Scaling Laws for Deepfake Detection

Wenhao Wang; Jusheng Zhang; Longqi Cai; Taihong Xiao; Yuxiao Wang; Ming-Hsuan Yang

cs.CVOct 18, 2025

Scaling Laws for Deepfake Detection

Authors: Wenhao Wang, Jusheng Zhang, Longqi Cai, Taihong Xiao, Yuxiao Wang, Ming-Hsuan Yang

Organizations: University of Technology Sydney · Google DeepMind

Abstract

This paper presents a systematic study of scaling laws for the deepfake detection task. Specifically, we analyze the model performance against the number of real image domains, deepfake generation methods, and training images. Since no existing dataset meets the scale requirements for this research, we construct ScaleDF, the largest dataset to date in this field, which contains over 5.8 million real images from 51 different datasets (domains) and more than 8.8 million fake images generated by 102 deepfake methods. Using ScaleDF, we observe power-law scaling similar to that shown in large language models (LLMs). Specifically, the average detection error follows a predictable power-law decay as either the number of real domains or the number of deepfake methods increases. This key observation not only allows us to forecast the number of additional real domains or deepfake methods required to reach a target performance, but also inspires us to counter the evolving deepfake technology in a data-centric manner. Beyond this, we examine the role of pre-training and data augmentations in deepfake detection under scaling, as well as the limitations of scaling itself.The ScaleDF dataset is available at https://huggingface.co/datasets/WenhaoWang/ScaleDF.

Figures & tables

Figure 1: Illustration of ScaleDF, which is the largest deepfake detection dataset across four dimensions: number of real domains, number of deepfake methods, number of videos, and number of images. To ensure a fair comparison, for datasets that do not explicitly provide the number of images, we estimate image counts by sampling one frame per second from the videos. Using this dataset, we present a systematic study on scaling laws for deepfake detection and reveal several insights.

Figure 2: Statistics of heights and widths of all real faces in the ScaleDF dataset.

	Real	Deepfake	Videos		Images
Dataset	Domains	Methods	Real	Fake	Real	Fake
DF-TIMIT ( Korshunov and Marcel, 2018 )	$1$	$2$	$320$	$640$	-	-
UADFV ( Yang et al., 2019 )	$1$	$1$	$49$	$49$	-	-
FaceForensics++ ( Rossler et al., 2019 )	$1$	$4$	$1,000$	$4,000$	-	-
DeepFakeDetection ( Google, 2019 )	$1$	$5$	$363$	$3,068$	-	-
Celeb-DF V2 ( Li et al., 2020 )	$1$	$1$	$590$	$5,639$	-	-

Table 1: Compared to existing datasets, ScaleDF includes more real domains, more deepfake methods, and a larger number of videos and images, enabling scaling law research in these dimensions.

Figure 3: Left: observed power-law scaling as the number of training real domains changes; Right: observed power-law scaling as the number of training deepfake methods changes.

\mu

represents the mean computed over

10

repetitions and

7

test datasets, while

\sigma

denotes the variance across the

10

repetitions.

Figure 4: Left: observed double-saturating power-law scaling as the number of training images changes;

\mu

represents the mean computed over

5

repetitions and

7

test datasets, while

\sigma

denotes the variance across the

5

repetitions. Right: performance changes with respect to model sizes.

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
ImageNet	$0.793$	$0.915$	$0.815$	$0.824$	$0.909$	$0.980$	$0.968$	$0.886$
CLIP	$0.795$	$0.894$	$0.802$	$0.848$	$0.954$	$0.979$	$0.981$	$0.893$
SigLIP 2	$0.750$	$0.890$	$0.785$	$0.855$	$0.951$	$0.986$	$0.983$	$0.886$

Table 2: Comparison of using different pre-training models: similar performance observed.

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
N/A	$0.667$	$0.805$	$0.781$	$0.756$	$0.808$	$0.937$	$0.975$	$0.818$
QC	$0.774$	$0.914$	$0.802$	$0.770$	$0.932$	$0.978$	$0.976$	$0.878$
QC + PT	$0.793$	$0.915$	$0.815$	$0.824$	$0.909$	$0.980$	$0.968$	$0.886$

Table 3: Effectiveness of image quality compression (QC) and perturbations (PT).

