face embedding geometry and generative modeling in computer vision, particularly if the non-collision guarantees hold under gallery growth.

major comments (3)
  1. [Abstract] Abstract: The central claims of demonstrating 10M non-colliding virtual identity embeddings against 360K real identities and realizing 1M photorealistic images via GapGen are presented without any quantitative metrics, error analysis, baseline comparisons, collision rate measurements, or separability margins. This absence directly undermines verification of the non-collision and photorealism guarantees that support the headline provisioning results.
  2. [Abstract] The repulsion-based allocation and GapGen curriculum are introduced as new components, yet the claimed provisioning counts (10M embeddings, 1M images) are not reduced to quantities derived from fitted parameters or empirical measurements on the same data; this creates a circularity risk for the capacity claims.
  3. [Abstract] The load-bearing assumption that GapGen maps allocated embedding gaps back to images whose re-embedded vectors remain non-colliding and photorealistic (without drift exceeding the separability margin) lacks supporting analysis of post-generation embedding distances, collision checks, or fidelity degradation as the real gallery scales.
minor comments (1)
  1. The notation for 'Biometric Identity Provisioning (BIP)' and 'GapGen' is introduced without a clear early definition or comparison to related packing or generative methods in the literature.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and indicate planned revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims of demonstrating 10M non-colliding virtual identity embeddings against 360K real identities and realizing 1M photorealistic images via GapGen are presented without any quantitative metrics, error analysis, baseline comparisons, collision rate measurements, or separability margins. This absence directly undermines verification of the non-collision and photorealism guarantees that support the headline provisioning results.

    Authors: We agree that the abstract, due to its length constraints, does not convey the supporting quantitative details. The experimental sections of the manuscript report collision rates, separability margins, baseline comparisons, and photorealism metrics. We will revise the abstract to include a concise summary of these key quantitative results to better support the central claims. revision: yes

  2. Referee: [Abstract] The repulsion-based allocation and GapGen curriculum are introduced as new components, yet the claimed provisioning counts (10M embeddings, 1M images) are not reduced to quantities derived from fitted parameters or empirical measurements on the same data; this creates a circularity risk for the capacity claims.

    Authors: The 10M and 1M figures are produced by running the repulsion allocation on the 360K gallery and then synthesizing a subset with GapGen; they are not presupposed. We will add an explicit derivation subsection that shows how the counts follow from the fitted subspace parameters and empirical allocation runs on the same data, eliminating any appearance of circularity. revision: yes

  3. Referee: [Abstract] The load-bearing assumption that GapGen maps allocated embedding gaps back to images whose re-embedded vectors remain non-colliding and photorealistic (without drift exceeding the separability margin) lacks supporting analysis of post-generation embedding distances, collision checks, or fidelity degradation as the real gallery scales.

    Authors: We acknowledge that stronger post-generation verification is needed to confirm the assumption holds under scaling. We will expand the experimental section with additional quantitative analysis of re-embedding distances, collision checks after synthesis, and fidelity degradation curves as gallery size increases. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation introduces independent allocation logic and generator

full rationale

The paper's core chain begins with an empirical geometric observation that real-face embeddings occupy a low-dimensional subspace on the hypersphere, from which it logically follows that virtual identities are placed in the remaining gaps via a newly defined repulsion-based packing procedure. This allocation is executed directly on the 360K real gallery to produce the 10M non-colliding points, without the counts being recovered from any fitted parameter of the same data. GapGen is introduced as a separate curriculum-trained model whose training objective and validation (1M images) are defined independently of the allocation counts. No self-citation is invoked as a load-bearing uniqueness theorem, no ansatz is smuggled, and no known result is merely renamed; the provisioning guarantees and v-LFW protocols rest on the explicit construction and synthesis steps rather than reducing to the inputs by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claims rest on the geometric premise that real faces leave usable gaps inside their manifold and on the untested ability of a curriculum generator to realize those gaps as photorealistic images.

axioms (1)
  • domain assumption Real face identities occupy a low-dimensional subspace of the embedding hypersphere, leaving usable gaps for virtual identities.
    Stated as the key geometric insight enabling the entire BIP framework.
invented entities (2)
  • Biometric Identity Provisioning (BIP) no independent evidence
    purpose: Framework for provisioning non-colliding virtual identities from real enrollment galleries
    Newly defined problem and solution approach.
  • GapGen no independent evidence
    purpose: Generator trained to synthesize faces from non-colliding gap embeddings
    Introduced to realize the allocated embeddings as images.

