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

The DOI matches its Crossref record; 51 of the 52 references it cites were found in a registry or an index; 9 of the 24 sources its claims cite were checked for support.

Improving Generalization of Deepfake Detection With Data Farming and Few-Shot Learning

Open access
Korshunov, Pavel, Marcel, Sebastien · 2022 · IEEE Transactions on Biometrics, Behavior, and Identity Science

Abstract

Recent advances in automated video and audio editing tools, generative adversarial networks (GANs), and social media allow creation and fast dissemination of high quality tampered videos, which are generally called deepfakes. Typically, in these videos, a face is swapped with someone else’s using GANs. Accessible open source software and apps for the face swapping led to a wide and rapid dissemination of the generated deepfakes, posing a significant technical challenge for their detection and filtering. In response to the threat, which deepfake videos can pose to our trust in video evidence, several large datasets of deepfake videos and several methods to detect them were proposed recently. However, the proposed methods suffer from a problem of overfitting on the training data and the lack of the generalization across different databases and the generative models. Therefore, in this paper, we investigate the techniques for improving the generalization of deepfake detection methods that can be employed in practical settings. We have selected two popular state of the art deepfake detectors: based on Xception and EfficientNet models, and we use five databases: from Google and Jigsaw, FaceForensics++, DeeperForensics, Celeb-DF, and our own publicly available large dataset DF-Mobio. To improve generalization, we apply different augmentation strategies used during training, including a proposed aggressive ‘data farming’ technique based on random patches. We also tested two few-shot tuning methods, when either a first convolutional layer or a last layer of a pre-trained model is tuned on 100 seconds from a training set of the test database. The experimental results clearly expose the generalization problem of deepfake detection methods, since the accuracy drops significantly when a model is trained on one dataset and evaluated on another. However, the silver lining is that an aggressive augmentation during training and a few-shot tuning on the test database can improve the accuracy of the detection methods in a cross-database scenario. As a side observation, we show the importance of database selection for training and evaluation, as FaceForensics++ is found to be better to use for training, while DeeperForensics is found to be significantly more challenging as a test database.

Overview

CoChat Summary

TL;DR

This paper evaluates Xception and EfficientNet deepfake detectors across five databases and shows severe cross-database generalization failure, with deepfakes misclassified as real rising from about 7% on average in same-database tests to well above 80% cross-database. It finds that aggressive “data farming” augmentation plus first-layer few-shot tuning on the target database improves cross-database HTER, while FaceForensics++ is best for training and DeeperForensics is most challenging for testing.

Key findings

  • Same-database performance is high, but cross-database performance collapses: the conclusion reports that the percentage of deepfake videos misclassified as real can go from a desirable 7% on average in a single-database scenario to well above 80% in the cross-database scenario. In Table 3, cross-database HTER values are roughly 38–46, with many FAR values above 80%.
  • The best generalization strategy is data farming (`augm-patches`: face augmentation plus background patches) combined with tuning the first layer on the test database. In Table 3, Xception with `augm-patches` and `first layer` tuning reaches HTER 38.14 and FAR 68.34; EfficientNet with the same strategy reaches HTER 39.69 and FAR 73.76. Removing DeeperForensics lowers Xception `augm-patches` + `first layer` to HTER 31% and FAR 50%, from 38% and 68%.
  • Last-layer tuning is a mixed bag and is not recommended: it sometimes improves and sometimes worsens FAR/HTER. For same-database evaluation, the best configurations are faces-only with no tuning or at most last-layer tuning and no augmentation, but these overfit and perform poorly cross-database.
  • Database choice matters: FaceForensics++ is the best database for training, DF-Mobio is next best, and DeeperForensics is the most challenging test database, with HTER averaging at 50% and FAR at nearly 99% independent of training or tuning strategies.

Methods

  • Two detectors were used: XceptionNet and EfficientNet variant B4. Five databases were used: FaceForensics++ (Deepfakes, Face Shifter, Face Swap subsets), Google and Jigsaw, DeeperForensics, Celeb-DF, and DF-Mobio. Table 1 reports, for example, FaceForensics++ with 1000 original videos and 3000 deepfakes, DeeperForensics with 1000 original and 2000 deepfakes, Celeb-DF with 590 original and 5639 deepfakes, and DF-Mobio with 31,950 original and 14,546 deepfakes.
  • Training used 299×299 input images, batch size 32, Adam optimizer with learning rate 0.0002, 20 epochs, early stopping on validation loss, and cross-entropy loss. The study trained 30 original non-tuned models, produced 240 tuned models, and evaluated 270 models total, yielding 630 sets of evaluation scores.
  • Input variability strategies were `face-only`, `face-augm` (hue, brightness, contrast, saturation, flipping, JPEG quality 50–90), and `augm-patches` data farming (adding a background patch for each genuine face, effectively doubling the genuine class). Few-shot tuning strategies were `none`, `first layer`, and `last layer`, using roughly 100 seconds or 5 clips of 10 seconds from each class from another database’s training set. Evaluation used FAR, FRR, HTER, and AUC, with the threshold θ_far set at 10% FAR on the original training database’s validation set.

