EdgeDoc: Hybrid CNN-Transformer Model for Accurate Forgery Detection and Localization in ID Documents

Anjith George; Sebastien Marcel

REVIEW 3 major objections 2 minor 19 references

EdgeDoc: Hybrid CNN-Transformer Model for Accurate Forgery Detection and Localization in ID Documents

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read EdgeDoc claims a hybrid CNN-transformer with noiseprint features detects and localizes forged ID documents better than baselines on the FantasyID benchmark.

desk verdict A modest, credible application paper with a concrete challenge result; the real uncertainty is whether the FantasyID evaluation supports the real-world claim, and the corrupted full text makes that unanswerable from this submission. read the letter →

arxiv 2508.16284 v1 pith:S2CA6LY2 submitted 2025-08-22 cs.CV

Anjith George , Sebastien Marcel This is my paper

classification cs.CV

keywords documentforgerydetectionlocalizationIDdocumentsnoiseprintCNN-transformerhybridKYCsecurityimageforensics

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 aims to show that a hybrid architecture—a lightweight convolutional transformer combined with auxiliary noiseprint features—can both flag a forged identity document and show where it was altered. This matters because KYC and remote-onboarding checks need to catch subtle edits that look natural to human reviewers, and localization gives investigators a concrete region to inspect. The evidence is empirical: EdgeDoc placed third in the ICCV 2025 DeepID Challenge and outperformed baseline approaches on the FantasyID dataset. If correct, the result points to a practical recipe for document forgery detection that fuses low-level sensor-noise cues with context-aware transformer processing.

What carries the argument

The load-bearing component is the EdgeDoc architecture, which fuses two input streams: the document image processed by a lightweight convolutional transformer, and auxiliary noiseprint features. Noiseprint is a per-image fingerprint of the sensor and processing pipeline that produced the image; any inserted or reworked region has an inconsistent fingerprint, so the noiseprint stream gives the transformer a localization signal aligned with tampered pixels. The design lets the transformer's self-attention combine layout and text semantics with these low-level forensics cues in one forward pass.

What would settle it

Construct a test set of forged ID documents that have been edited and then printed, re-scanned, or saved with strong JPEG/WebP recompression—manipulations common in real fraud attempts. If EdgeDoc's detection accuracy on this set falls to the level of a plain CNN baseline, the central claim that noiseprint-augmented hybrid features are decisive would be falsified.

Watch

Extended reading notes

Core claim

On its own terms, this paper's claim is that fusing auxiliary noiseprint features into a lightweight convolutional transformer yields a single model that both detects whether an ID document has been forged and localizes the altered region. The paper reports that the resulting system, EdgeDoc, placed third in the ICCV 2025 DeepID Challenge and outperformed the baseline approaches it was compared with on the FantasyID dataset. The stated reason the combination works is that the transformer captures document structure and context while the noiseprint stream exposes the traces left by splicing or re-editing, traces that are hard to see in pixel space alone.

Load-bearing premise

The paper's claim that the method works in real-world scenarios depends on the FantasyID dataset and the DeepID Challenge evaluation representing the forgeries that actually occur in KYC fraud; if that benchmark is narrow or synthetic, the reported advantage may not carry over.

Editorial extensions

If this is right

KYC and remote-onboarding providers could use EdgeDoc to flag suspect documents and automatically highlight the specific forged region for manual review.
Adding a noiseprint stream is a general recipe that strengthens a lightweight transformer for forgery detection without a separate forensics classifier.
Because detection and localization happen in one model, deployment cost stays close to a single backbone instead of an ensemble of detectors and segmenters.
The DeepID Challenge ranking is evidence that the method generalizes beyond its training distribution at least as well as the specific baselines it was compared with.

Reading between the lines

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

The gain from noiseprint is likely to shrink when forgeries are printed, re-scanned, or aggressively recompressed, because those operations destroy the sensor-noise footprint; the paper does not test this regime.
Because ID documents are highly standardized layouts, the same twin-stream architecture should transfer to passports, driver's licenses, visas, and diplomas with little change.
An ablation that removes the noiseprint branch on FantasyID would isolate how much of EdgeDoc's edge comes from the auxiliary features versus the transformer backbone; the paper does not report this split explicitly.

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.

Desk Editor's Note

read the letter

Honest take: this is a standard, workmanlike application paper with a real external validation signal. The new bit is the specific combination of a lightweight CNN-transformer with noiseprint features applied to ID-document forgery detection and localization. Both ingredients are known in forensics, so the contribution is bounded, but a new combination targeting a real KYC problem can still be a legitimate contribution. Third place in the ICCV 2025 DeepID Challenge is a concrete datum, and the FantasyID results, if they match the abstract, put the method above its baselines.