pith-pipeline@v0.9.0 · 5825 in / 1441 out tokens · 40105 ms · 2026-05-20T10:47:56.764520+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 2 internal anchors

  1. [1]

    Partial FC: Training 10 million identities on a single machine

    Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial FC: Training 10 million identities on a single machine. InICCV, 2021

    work page 2021

  2. [2]

    Digiface-1m: 1 million digital face images for face recognition

    Gwangbin Bae, Martin de La Gorce, Tadas Baltrušaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. InWACV, 2023

    work page 2023

  3. [3]

    Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model

    Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. InICCV, 2023

    work page 2023

  4. [4]

    Sface: Privacy- friendly and accurate face recognition using synthetic data

    Fadi Boutros, Marco Huber, Patrick Siebke, Tim Rieber, and Naser Damer. Sface: Privacy- friendly and accurate face recognition using synthetic data. InIJCB, 2022

    work page 2022

  5. [5]

    Securing autonomous AI agents

    Cloud Security Alliance. Securing autonomous AI agents. https:// cloudsecurityalliance.org/artifacts/securing-autonomous-ai-agents , Febru- ary 2026. Accessed April 29, 2026

    work page 2026

  6. [6]

    Id-reveal: Identity-aware deepfake video detection

    Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. InICCV, 2021

    work page 2021

  7. [7]

    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. InCVPR, 2019

    work page 2019

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InICLR, 2021

    work page 2021

  9. [9]

    Region-aware temporal inconsistency learning for deepfake video detection

    Zhihao Gu, Taiping Yao, Yang Chen, Ran Yi, Shouhong Ding, and Lizhuang Ma. Region-aware temporal inconsistency learning for deepfake video detection. InIJCAI, 2022

    work page 2022

  10. [10]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. InNeurIPS, 2017

    work page 2017

  11. [11]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020

    work page 2020

  12. [12]

    Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller

    Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007

    work page 2007

  13. [13]

    Billion-scale similarity search with GPUs

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535–547, 2021

    work page 2021

  14. [14]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InCVPR, 2019

    work page 2019

  15. [15]

    AdaFace: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. AdaFace: Quality adaptive margin for face recognition. InCVPR, 2022

    work page 2022

  16. [16]

    DCFace: Synthetic face generation with dual condition diffusion model

    Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. DCFace: Synthetic face generation with dual condition diffusion model. InCVPR, 2023

    work page 2023

  17. [17]

    VIGFace: Virtual identity generation for privacy-free face recognition dataset

    Minsoo Kim, Min-Cheol Sagong, Gi Pyo Nam, Junghyun Cho, and Ig-Jae Kim. VIGFace: Virtual identity generation for privacy-free face recognition dataset. InICCV, 2025

    work page 2025

  18. [18]

    SELFI: Se- lective fusion of identity for generalizable deepfake detection.arXiv preprint arXiv:2506.17592, 2025

    Younghun Kim, Minsuk Jang, Myung-Joon Kwon, Wonjun Lee, and Changick Kim. SELFI: Se- lective fusion of identity for generalizable deepfake detection.arXiv preprint arXiv:2506.17592, 2025. 10

    work page arXiv 2025

  19. [19]

    Preserving fairness generalization in deepfake detection

    Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. InCVPR, 2024

    work page 2024

  20. [20]

    AI-Face: A million-scale demographi- cally annotated ai-generated face dataset and fairness benchmark

    Li Lin, Santosh, Mingyang Wu, Xin Wang, and Shu Hu. AI-Face: A million-scale demographi- cally annotated ai-generated face dataset and fairness benchmark. InCVPR, 2025

    work page 2025

  21. [21]

    Spatial-phase shallow learning: rethinking face forgery detection in frequency domain

    Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: rethinking face forgery detection in frequency domain. InCVPR, 2021

    work page 2021

  22. [22]

    Sphereface: Deep hypersphere embedding for face recognition

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Sphereface: Deep hypersphere embedding for face recognition. InCVPR, 2017

    work page 2017

  23. [23]