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF
DFD	$-$	$0.802$	$0.738$	$0.610$	$0.545$	$0.594$	$0.524$
CDF V2	$0.709$	$-$	$0.777$	$0.640$	$0.599$	$0.665$	$0.612$
Wild	$0.643$	$0.757$	$-$	$0.555$	$0.489$	$0.562$	$0.556$
Forgery.	$0.813$	$0.913$	$0.821$	$-$	$0.687$	$0.833$	$0.657$
DFF	$0.583$	$0.568$	$0.552$	$0.578$	$-$	$0.669$	$0.622$
DF40	$0.587$	$0.800$	$0.684$	$0.677$	$0.607$	$-$	$0.667$

Table 4: Comparison in cross-benchmark setting: with scaling, we achieve the best performance.

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
w/o FF++	$0.758$	$0.868$	$0.796$	$0.807$	$0.905$	$0.978$	$0.969$	$0.869$
w FF++	$0.793$	$0.915$	$0.815$	$0.824$	$0.909$	$0.980$	$0.968$	$0.886$

Table 5: Comparison of whether includes FaceForensics++ ( Rössler et al., 2019 ) in the ScaleDF.

No.	Dataset	Format	Vol.	Link
$0$	GRID ( Cooke et al., 2006 )	Video	$16$ K+	Link
$1$	MORPH - 2 ( Ricanek and Tesafaye, 2006 )	Image	$49$ K+	Link
$2$	LFW ( Huang et al., 2008 )	Image	$13$ K+	Link
$3$	Multi-PIE ( Gross et al., 2008 )	Image	$0.1$ M+	Link
$4$	GENKI - 4K ( MPLab, 2009 )	Image	$3.8$ K+	Link
$5$	YouTubeFaces ( Wolf et al., 2011 )	Image	$0.2$ M+	Link

Table 6: Real datasets included in the ScaleDF, with the testing ones highlighted in .

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Left Top: observed power-law scaling as the number of training real domains changes; Right Top: observed power-law scaling as the number of training deepfake methods changes; Left Bottom: observed double-saturating power-law scaling as the number of training images changes; Right Bottom: performance changes with respect to model sizes.

\mu

represents the mean computed over repetitions and test datasets, while

\sigma

denotes the variance across repetitions.

$\mathrm{EER}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
ImageNet	$0.281$	$0.162$	$0.268$	$0.260$	$0.161$	$0.063$	$0.093$	$0.184$
CLIP	$0.278$	$0.179$	$0.279$	$0.231$	$0.085$	$0.065$	$0.052$	$0.167$
SigLIP 2	$0.316$	$0.193$	$0.290$	$0.229$	$0.094$	$0.052$	$0.052$	$0.175$

Appendix

Table 7: Comparison of using different pre-training models: similar performance observed.

Figure 7: Perceived demographic distribution of faces in ScaleDF.

No.	Method	Cat.	Arch.	Link
$0$	Faceswap ( Earl, 2015 )	FS	Affine	Code
$1$	FaceSwap ( Kowalski, 2016 )	FS	3D	Code
$2$	DeepFakes ( deepfakes, 2017 )	FS	VAE	Code
$3$	FSGAN ( Nirkin et al., 2019 )	FS	GAN	Code
$4$	SimSwap ( Chen et al., 2020 )	FS	GAN	Code
$5$	HifiFace ( Wang et al., 2021c )	FS	3D	Code

Appendix

Table 11: Deepfake methods included in the ScaleDF, with the testing ones highlighted in .

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
$46$	$0.793$	$0.915$	$0.815$	$0.824$	$0.909$	$0.980$	$0.968$	$0.886$
$32$	$0.752$	$0.845$	$0.716$	$0.755$	$0.881$	$0.927$	$0.966$	$0.835$
$32$	$0.764$	$0.917$	$0.747$	$0.818$	$0.907$	$0.985$	$0.968$	$0.872$
$32$	$0.779$	$0.904$	$0.727$	$0.817$	$0.920$	$0.981$	$0.967$	$0.871$
$32$	$0.765$	$0.897$	$0.817$	$0.817$	$0.865$	$0.977$	$0.960$	$0.871$
$32$	$0.760$	$0.904$	$0.732$	$0.811$	$0.863$	$0.980$	$0.960$	$0.859$

Appendix

Table 12: Full experimental results measured by

\mathrm{AUC}

for scaling law with varying numbers of real domains . The concluded scaling law is shown in Fig. 3 (Left).