I could not inspect the body—the full text in this submission is corrupted (all replacement characters), so my read is restricted to the abstract. That means I cannot verify the architecture details, metric definitions, baseline choices, or whether the numbers come with error bars or ablations. The central claim as stated is modest: competitive rather than SOTA. So I would not call the paper unsound from what is visible.

The soft spots, in proportion: the abstract's last sentence—"highlighting its effectiveness in realworld scenarios"—overreaches relative to the evidence shown. One challenge and one dataset are not real-world deployment. If FantasyID is synthetic or dominated by classic copy-move/splicing manipulations, that is exactly where noiseprint features are strongest, and the method may not generalize to GAN- or diffusion-based forgeries, which are the emerging threat in KYC fraud. That concern is real but it is about the evidence, not necessarily the method. A good reviewer should ask the authors to report per-manipulation-type results and, ideally, evaluate on a second dataset. The "novel approach" phrasing in the abstract also oversells what is an incremental combination, but that is a minor framing issue.

Who is this for? People building document-forensics pipelines or working on the DeepID challenge line. They will get a useful baseline architecture recipe and a competitive point of reference. It does not deserve a desk reject; it deserves a normal peer review with attention to dataset representativeness and the real-world generalization language. I would not cite it in my own work before seeing the body, but I would not be surprised if it holds up.

Referee Report

3 major / 2 minor

Summary. The paper proposes EdgeDoc, a hybrid CNN-transformer model augmented with auxiliary noiseprint features for detecting and localizing forgeries in ID documents. The abstract reports that EdgeDoc placed third in the ICCV 2025 DeepID Challenge and outperforms baseline methods on the FantasyID dataset, interpreting this as evidence of effectiveness in real-world KYC scenarios. The provided manuscript body is unreadable due to widespread encoding corruption; only the abstract is intelligible. Consequently, the architecture, the noiseprint feature formulation, the training and evaluation protocols, the dataset description, and the quantitative results are not available for verification.

Significance. If substantiated, the paper would offer a useful, competitive approach for document forgery detection: the combination of a lightweight convolutional transformer with noiseprint features is a sensible direction, and an external challenge placement provides some signal of practical competitiveness. However, the significance depends entirely on the evaluation's validity and generalizability. Because the body text is corrupted and the abstract contains no numbers, no quantitative claim can currently be checked. The paper's central claim is plausible but unverified at this stage.

major comments (3)

[Full text (all body sections)] The manuscript body is unreadable: it consists of mojibake with no parseable sentences, equations, tables, or figure captions. As a result, the architecture details, the noiseprint feature computation, the training procedure, dataset statistics, baseline definitions, and metric definitions are all missing. The central claim of outperforming baselines on FantasyID cannot be verified from the supplied file. This is the load-bearing evidence for the paper and must be addressed by providing a readable manuscript.
[Abstract, paragraph 3] The abstract reports 'third place' and 'outperforms baseline approaches' but gives no numerical values, no metrics (e.g., F1, IoU, AUC), no error bars, and no comparison protocol. Even if the body were readable, the abstract alone does not support the effectiveness claim. The authors should state the quantitative results and evaluation metrics in the abstract or, at minimum, clearly in the results section.
[Abstract, paragraph 3 / Results (missing)] The claim that results 'highlight its effectiveness in realworld scenarios' generalizes from a single dataset (FantasyID). There is no visible evidence that FantasyID is representative of real-world KYC forgeries or of the manipulation types encountered in practice. The manuscript should discuss the dataset's construction (synthetic vs. real, manipulation types, document types) and should address whether noiseprint-based features are sufficient for modern generative manipulations. Without this, the real-world generalization claim is unsupported.

minor comments (2)

[Abstract, last line] The project page URL contains a space: 'https://www.idiap. ch/paper/edgedoc/'. Also, 'realworld' should be 'real-world'.
[References] The FantasyID dataset and the ICCV 2025 DeepID Challenge are not cited with full references in the visible text. Provide citations and, if available, dataset documentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical results on an external challenge and dataset, not derivations that reduce to their own inputs.

full rationale

The available evidence (abstract plus the corrupted full-text fragments) does not exhibit any circular step. The central claims are (1) EdgeDoc placed third in the ICCV 2025 DeepID Challenge and (2) it outperforms baselines on the FantasyID dataset. These are empirical claims evaluated against external benchmarks, not quantities defined in terms of the paper's own outputs. The method description combines a lightweight convolutional transformer with auxiliary noiseprint features; noiseprint is a prior, externally established feature extractor, not a parameter fitted here and then renamed as a prediction. There is no equation in the visible text showing a reported metric being equal by construction to a fitted input, no subset-fitting-then-prediction pattern, and no load-bearing self-citation chain. The abstract's phrase 'highlighting its effectiveness in realworld scenarios' is a generalizability inference that depends on the representativeness of FantasyID, but that is a concern about external validity and dataset appropriateness, not circularity. The corruption of the full text prevents independent verification of dataset construction or baseline details, but absence of verifiable detail is not evidence of circularity. Under the stated review rules, a non-finding with score 0 is the appropriate outcome.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

