PaperlayerView exact revision on arXiv

Source-linked AI summary

DF40: Toward Next-Generation Deepfake Detection

Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, Li Yuan

arXiv:2406.13495v2cs.CV

TL;DR

Existing deepfake detectors may fail to cover realistic and diverse forgeries because datasets contain limited forgery types, outdated techniques, and narrow evaluation settings. DF40 addresses this gap with 40 deepfake techniques and broad evaluations across four protocols and eight detectors, revealing that current benchmark winners may not generalize universally.

  • Problem

    Existing detection methods often handle only subsets of deepfake types, while prevailing datasets and evaluation protocols provide limited coverage of diverse and realistic forgeries.

  • Method

    DF40 comprises 40 deepfake techniques and evaluates eight representative detectors under four standard protocols using more than 2,000 evaluations.

  • Results

    Current SoTA performance on existing datasets may not identify a definitive winner, because CFDA performs comparably or worse than other detectors on DF40.

  • Takeaways & Limitations

    Diverse datasets such as DF40 support broader assessment of deepfake detectors against realistic deepfake techniques and types.

  • Takeaways & Limitations

    Blending-based training remains limited in generalizing to all deepfake types, leaving the role and coverage of blending data unresolved.

Abstract

from arXiv · show

We propose a new comprehensive benchmark to revolutionize the current deepfake detection field to the next generation. Predominantly, existing works identify top-notch detection algorithms and models by adhering to the common practice: training detectors on one specific dataset (e.g., FF++) and testing them on other prevalent deepfake datasets. This protocol is often regarded as a "golden compass" for navigating SoTA detectors. But can these stand-out "winners" be truly applied to tackle the myriad of realistic and diverse deepfakes lurking in the real world? If not, what underlying factors contribute to this gap? In this work, we found the dataset (both train and test) can be the "primary culprit" due to: (1) forgery diversity: Deepfake techniques are commonly referred to as both face forgery and entire image synthesis. Most existing datasets only contain partial types of them, with limited forgery methods implemented; (2) forgery realism: The dominated training dataset, FF++, contains out-of-date forgery techniques from the past four years. "Honing skills" on these forgeries makes it difficult to guarantee effective detection generalization toward nowadays' SoTA deepfakes; (3) evaluation protocol: Most detection works perform evaluations on one type, which hinders the development of universal deepfake detectors. To address this dilemma, we construct a highly diverse deepfake detection dataset called DF40, which comprises 40 distinct deepfake techniques. We then conduct comprehensive evaluations using 4 standard evaluation protocols and 8 representative detection methods, resulting in over 2,000 evaluations. Through these evaluations, we provide an extensive analysis from various perspectives, leading to 7 new insightful findings. We also open up 4 valuable yet previously underexplored research questions to inspire future works. Our project page is https://github.com/YZY-stack/DF40.

1 Introduction

Deepfake detection is challenged by diverse, increasingly realistic manipulations that existing methods and datasets do not comprehensively cover. DF40 addresses this gap with 40 techniques and broader evaluation protocols.

  • Motivation: Deepfakes can manipulate identity through face-swapping or control facial expressions and movements through face-reenactment.The paper links these manipulations to digital crimes and weakened social trust.
  • Limitations of Existing Detection: Existing detectors often handle only subsets of deepfake types and may rely on blending artifacts absent from newer non-blending methods.Simswap and FaceDancer directly generate all content, including the background.
  • DF40 Contribution: DF40 implements 40 deepfake techniques spanning face-swapping, face-reenactment, entire face synthesis, and face editing.Its methods include recent generators and widely used software intended to simulate real-world deepfakes.
  • DF40 Contribution: DF40 introduces four standard evaluation protocols and reports eight insightful findings to support broader deepfake-detection research.The benchmark is designed to move beyond training and testing solely on face-swapping fakes.

2 Background

Deepfake datasets have historically emphasized blending-based face-swapping, while newer resources cover either limited manipulation types or only entire-image synthesis. DF40 expands coverage across four deepfake categories and 40 methods.

  • Deepfake Taxonomy: Face-swapping methods divide into DF-family approaches that blend masked faces and FS-family approaches that directly generate all content, including backgrounds.FF-DF, CDF, and DFDC exemplify the blending-based line.
  • Existing Datasets: Earlier public datasets mostly used blending-based face-swapping, with no more than four specific manipulation approaches.DFDC expanded to seven approaches, and ForgeryNet to 15 methods, while later datasets addressed diffusion or entire-image synthesis.
  • Dataset Construction: HeyGen and DeepFaceLab are widely used and highly realistic, but their one-to-one implementation requires substantially more time and resources.DF40 therefore creates only 50 and 100 examples for these methods, respectively.
  • Existing Datasets: DiffusionDeepfake and GenImage focus on entire-image synthesis and may not guarantee detection of face-swapping deepfakes.Their scopes differ from a benchmark intended to cover multiple deepfake categories.

3 DF40 Benchmark

DF40 is a diverse face-deepfake benchmark spanning four forgery types, multiple data domains, and both video and image formats. Its construction supports cross-forgery and cross-domain evaluation while modeling distinct generation processes.