$\mathrm{AUC}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
$88$	$0.793$	$0.915$	$0.815$	$0.824$	$0.909$	$0.980$	$0.968$	$0.886$
$64$	$0.774$	$0.914$	$0.776$	$0.815$	$0.874$	$0.982$	$0.965$	$0.871$
$64$	$0.756$	$0.912$	$0.802$	$0.820$	$0.893$	$0.978$	$0.955$	$0.874$
$64$	$0.790$	$0.917$	$0.802$	$0.836$	$0.888$	$0.981$	$0.963$	$0.882$
$64$	$0.782$	$0.915$	$0.796$	$0.823$	$0.888$	$0.977$	$0.963$	$0.878$
$64$	$0.766$	$0.908$	$0.806$	$0.802$	$0.894$	$0.978$	$0.965$	$0.874$

Appendix

Table 13: Full experimental results measured by

\mathrm{AUC}

for scaling law with varying numbers of deepfake methods . The concluded scaling law is shown in Fig. 3 (Right).

Table 18

$\mathrm{EER}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
$46$	$0.281$	$0.162$	$0.268$	$0.260$	$0.161$	$0.063$	$0.093$	$0.184$
$32$	$0.317$	$0.224$	$0.337$	$0.310$	$0.189$	$0.149$	$0.092$	$0.231$
$32$	$0.308$	$0.162$	$0.313$	$0.264$	$0.162$	$0.055$	$0.092$	$0.194$
$32$	$0.293$	$0.173$	$0.326$	$0.265$	$0.148$	$0.063$	$0.093$	$0.194$
$32$	$0.307$	$0.179$	$0.257$	$0.269$	$0.209$	$0.070$	$0.105$	$0.199$
$32$	$0.308$	$0.173$	$0.334$	$0.270$	$0.210$	$0.064$	$0.104$	$0.209$

Appendix

Table 16: Full experimental results measured by

\mathrm{EER}

for scaling law with varying numbers of real domains . The concluded scaling law is shown in Fig. 5 (Left Top).

$\mathrm{EER}$	DFD	CDF V2	Wild	Forgery.	DFF	DF40	ScaleDF	Mean
$88$	$0.281$	$0.162$	$0.268$	$0.260$	$0.161$	$0.063$	$0.093$	$0.184$
$64$	$0.299$	$0.165$	$0.301$	$0.268$	$0.200$	$0.061$	$0.096$	$0.199$
$64$	$0.312$	$0.167$	$0.279$	$0.264$	$0.175$	$0.066$	$0.112$	$0.196$
$64$	$0.283$	$0.162$	$0.277$	$0.249$	$0.186$	$0.064$	$0.099$	$0.189$
$64$	$0.293$	$0.162$	$0.272$	$0.261$	$0.183$	$0.067$	$0.100$	$0.191$
$64$	$0.304$	$0.172$	$0.278$	$0.278$	$0.178$	$0.068$	$0.096$	$0.196$

Appendix

Table 17: Full experimental results measured by

\mathrm{EER}

for scaling law with varying numbers of deepfake methods . The concluded scaling law is shown in Fig. 5 (Right Top).

Table 21

Panel of Table 20 — Table 20: Visualization of processed faces in ScaleDF.

Panel of Table 21 — Table 21: Visualization of processed faces in ScaleDF.

Panel of Table 22 — Table 22: Visualization of processed faces in ScaleDF.

Panel of Table 23 — Table 23: Visualization of processed faces in ScaleDF.

Panel of Table 24 — Table 24: Visualization of processed faces in ScaleDF.

Panel of Table 25 — Table 25: Visualization of processed faces in ScaleDF.

Panel of Table 26 — Table 26: Visualization of processed faces in ScaleDF.

Panel of Table 27 — Table 27: Visualization of processed faces in ScaleDF.

Panel of Table 28 — Table 28: Visualization of processed faces in ScaleDF.

Panel of Table 29 — Table 29: Visualization of processed faces in ScaleDF.

Panel of Table 30 — Table 30: Visualization of processed faces in ScaleDF.

Panel of Table 31 — Table 31: Visualization of processed faces in ScaleDF.

Panel of Table 32 — Table 32: Visualization of processed faces in ScaleDF.

Panel of Table 33 — Table 33: Visualization of processed faces in ScaleDF.

Panel of Table 34 — Table 34: Visualization of processed faces in ScaleDF.

Panel of Table 35 — Table 35: Visualization of processed faces in ScaleDF.

Panel of Table 36 — Table 36: Demonstration of the perturbations used when training models on the ScaleDF dataset.

Panel of Table 37 — Table 37: Demonstration of the perturbations used when training models on the ScaleDF dataset.

