Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 69 of the 82 references it cites were found in a registry or an index; 4 of the 8 sources its claims cite were checked for support.

Deepfake Detection: Analyzing Model Generalization Across Architectures, Datasets, and Pre-Training Paradigms

Open access
Khan, Sohail Ahmed, Dang-Nguyen, Duc-Tien·2024·IEEE Access

Abstract

As deepfake technology gains traction, the need for reliable detection systems is crucial. Recent research has introduced various deep learning-based detection systems, yet they exhibit limitations in generalizing effectively across diverse data distributions that differ from the training data. Our study focuses on understanding the generalization challenges by exploring specific aspects such as deep learning model architecture, pre-training strategy and datasets. Through a comprehensive comparative analysis, we evaluate multiple supervised and self-supervised deep learning models for deepfake detection. Specifically, we evaluate eight supervised deep learning architectures and two transformer-based models pre-trained using self-supervised strategies (DINO, CLIP) on four different deepfake detection benchmarks (FakeAVCeleb, CelebDF-V2, DFDC and FaceForensics++). Our analysis includes intra-dataset and inter-dataset evaluations, examining the best performing models, generalisation capabilities and impact of augmentations. We also investigate the trade-off between model size, efficiency and performance. Our main goal is to provide insights into the effectiveness of different deep learning architectures (transformers, CNNs), training strategies (supervised, self-supervised) and deepfake detection benchmarks. Through our extensive analysis, we established that Transformer models outperform CNN models in deepfake detection. Also, we show that FaceForensics++ and DFDC datasets equip models with comparably better generalisation capabilities, as compared to FakeAVCeleb and CelebDF-V2 datasets. Our analysis also show that image augmentations can be helpful in achieving better performance, at least for the Transformer models.

Overview

CoChat Summary

TL;DR

This paper comparatively evaluates supervised CNN/Transformer architectures and self-supervised DINO/CLIP ViT models for deepfake detection across four benchmarks, with intra- and inter-dataset tests. The abstract reports that Transformer models outperform CNNs, that FaceForensics++ and DFDC support better generalisation than FakeAVCeleb and CelebDF-V2, and that image augmentations can help Transformer models.

Key findings

  • Transformer models outperform CNN models in deepfake detection, according to the abstract’s stated conclusion.
  • FaceForensics++ and DFDC datasets equip models with comparably better generalisation capabilities than FakeAVCeleb and CelebDF-V2.
  • Image augmentations can be helpful for achieving better performance, at least for the Transformer models.
  • The paper evaluates eight supervised deep learning architectures and two transformer-based models pre-trained with self-supervised strategies (DINO, CLIP) on FakeAVCeleb, CelebDF-V2, DFDC, and FaceForensics++.

Methods

  • Design: comprehensive comparative analysis with intra-dataset and inter-dataset evaluations; models are trained on each of four deepfake detection benchmarks and tested on the same dataset and on the remaining datasets. Models are also trained twice, first without image augmentations and then with various image augmentations.
  • Data: four benchmarks — FakeAVCeleb, CelebDF-V2, DFDC, and FaceForensics++. Image models use a single cropped face image `[3, 224, 224]`; video models use a tensor of 8 consecutive cropped face images `[8, 3, 224, 224]`. Table 1 reports balanced train/validation/test face-frame counts, e.g., FakeAVCeleb 47,808 real and 47,808 fake training frames, CelebDF-V2 50,000 each, DFDC 50,000 each, and FaceForensics++ 50,000 each.
  • Analysis: supervised models update both feature extractor and classification head; self-supervised DINO and CLIP models freeze the feature extractor and train only a new classification head. For DINO, features are extracted from the last four encoder blocks; for CLIP, from the last encoder block. The study also examines model size, efficiency, and performance trade-offs.

Limitations

  • Paper-stated: DFDC is huge, so only a subset is used for training/evaluation; audio features are not used; FakeAVCeleb uses only two subsets (FakeVideo/FakeAudio and RealVideo/RealAudio); FakeAVCeleb lacks predefined train/validation/test splits, requiring manual partitioning; datasets are imbalanced, addressed by extracting balanced face frames; FaceForensics++ experiments use high-quality videos only.
  • Paper-stated: the smaller FakeAVCeleb training/validation sets are assumed to have minimal impact, supported by observations from training with even fewer frames (approximately 25k real and 25k fake frames).
  • Evident from the provided text: the supplied full text is truncated before the results/discussion section, so the paper’s numerical performance outcomes are not available in the provided material.

Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/access.2023.3348450

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 3 Not valid · 2 Not checked · 8 Own results and interpretations

1 more claim

Create a free account to see all 14 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 3
  • Probably not a substantial claim. 6

Create a free account to see these claims.

Citations

Cites

Show all 82 references

Research works printed but not linked: 7

  1. Thompson, 2023, Making Deepfakes Gets Cheaper and Easier Thanks to A.I
  2. Dosovitskiy, An image is worth 16✗16 words: Transformers for image recognition at scale
  3. Dufour, 2019, Contributing Data to Deepfake Detection Research
  4. 2022, Faceswap Github
  5. 2022, Deepfakes Github
  6. Radford, 2019, Language models are unsupervised multitask learners, OpenAI Blog, 1, 9
  7. 2021, Introducing ChatGPT

Cited by

See all 53

Metadata

Authors
Khan, S. A., & Dang-Nguyen, D.-T.
Affiliations
  • University of Bergen (NO)
  • Department of Information Science and Media Studies, University of Bergen, Bergen, MediaFutures, Norway
Venue
IEEE Access
Published
2024
Volume
12
Pages
1880-1908
ISSN
2169-3536 (ISSN-L)
OpenAlex
W4390422207
Cited by
53
References
82
Citations by year
2026: 23 · 2025: 24 · 2024: 6
FWCI
4.79
Type
journal article
Language
en
Open access
Locations
  • IEEE Access (publishedVersion, cc-by)
  • NORA - Norwegian Open Research Archives (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Industry Partners
  • Research Council of Norway, MediaFutures
  • Research Centre for Responsible Media Technology and Innovation, through the Centers for Research-Based Innovation Scheme (309339)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
Keywords
deepfake detection, model generalization, self-supervised pre-training, DINO, CLIP, transformer models, CNN models, FaceForensics++, DFDC, image augmentation, model efficiency, benchmark comparison

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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: Deepfake Detection: Analyzing Model Generalization Across Architectures, Datasets, and Pre-Training Paradigms. https://app.cochat.ai/papers/10.1109/access.2023.3348450
The paper
Khan, S. A., & Dang-Nguyen, D.-T. (2024). Deepfake Detection: Analyzing Model Generalization Across Architectures, Datasets, and Pre-Training Paradigms. IEEE Access. https://doi.org/10.1109/access.2023.3348450

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.