At the abstract level, no new physical or mathematical entities are introduced. The model's learned weights and architecture hyperparameters are fitted to data, and the reported performance is benchmarked against an external dataset, so no explicit free parameters are disclosed. The main assumptions are about dataset representativeness and the validity of noiseprint features.

free parameters (1)

Network weights and architecture hyperparameters (depth, heads, fusion weights) = Not disclosed in abstract
The model is trained on data, so its weights and architecture choices are fitted or chosen by hand. The abstract does not report them, yet the claimed superiority depends on these choices.

assumptions (2)

domain assumption The FantasyID dataset is representative of real-world ID document forgeries.
The abstract claims effectiveness in real-world scenarios based on results on FantasyID. If the dataset is not representative, the central claim does not generalize.
domain assumption Noiseprint features reliably indicate local image manipulation in ID documents.
The method's auxiliary branch relies on noiseprint traces as a signal for forgery localization. This is a known forensics assumption inherited from prior work, but it is load-bearing for the proposed architecture.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EdgeDoc: Hybrid CNN-Transformer Model for Accurate Forgery Detection and Localization in ID Documents." pith.science (2026). https://pith.science/paper/S2CA6LY2

@misc{pith2026250816284,
  author       = {Pith},
  title        = {Pith review of: EdgeDoc: Hybrid CNN-Transformer Model for Accurate Forgery Detection and Localization in ID Documents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S2CA6LY2}},
  note         = {Machine review of arXiv:2508.16284}
}

read the original abstract

The widespread availability of tools for manipulating images and documents has made it increasingly easy to forge digital documents, posing a serious threat to Know Your Customer (KYC) processes and remote onboarding systems. Detecting such forgeries is essential to preserving the integrity and security of these services. In this work, we present EdgeDoc, a novel approach for the detection and localization of document forgeries. Our architecture combines a lightweight convolutional transformer with auxiliary noiseprint features extracted from the images, enhancing its ability to detect subtle manipulations. EdgeDoc achieved third place in the ICCV 2025 DeepID Challenge, demonstrating its competitiveness. Experimental results on the FantasyID dataset show that our method outperforms baseline approaches, highlighting its effectiveness in realworld scenarios. Project page : https://www.idiap. ch/paper/edgedoc/

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Generated faces in the wild: Quantitative comparison of stable diffusion, midjourney and dall-e 2

Ali Borji. Generated faces in the wild: Quantitative comparison of stable diffusion, midjourney and dall-e 2. arXiv preprint arXiv:2210.00586, 2022

arXiv 2022
[3]

Noiseprint: A cnn-based camera model fingerprint

Davide Cozzolino and Luisa Verdoliva. Noiseprint: A cnn-based camera model fingerprint. IEEE Transactions on Information Forensics and Security, 15: 0 144--159, 2019

work page 2019
[4]

Edgeface: Efficient face recognition model for edge devices

Anjith George, Christophe Ecabert, Hatef Otroshi Shahreza, Ketan Kotwal, and S \'e bastien Marcel. Edgeface: Efficient face recognition model for edge devices. IEEE Transactions on Biometrics, Behavior, and Identity Science, 6 0 (2): 0 158--168, 2024

work page 2024
[5]

Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization

Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20606--20615, 2023

work page 2023
[6]

Fantasyid: A dataset for detecting digital manipulations of id-documents

Pavel Korshunov, Amir Mohammadi, Vidit Vidit, Christophe Ecabert, and S \' e bastien Marcel. Fantasyid: A dataset for detecting digital manipulations of id-documents. In IEEE International Joint Conference on Biometrics (IJCB), 2025 a

work page 2025
[7]

Deepid challenge of detecting synthetic manipulations in id documents

Pavel Korshunov, Vidit, Amir Mohammadi, Christophe Ecabert, Nevena Shamoska, S\' e bastien Marcel, Zeqin Yu, Ye Tian, Jiangqun Ni, Lazar Lazarevic, Renat Khizbullin, Anastasiia Evteeva, Alexey Tochin, Aleksei Grishin, Anjith George, Daniel DeAlcala, Tam\' a s Endrei, Javier Mu\ n oz Haro, Ruben Tolosana, Ruben Vera-Rodriguez, Aythami Morales, Julian Fierr...

work page 2025
[8]

Forgery-aware adaptive transformer for generalizable synthetic image detection

Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10770--10780, 2024

work page 2024

Show all 19 references

[9]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

2017 arXiv
[10]

Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications

Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. In European conference on computer vision, pages 3--20...