  • Dataset scope: DF40 provides 40 deepfake approaches across face-swapping, face-reenactment, entire face synthesis, and face editing.It includes over 0.1M video clips for FS and FR and over 1M images for EFS and FE.
  • Data construction: The dataset uses FF++ and CDF as original data to retain prior train-on-FF++ and test-on-CDF evaluation settings.Additional real data from datasets such as CelebA support unknown-domain evaluations.
  • Evaluation protocols: Protocol-1 evaluates different forgery types within the same data domain, while Protocol-2 evaluates the same forgery types across different data domains.Table 3 focuses on the FF domain, whereas Table 4 evaluates on the CDF domain and distinguishes FF-created from CDF-created data.
  • Forgery categories: Face-swapping replaces target content while preserving the source identity, whereas face-reenactment preserves target identity and changes attributes such as pose, mouth, or expression.These operations are represented through distinct generation functions in the DF40 construction.
  • Forgery categories: Entire face synthesis fine-tunes a generative model on FF++ and CDF real data, then generates new faces from noise.Face editing instead alters external attributes such as facial hair or age using a conditional source.

4 Evaluations and Analysis

The evaluations show that detector performance depends strongly on forgery type, data domain, manipulated region, and model pretraining, challenging conclusions drawn from narrow benchmarks. DF40 further exposes transfer asymmetries, domain–forgery interactions, and several open research questions.

  • Experimental Setup: 40 distinct deepfake techniques are evaluated with standardized preprocessing, five Xception-based detectors, and cross-type, cross-domain protocols.The selected methods include Xception, SRM, SPSL, RECCE, and RFM.
  • Findings: FS-trained models reach around 0.8 AUC on EFS, whereas EFS-trained models achieve about 0.6 AUC on FS, showing asymmetric transfer across forgery types.The authors attribute this pattern partly to FS containing both localized and global forgeries, while EFS is restricted to global forgeries.
  • Findings: SRM, SPSL, RECCE, and RFM perform similarly to Xception, despite achieving higher results on previous datasets such as DFDC.The authors suggest these SoTA detectors may learn sub-optimal forgery features under the DF40 setting.
  • Findings: CLIP-base and CLIP-large outperform Xception and other SoTA detectors across evaluated scenarios, with pretraining producing more informative real-face features.The analysis links CLIP’s advantage to more structured real-sample representations and reduced reliance on identity-related artifacts.
  • Findings: Changing both forgery method and domain causes a significant performance drop, with Xception scoring 0.657 on FR (CDF) and 0.642 on EFS (CDF) after FS (FF) training.These results are nearly 20% lower than settings crossing only one factor.
  • Findings: FR methods generally transfer across domains, but Wav2Lip is an exception because its mouth-region artifacts are more localized than other FR methods.The authors caution that transfer among FR methods does not eliminate the value of collecting diverse FR techniques.
  • Further Analysis: A diverse evaluation set changes apparent rankings: CFDA’s advantages on existing datasets disappear or reverse on DF40, while CLIP-large reaches 0.746 AUC on unseen non-face deepfakes.The findings also report that super-resolution improves SRI detection by more than 20 percentage points for models trained on FS and EFS.
  • Open Questions: The paper opens questions about blending data, jointly learning diverse forgeries, classifying by FS/FR/EFS/FE, and building domain-invariant detectors.These questions follow the observed limits of blending-based generalization and the joint influence of domain and forgery method.

5 Conclusions, Board Impacts, and Limitations

The paper presents DF40 as a 40-technique benchmark for broad deepfake detection evaluation, reporting more than 2,000 evaluations, seven findings, and four open questions. It emphasizes realistic data for contemporary detection while acknowledging limited analysis of video-level detectors.

  • Conclusions: DF40 comprises 40 distinct deepfake techniques and supports detection research through a highly diverse benchmark.The benchmark is intended to support next-generation deepfake detection.
  • Conclusions: More than 2,000 evaluations use eight representative detectors and four standard protocols, producing seven findings and four open questions.The evaluations are designed to analyze detection from multiple perspectives.
  • Board Impacts: DF40 provides realistic techniques intended to facilitate detection of today’s real-world deepfakes and help safeguard societal trust.The authors also frame the benchmark as supporting responsible technology use.
  • Limitations: A limitation is the lack of comprehensive analysis of video-level detectors, including whether they capture temporal and spatial artifacts effectively.Video-model evaluations are included in the Appendix, but the issues are not deeply analyzed.

Checklist

The checklist records the paper’s disclosures on contributions, limitations, ethics, assets, reproducibility, and experimental reporting. It also states that the work has no theoretical results.

