CoChat

GenConViT: Deepfake Video Detection Using Generative Convolutional Vision Transformer

Open access
Deressa, Deressa Wodajo, Mareen, Hannes, Lambert, Peter, Atnafu, Solomon, Akhtar, Zahid, Van Wallendael, Glenn · 2025 · Applied Sciences

The DOI matches its Crossref record; 71 of the 76 references it cites were found in a registry or an index, and 5 are not checked yet; 1 of the 8 sources its claims cite were checked for support.

Abstract

Deepfakes have raised significant concerns due to their potential to spread false information and compromise the integrity of digital media. Current deepfake detection models often struggle to generalize across a diverse range of deepfake generation techniques and video content. In this work, we propose a Generative Convolutional Vision Transformer (GenConViT) for deepfake video detection. Our model combines ConvNeXt and Swin Transformer models for feature extraction, and it utilizes an Autoencoder and Variational Autoencoder to learn from latent data distributions. By learning from the visual artifacts and latent data distribution, GenConViT achieves an improved performance in detecting a wide range of deepfake videos. The model is trained and evaluated on DFDC, FF++, TM, DeepfakeTIMIT, and Celeb-DF (v2) datasets. The proposed GenConViT model demonstrates strong performance in deepfake video detection, achieving high accuracy across the tested datasets. While our model shows promising results in deepfake video detection by leveraging visual and latent features, we demonstrate that further work is needed to improve its generalizability when encountering out-of-distribution data. Our model provides an effective solution for identifying a wide range of fake videos while preserving the integrity of media.

Verified 2026-10-05 · 5 references not checked yet

Overview

CoChat Summary

TL;DR

GenConViT combines ConvNeXt-Swin hybrid feature extraction with Autoencoder and Variational Autoencoder latent features for deepfake video detection. It achieves strong in-domain results across five datasets, but an out-of-distribution ablation shows substantial generalization failures.

Key findings

  • GenConViT reports an average accuracy of 95.8% and an AUC of 99.3% across the tested datasets.
  • On DFDC, it reports 98.5% accuracy, 99.9% AUC, and 99.1% F1; on FF++, it reports 97.0% accuracy, 99.6% AUC, and 97.1% F1.
  • The full ensemble outperforms its individual networks: networks A and B average 94% accuracy individually, while GenConViT improves over network A by 2.2% and network B by 1.79%.
  • In leave-one-dataset-out ablation, performance drops sharply on unseen data: Celeb-DF (v2) held-out accuracy is 55.67% with 11.56% fake accuracy; DFDC held-out is 62.08% with 28.84% fake accuracy; FF++ held-out is 64.86% with 44.20% fake accuracy.

Methods

  • Architecture: two independently trained networks. Network A uses an Autoencoder plus ConvNeXt-Swin; network B uses a Variational Autoencoder plus ConvNeXt-Swin. Their real/fake predictions are averaged.
  • Data: trained and evaluated on DFDC, FF++, TrustedMedia (TM), DeepfakeTIMIT, and Celeb-DF (v2). Preprocessing yielded 1,004,810 face images split approximately 80:15:5 into 826,756 training, 130,948 validation, and 47,106 test images; video evaluation held back 3,972 videos and averaged predictions over 15 frames per video.
  • Training and analysis: Adam optimizer with learning rate 0.0001 and weight decay 0.0001, 30 epochs, batch size 32 for network A and 16 for network B, with Albumentation augmentation. Metrics include accuracy, F1, ROC, and AUC; generalization was tested via leave-one-dataset-out ablation on approximately 5% subsets of training data.

Limitations

  • The paper states that GenConViT struggles to generalize to out-of-distribution data, especially more realistic deepfakes, and that deeper or wider CNN architectures did not resolve the domain-shift drop.
  • The paper states lower performance on Celeb-DF (v2), attributed to overlapping real/fake latent embeddings in t-SNE visualization, and notes that GenConViT lacks temporal information because it processes frames independently.
  • The paper notes that the OOD ablation used randomly selected subsets (about 10k samples per dataset) rather than the full >1M-image training set, and that DFDC/FF++ test subdatasets are usually not shared, so comparisons with prior work are approximate.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/app15126622

Cites

71 of 76 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 0 Not valid · 5 Not checked · 50 Own results and interpretations

33 more claims

Create a free account to see all 56 claims and to verify the 7 references still open.