2022
[11]

V-net: Fully convolutional neural networks for volumetric medical image segmentation

Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565--571. Ieee, 2016

2016
[12]

Exploring a patch-wise approach for privacy-preserving fake id detection

Javier Mu \ n oz-Haro, Ruben Tolosana, Ruben Vera-Rodriguez, Aythami Morales, and Julian Fierrez. Exploring a patch-wise approach for privacy-preserving fake id detection. arXiv preprint arXiv:2504.07761, 2025

2025 arXiv
[13]

Towards universal fake image detectors that generalize across generative models

Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480--24489, 2023

2023
[14]

Few-shot learning: Expanding id cards presentation attack detection to unknown id countries

Alvaro Sanchez, Juan M Esp \' n, and Juan E Tapia. Few-shot learning: Expanding id cards presentation attack detection to unknown id countries. In 2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1--9. IEEE, 2024

2024
[15]

First competition on presentation attack detection on id card

Juan E Tapia, Naser Damer, Christoph Busch, Juan M Espin, Javier Barrachina, Alvaro S Rocamora, Kri s tof Ocvirk, Leon Alessio, Borut Batagelj, Sushrut Patwardhan, et al. First competition on presentation attack detection on id card. In 2024 IEEE International Joint Conference...

2024
[16]

Exploring multi-modal fusion for image manipulation detection and localization

Konstantinos Triaridis and Vasileios Mezaris. Exploring multi-modal fusion for image manipulation detection and localization. In International conference on multimedia modeling, pages 198--211. Springer, 2024

2024
[17]

Forensics-bench: A comprehensive forgery detection benchmark suite for large vision language models

Jin Wang, Chenghui Lv, Xian Li, Shichao Dong, Huadong Li, Kelu Yao, Chao Li, Wenqi Shao, and Ping Luo. Forensics-bench: A comprehensive forgery detection benchmark suite for large vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference,...

2025
[18]

Research on identity document image tampering detection based on texture understanding and multistream networks

Lixin Wang, Zhenjiang Li, and Wenqi Zhao. Research on identity document image tampering detection based on texture understanding and multistream networks. Journal of Electronic Imaging, 34 0 (4): 0 043018--043018, 2025 b

2025
[19]

Deep learning-based forgery attack on document images

Lin Zhao, Changsheng Chen, and Jiwu Huang. Deep learning-based forgery attack on document images. IEEE Transactions on Image Processing, 30: 0 7964--7979, 2021

2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

write newline

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

[2] [2]

Generated faces in the wild: Quantitative comparison of stable diffusion, midjourney and dall-e 2

Ali Borji. Generated faces in the wild: Quantitative comparison of stable diffusion, midjourney and dall-e 2. arXiv preprint arXiv:2210.00586, 2022

arXiv 2022

[3] [3]

Noiseprint: A cnn-based camera model fingerprint

Davide Cozzolino and Luisa Verdoliva. Noiseprint: A cnn-based camera model fingerprint. IEEE Transactions on Information Forensics and Security, 15: 0 144--159, 2019

work page 2019

[4] [4]

Edgeface: Efficient face recognition model for edge devices

Anjith George, Christophe Ecabert, Hatef Otroshi Shahreza, Ketan Kotwal, and S \'e bastien Marcel. Edgeface: Efficient face recognition model for edge devices. IEEE Transactions on Biometrics, Behavior, and Identity Science, 6 0 (2): 0 158--168, 2024

work page 2024

[5] [5]

Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization

Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20606--20615, 2023

work page 2023

[6] [6]

Fantasyid: A dataset for detecting digital manipulations of id-documents

Pavel Korshunov, Amir Mohammadi, Vidit Vidit, Christophe Ecabert, and S \' e bastien Marcel. Fantasyid: A dataset for detecting digital manipulations of id-documents. In IEEE International Joint Conference on Biometrics (IJCB), 2025 a

work page 2025

[7] [7]

Deepid challenge of detecting synthetic manipulations in id documents

Pavel Korshunov, Vidit, Amir Mohammadi, Christophe Ecabert, Nevena Shamoska, S\' e bastien Marcel, Zeqin Yu, Ye Tian, Jiangqun Ni, Lazar Lazarevic, Renat Khizbullin, Anastasiia Evteeva, Alexey Tochin, Aleksei Grishin, Anjith George, Daniel DeAlcala, Tam\' a s Endrei, Javier Mu\ n oz Haro, Ruben Tolosana, Ruben Vera-Rodriguez, Aythami Morales, Julian Fierr...

work page 2025

[8] [8]

Forgery-aware adaptive transformer for generalizable synthetic image detection

Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10770--10780, 2024

work page 2024