Limitations

  • The paper states that detailed analysis of all results is not feasible in scope, so it focuses on general trends; DF-Mobio’s public availability is pending journal publication; the Facebook Deepfake Detection Challenge dataset was excluded due to strict licensing; and data/training fusion is left for future work.
  • The paper states that generalization remains far from solved: even with aggressive data farming and first-layer tuning, cross-database FAR values are close to 70%, DeeperForensics is smaller and models may overfit more on it, and last-layer tuning is not recommended.
  • Evident from the text: the evaluation covers only two architectures and five databases, fixes the threshold from the original training database, limits few-shot tuning to about 100 seconds, and lets DeeperForensics strongly affect aggregate cross-database averages.

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

Integrity signals

Registry

Crossref · 10.1109/tbiom.2022.3143404

Cites

51 of 51 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
From the PDF
Image 2
From the PDF
(b) DF-Mobio
(b) DF-MobioFrom the PDF
(d) Celeb-DF
(d) Celeb-DFFrom the PDF
Fig. 2 Different augmentation techniques used in the experiments for increasing the variability of the training data: a) use detected faces as is with no augmentation, b) apply typical augmentation techniques to faces c) use a more drastic data farming that combines faces, augmentation, and background patches.
Fig. 2 Different augmentation techniques used in the experiments for increasing the variability of the training data: a) use detected faces as is with no augmentation, b) apply typical augmentation techniques to faces c) use a more drastic data farming that combines faces, augmentation, and background patches.From the PDF
Image 6
From the PDF
Image 7
From the PDF
Image 8
From the PDF
Image 9
From the PDF
(e) DF End-to-end
(e) DF End-to-endFrom the PDF
(f) DF Reenact
(f) DF ReenactFrom the PDF
(a) Original 1
(a) Original 1From the PDF
(a) Face, no augmentation
(a) Face, no augmentationFrom the PDF
Fig. 6 ROC curves with the corresponding AUC value of Xception model trained on Google, Celeb-DF, and DF-Mobio databases and evaluated on the test set of Celeb-DF.
Fig. 6 ROC curves with the corresponding AUC value of Xception model trained on Google, Celeb-DF, and DF-Mobio databases and evaluated on the test set of Celeb-DF.From the PDF
Image 15
From the PDF
Image 16
From the PDF
Image 17
From the PDF
Image 18
From the PDF
Image 19
From the PDF
(d) FaceForensics++
(d) FaceForensics++From the PDF
Image 21
From the PDF
(g) tuned, DeeperForensics
(g) tuned, DeeperForensicsFrom the PDF
(h) tuned, Google
(h) tuned, GoogleFrom the PDF

Verification

Claims the paper makes

7 Valid · 2 Not valid · 7 Not checked · 47 Own results and interpretations

28 more claims

Create a free account to see all 63 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 15
  • It may state two separate claims as one. 21

Create a free account to see these claims.

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.

2 of 2

Citations

Cites

Show all 52 references

Research works printed but not linked: 3

  1. mama, 2019, Towards Deepfake Detection that Actually Works
  2. 10.1109/CVPRW50498.2020.00329 · DOI
  3. dufour, 2019, Contributing data to deepfake detection research

Cited by

See all 50

Metadata

Authors
Korshunov, P., & Marcel, S.
Affiliations
  • Idiap Research Institute (CH)
  • Biometrics Security and Privacy Group, Idiap Research Institute, Martigny, Switzerland
Venue
IEEE Transactions on Biometrics, Behavior, and Identity Science
Published
2022-07
Volume
4
Issue
3
Pages
386-397
ISSN
2637-6407 (ISSN-L)
OpenAlex
W4205421468
Cited by
50
References
52
Citations by year
2026: 8 · 2025: 17 · 2024: 14 · 2023: 7 · 2022: 4
FWCI
3.44
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Biometrics Behavior and Identity Science (publishedVersion)
  • IRIS (submittedVersion)
Funders
  • Hasler Foundation’s VERIFAKE Project
  • Swiss Center for Biometrics Research and Testing
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 detection, cross-dataset generalization, data farming, few-shot learning, cross-database evaluation, face swapping, generative adversarial networks, Xception, EfficientNet, FaceForensics++, Celeb-DF, database selection, overfitting

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 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (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 and gpt-6-sol).

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: Improving Generalization of Deepfake Detection With Data Farming and Few-Shot Learning. https://app.cochat.ai/papers/10.1109/tbiom.2022.3143404
The paper
Korshunov, P., & Marcel, S. (2022). Improving Generalization of Deepfake Detection With Data Farming and Few-Shot Learning. IEEE Transactions on Biometrics, Behavior, and Identity Science. https://doi.org/10.1109/tbiom.2022.3143404

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.