See all claims — free account
  • Probably not a substantial claim. 8
  • It may state two separate claims as one. 11
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 5

Citations

Cites

Show all 76 references

Research works printed but not linked: 14

  1. Liu, P., Lin, Y., He, Y., Wei, Y., Zhen, L., Zhou, J.T., Goh, R.S.M., and Liu, J. (2021). Automated deepfake detection. arXiv.
  2. Mirza, M., and Osindero, S. (2014). Conditional generative adversarial nets. arXiv.
  3. Kingma, D.P., and Welling, M. (2022). Auto-Encoding Variational Bayes. arXiv.
  4. Yan, Z., Sun, P., Lang, Y., Du, S., Zhang, S., Wang, W., and Liu, L. (2022). Multimodal graph learning for deepfake detection. arXiv.
  5. Li, B., Sun, J., Poskitt, C.M., and Wang, X. (2024). How Generalizable are Deepfake Image Detectors? An Empirical Study. arXiv.
  6. O’shea, K., and Nash, R. (2015). An introduction to convolutional neural networks. arXiv.
  7. Li, L., Bao, J., Yang, H., Chen, D., and Wen, F. (2020). FaceShifter: Towards High Fidelity And Occlusion Aware Face Swapping. arXiv.
  8. Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2018). Progressive Growing of GANs for Improved Quality, Stability, and Variation. arXiv.
  9. Heo, Y.J., Choi, Y.J., Lee, Y.W., and Kim, B.G. (2021). Deepfake Detection Scheme Based on Vision Transformer and Distillation. arXiv.
  10. Dolhansky, B., Howes, R., Pflaum, B., Baram, N., and Ferrer, C.C. (2019). The Deepfake Detection Challenge (DFDC) Preview Dataset. arXiv.
  11. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., and Ferrer, C.C. (2020). The DeepFake Detection Challenge (DFDC) Dataset. arXiv.
  12. Korshunov, P., and Marcel, S. (2018). DeepFakes: A New Threat to Face Recognition? Assessment and Detection. arXiv.
  13. Bazarevsky, V., Kartynnik, Y., Vakunov, A., Raveendran, K., and Grundmann, M. (2019). BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs. arXiv.
  14. Maiano, L., Casadei, F., and Amerini, I. (2025). Enhancing Abnormality Identification: Robust Out-of-Distribution Strategies for Deepfake Detection. arXiv.

Cited by

See all 46

Metadata

Authors
Deressa, D. W., Mareen, H., Lambert, P., Atnafu, S., Akhtar, Z., & Van Wallendael, G.
Affiliations
  • IDLab, Department of Electronics and Information Systems, Ghent University—imec, 9000 Gent, Belgium
  • Ghent University (BE)
  • Addis Ababa University (ET)
  • SUNY Polytechnic Institute (US)
Venue
Applied Sciences
Publisher
MDPI AG
Published
2025-06-12
Volume
15
Issue
12
Pages
6622
ISSN
2076-3417 (ISSN-L)
OpenAlex
W4411241923
Cited by
46
References
66
Citations by year
2026: 31 · 2025: 12
FWCI
14.88
Type
journal article
Language
en
Open access
Locations
  • Applied Sciences (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • Ghent University Academic Bibliography (Ghent University) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
  • Addis Ababa University Research Grant for the Adaptive Problem-Solving Research (AR/048/2021, G0A2523N)
  • Research Foundation—Flanders (AR/048/2021, G0A2523N)
  • Flemish government (AR/048/2021, G0A2523N)
  • IDLab (Ghent University—imec) (AR/048/2021, G0A2523N)
  • Flanders Innovation and Entrepreneurship (VLAIO) (AR/048/2021, G0A2523N)
  • European Union (AR/048/2021, G0A2523N)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
deepfake video detection, ConvNeXt, Swin Transformer, variational autoencoder, autoencoder, latent space distribution, visual artifacts, media integrity, out-of-distribution generalization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-04, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-04 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: GenConViT: Deepfake Video Detection Using Generative Convolutional Vision Transformer. https://app.cochat.ai/papers/10.3390/app15126622
The paper
Deressa, D. W., Mareen, H., Lambert, P., Atnafu, S., Akhtar, Z., & Van Wallendael, G. (2025). GenConViT: Deepfake Video Detection Using Generative Convolutional Vision Transformer. Applied Sciences. https://doi.org/10.3390/app15126622

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.