Panel of Table 38 — Table 38: Demonstration of the perturbations used when training models on the ScaleDF dataset.

Explore similar work

Aug 10, 2026cs.CV

Foundation Models are Implicit Deepfake Detectors

Pretrained self-supervised representations have emerged as a core component of current deepfake detection methods, yet it remains unclear which of their properties make real and fake media distinguishable. In this work, we uncover a surprisingly consistent phenomenon: across multiple pretrained models, datasets, and both image and video domains, fake samples systematically produce lower-magnitude representations than their real counterparts. Motivated by this finding, we formulate deepfake detection as an anomaly detection problem and show that simple statistics of feature magnitude achieve competitive performance with far more sophisticated deepfake detection methods. We further investigate the origin of this effect and demonstrate that reduced feature magnitude is primarily associated with semantic shifts introduced by fake content, while low-level generative fingerprints play a comparatively smaller role. Finally, we show that this discriminative signal strengthens as the size of the underlying foundation model grows, suggesting that advances in representation learning naturally translate into stronger zero-shot deepfake detectors.

Stefan Smeu, Dragos-Alexandru Boldisor, Elisabeta Oneata +1

1Bitdefender, Romania · 2POLITEHNICA Bucharest

Deepfake Detection Unsupervised Detection Self-Supervised Representations Fake News Fingerprint Foundation Model

Ask PDF

Nov 29, 2024cs.CV

Deepfake Media Generation and Detection in the Generative AI Era: A Survey and Outlook

We survey deepfake generation and detection techniques, covering all deepfake media types: image, video, audio and multimodal content. We identify various kinds of deepfakes and construct taxonomies of deepfake generation and detection methods, illustrating the important groups of methods. Next, we gather datasets used for deepfake detection and provide updated rankings of the best performing detectors on the most popular datasets. In addition, we develop a novel multimodal benchmark to evaluate deepfake detectors on out-of-distribution content. The results indicate that state-of-the-art detectors fail to generalize to deepfakes generated by unseen generators. Our project page and new benchmark are available at https://github.com/CroitoruAlin/biodeep.

Florinel-Alin Croitoru, Andrei-Iulian Hiji, Vlad Hondru +7

Department of Computer Science, University of Bucharest, Romania · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE · Department of Computer Science, University of Central Florida, Orlando, US

Deepfake Detection Unsupervised Detection Multimodal Generation Generative Artificial Intelligence Multimodal Deep Learning Multimodal Benchmarks

Ask PDF

Oct 6, 2026cs.CV

CCDF: A Benchmark Dataset for Deepfake Detection in Real-World Surveillance Footage

Due to rapid advances in Generative AI, commercial video generation tools can be used to produce fabricated surveillance footage that can fool both human viewers and automated synthetic video detectors. Since these tools are so widely accessible, a malicious user can create a harmful video clip at minimal cost. The production and dissemination of such videos in high-stakes settings, such as crime reporting and elections, can misdirect emergency response efforts or distort political discourse. Existing deepfake video datasets, used by the research community to develop deepfake detection algorithms, exhibit two limitations: (1) they emphasize benign web content rather than footage of possibly malicious activity, and (2) they rely on older or open-source generators that do not represent recent advances in generative systems. We assemble CCtv DeepFakes (CCDF), a video deepfake dataset, to address both gaps. CCDF contains 1840 videos (460 real and 1380 generated) spanning 16 crime and accident categories, with generated content produced using three leading commercial systems: Grok Imagine, Google VEO 3.1, and OpenAI Sora 2. CCDF is a highly realistic, small-scale, manually annotated dataset targeting evaluation of detection models. We release three versions of the dataset: the raw generated data, a cleaned version in which video metadata are standardized between real and synthetic samples to prevent detectors from exploiting trivial cues, and an altered version simulating low-effort post-processing attacks. We evaluate CCDF with ten recent state-of-the-art detectors covering different detection approaches. Our results suggest that these approaches do not reliably distinguish CCDF's generated videos from real ones, despite their strong reported performance on existing datasets. These results further confirm that existing datasets are not well-suited to evaluating certain threats.

Baptiste Chopin, Thomas Swearingen, Arun Ross +2

da/sec – Biometrics and Security Research Group, Hochschule Darmstadt · Computer Science and Engineering Department, Michigan State University · STARS team, Inria Center at Université Côte d’Azur

Deepfake Detection Ai-Generated Video Detection Video Dataset

Ask PDF