  • Disclosures: The paper reports that its contributions, research scope, and limitations are described in the introduction, benchmark section, and Section 5.The checklist marks these disclosures as present.
  • Ethics: The paper states that it discusses potential negative societal impacts and proposes controlled access.These disclosures are referenced in Section 5 and the Appendix.
  • Theory: The work reports no theoretical results and therefore provides neither theoretical assumptions nor complete proofs.The checklist explicitly marks both theoretical-result items as not applicable.
  • Reproducibility: The authors provide reproduction instructions, training details, fixed-seed settings, and compute-resource information in the Appendix.The fixed seed matches DeepfakeBench to facilitate fair comparison.
  • Assets: The paper cites existing assets, provides licensing information, and links its code, dataset, and pretrained models.The checklist marks these asset and release disclosures as present.
  • Data and Participants: The authors state that real people were drawn from existing research works and report disclosures concerning consent, personally identifiable information, and offensive content.The checklist marks these disclosures as present.
  • Participants: The checklist indicates that participant-risk and compensation information were not provided because the corresponding crowdsourcing or human-subjects items were marked not applicable.Both participant-risk and hourly-wage/compensation questions are marked no.

A.1 Content Structure in Appendix

The Appendix organizes dataset generation, detection methods, experimental setup, full results, further analyses, and dataset-publication information. It also includes sections on controlled access, fairness, and related safeguards.

  • Dataset Generation: Section A.2 covers dataset-generation methods, original data, manipulation formulations, EFS fine-tuning, and source datasets.It includes details of all 40 implemented synthesis methods.
  • Detection Methods: Section A.3 introduces the detection methods used in the study.The outline separates detector descriptions from dataset-generation details.
  • Experiments: Section A.4 presents the experimental setup, implementation details, full experimental results, and further discussion.The section is divided into setup/details and full results/discussion subsections.
  • Further Analysis: Section A.5 reports further analyses, including fake regions, non-face AIGCs, external datasets, CLIP variants, anomalous values, video versus image models, and feature visualizations.The outline also lists analysis of deepfake artifacts.
  • Publication Information: Section A.6 provides dataset-publication information, including hosting links, controlled access, discrimination, bias, and fairness.These topics are listed as additional publication information.

A.2.1 Brief Introduction of Generation Methods

Deepfake generation methods are classified into face-swapping, face-reenactment, entire face synthesis, and face editing. These categories differ in their manipulation targets and commonly used techniques.

  • Deepfake techniques are classified into face-swapping, face-reenactment, entire face synthesis, and face editing.
  • Face-swapping: Face-swapping includes DF-family mask-and-blending methods and FS-family methods, with FF-DF, Celeb-DF, and DFDC representing the former line.
  • Face-reenactment: Face-reenactment modifies source faces to imitate another face’s actions or expressions and is rarely represented in existing datasets.Face2Face and NeuralTextures are commonly used reenactment-based forgeries implemented in FF++.
  • Entire Face Synthesis: Entire face synthesis, or Face AIGC, uses GANs such as StarGAN and diffusion models such as StableDiffusion to generate faces.
  • Face Editing: Face editing modifies facial attributes such as age and gender, commonly using StyleGAN latent codes during GAN inversion.

A.2.2 Details of Implemented Forgery Methods in DF40

DF40 implements a broad set of deepfake generation and synthesis methods, totaling 40 distinct techniques.

  • DF40 uses 40 distinct deepfake generation and synthesis methods.
  • The manuscript provides details for the implemented synthesis methods in Table 2.
  • Each synthesis method is briefly explained in the accompanying method descriptions.

1. FSGAN [52]

FSGAN combines face reenactment with face swapping by matching source-video frames to the target’s pose, angle, and expression before transferring and blending facial content.

  • FSGAN combines reenactment and face swapping as its key generation strategy.
  • It selects source-video frames that best correspond to the target video’s pose, angle, and expression.
  • FSGAN transfers missing facial parts and blends them with the target video after reenactment.

2. FaceSwap [1]

FaceSwap is a traditional, landmark- and graphics-based face-swapping method that uses facial landmark detection and image warping rather than neural networks.

  • FaceSwap performs face swapping without employing a neural network.
  • It relies on facial landmark detection and image warping to produce swapped faces.
  • FaceSwap is considered the relatively poor-quality prototype of face swapping in FF++.

3. SimSwap [9]

SimSwap is a state-of-the-art single-frame face-swapping method that produces high-quality results without prior exposure to academic datasets.

  • SimSwap is described as a state-of-the-art method for single-frame face swapping.
  • The facefusion software encapsulates SimSwap for direct use in deepfake generation.
  • SimSwap uses attention mechanisms and GANs to produce visually appealing and realistic single-frame face swaps.

4. InSwapper [28]

DF40 evaluates diverse deepfake detection methods across manipulation types and domains, revealing strong specialization, inconsistent transfer, and weaker performance for face editing and cross-domain testing.

  • Face-swapping methods: FaceDancer is a recent single-stage, identity-based face-swapping method for unknown identities.
  • Face-swapping methods: DeepFaceLab provides a complete end-to-end pipeline and is identified as a leading method for generating deepfake videos.
  • Dataset and evaluation: DF40 implements 40 distinct deepfake techniques selected across modeling types, conditional sources, forgery effects, and functions.
  • Detection results: 0.992 AUC is reported for FaceSwap detection by its own detection method, while Wav2Lip reaches 0.380 AUC on FaceSwap deepfakes.
  • Detection results: Detection performance generally declines on face-editing deepfakes, and AUC tends to be lower when testing crosses FF and CDF domains.
Loading 2406.13495v2…