    What is Microsoft Entra Agent ID? https://learn.microsoft.com/en-us/ entra/agent-id/what-is-microsoft-entra-agent-id , April 2026

    Microsoft. What is Microsoft Entra Agent ID? https://learn.microsoft.com/en-us/ entra/agent-id/what-is-microsoft-entra-agent-id , April 2026. Accessed April 29, 2026

    work page 2026

  24. [24]

    Arc2face: A foundation model for ID-consistent human faces

    Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for ID-consistent human faces. InECCV, 2024

    work page 2024

  25. [25]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. InECCV, 2020

    work page 2020

  26. [26]

    Synface: Face recognition with synthetic data

    Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. InICCV, 2021

    work page 2021

  27. [27]

    Faceforensics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. InICCV, 2019

    work page 2019

  28. [28]

    HyperFace: Generating synthetic face recognition datasets by exploring face embedding hypersphere

    Hatef Otroshi Shahreza and Sébastien Marcel. HyperFace: Generating synthetic face recognition datasets by exploring face embedding hypersphere. InICLR, 2025

    work page 2025

  29. [29]

    Detecting deepfakes with self-blended images

    Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, 2022

    work page 2022

  30. [30]

    Identity management for agentic ai: The new frontier of authorization, authentication, and security for an ai agent world,

    Tobin South, Subramanya Nagabhushanaradhya, Ayesha Dissanayaka, Sarah Cecchetti, George Fletcher, Victor Lu, Aldo Pietropaolo, Dean H Saxe, Jeff Lombardo, Abhishek Maligehalli Shivalingaiah, et al. Identity management for agentic AI: The new frontier of authorization, authentication, and security for an AI agent world.arXiv preprint arXiv:2510.25819, 2025

    work page arXiv 2025

  31. [31]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InICML, 2019

    work page 2019

  32. [32]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InNeurIPS, 2017

    work page 2017

  33. [33]

    Cosface: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. InCVPR, 2018

    work page 2018

  34. [34]

    InstantID: Zero-shot Identity-Preserving Generation in Seconds

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Iarpa janus benchmark- b face dataset

    Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark- b face dataset. InCVPRW, 2017

    work page 2017

  36. [36]

    Vec2face+ for face dataset generation.arXiv preprint arXiv:2507.17192, 2025

    Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face+ for face dataset generation.arXiv preprint arXiv:2507.17192, 2025

    work page arXiv 2025

  37. [37]

    Vec2face: Scaling face dataset generation with loosely constrained vectors

    Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face: Scaling face dataset generation with loosely constrained vectors. InICLR, 2025. 11

    work page 2025

  38. [38]

    Identity- driven multimedia forgery detection via reference assistance

    Junhao Xu, Jingjing Chen, Xue Song, Feng Han, Haijun Shan, and Yu-Gang Jiang. Identity- driven multimedia forgery detection via reference assistance. InACMMM, 2024

    work page 2024

  39. [39]

    Ucf: Uncovering common features for generalizable deepfake detection

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. InICCV, 2023

    work page 2023

  40. [40]

    Learning Face Representation from Scratch

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

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  41. [41]

    The unreason- able effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreason- able effectiveness of deep features as a perceptual metric. InCVPR, 2018

    work page 2018

  42. [42]

    WebFace260M: A benchmark unveiling the power of million-scale deep face recognition

    Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. WebFace260M: A benchmark unveiling the power of million-scale deep face recognition. InCVPR, 2021. 12 Appendix This appendix provides full details supporting the main paper. • Sec. A: Broader Impact and Ethics. • Sec. B: Verification...

    work page 2021

  43. [43]

    candid color portrait photo of a person, natural lighting

    = 1 π arcsin( √ 0.75) = 1 3, matching the120 ◦/360◦ arc fraction for a60 ◦ half-angle cap.✓ Monotonicity. µ(τ, d) isstrictly decreasingin τ: larger τ defines a smaller cap. µ(τ, d) is also strictly decreasing indby concentration of measure: for ˜d < d,µ(τ, ˜d)≥µ(τ, d). Gaussian Approximation (Intuition Only). µ(τ, d)≈Q(τ √ d) by the CLT, but in the far-ta...

    work page