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

The DOI matches its Crossref record; 54 of the 54 references it cites were found in a registry or an index; 6 of the 11 sources its claims cite were checked for support.

Deepfake Video Detection with Spatiotemporal Dropout Transformer

Open access
Zhang, DaiChi, Lin, Fanzhao, Hua, Yingying, Wang, Pengju, Zeng, Dan, Ge, Shiming · 2022 · Proceedings of the 30th ACM International Conference on Multimedia

Abstract

While the abuse of deepfake technology has caused serious concerns recently, how to detect deepfake videos is still a challenge due to the high photo-realistic synthesis of each frame. Existing image-level approaches often focus on single frame and ignore the spatiotemporal cues hidden in deepfake videos, resulting in poor generalization and robustness. The key of a video-level detector is to fully exploit the spatiotemporal inconsistency distributed in local facial regions across different frames in deepfake videos. Inspired by that, this paper proposes a simple yet effective patch-level approach to facilitate deepfake video detection via spatiotemporal dropout transformer. The approach reorganizes each input video into bag of patches that is then fed into a vision transformer to achieve robust representation. Specifically, a spatiotemporal dropout operation is proposed to fully explore patch-level spatiotemporal cues and serve as effective data augmentation to further enhance model's robustness and generalization ability. The operation is flexible and can be easily plugged into existing vision transformers. Extensive experiments demonstrate the effectiveness of our approach against 25 state-of-the-arts with impressive robustness, generalizability, and representation ability.

Overview

CoChat Summary

TL;DR

The paper proposes a Spatiotemporal Dropout Transformer (STDT) that reorganizes deepfake videos into bags of facial patches and applies temporal and spatial dropout to capture patch-level spatiotemporal inconsistency. It reports state-of-the-art AUC on FaceForensics++, DFDC, and Celeb-DF(v2), with strong robustness and cross-dataset generalization.

Key findings

  • STDT achieves the highest AUC among 25 compared state-of-the-arts on all three benchmarks: 99.8% on FaceForensics++, 99.1% on DFDC, and 97.2% on Celeb-DF(v2), which is at least 0.3%, 0.6%, and 2.8% higher than the state-of-the-arts, respectively.
  • Intra-dataset results show all ACC above 90% and AUC above 97%; on the Deepfakes, Face2Face, FaceShifter, and FaceSwap subsets of FF++, all AUC are above 99%.
  • Ablation on Celeb-DF(v2) shows spatiotemporal dropout (S+T) gives the best performance: ACC 91.70%, AUC 97.21%, REC 95.01%, PRE 92.55%, F1 93.76%, compared with no dropout at ACC 87.65%, AUC 93.43%, REC 94.71%, PRE 85.70%, F1 90.96%.
  • Incorporating STD into four ViT backbones consistently improves performance, averaging ACC improvement of 11.15% and AUC improvement of 7.76%; robustness tests under Flip, Blur, Bright, Compress, and Gaussian Noise remain close to origin, e.g., Flip gives ACC 90.54%, AUC 96.56% versus origin ACC 91.70%, AUC 97.21%.

Methods

  • The approach extracts 24 raw frames per video, performs face detection and alignment, then applies temporal dropout with rate α = 1/4 and spatial dropout with rate β = 17/18 after cropping each face into 6 × 6 non-overlapping patches; remaining patches are reorganized into a bag of patches.
  • The bag of patches is fed into a ViT-Base-16 backbone with 12 layers, 12 self-attention heads, dropout rate 0.1, attention dropout 0.0, projected token dimension 768, and MLP header dimension 3072; training uses Binary Cross Entropy, SGD, OneCycleLR, global learning rate 10^-3, and weight decay 10^-4.
  • Experiments compare against 25 state-of-the-art methods on FaceForensics++ (FF++), DFDC, and Celeb-DF(v2), using AUC as the main metric, plus robustness tests, cross-dataset generalization tests, t-SNE visualization, and ablation studies on dropout operation, dropout rates, and ViT backbones.

Limitations

  • The paper states future work will extend the approach to more video understanding tasks and enhance its interpretability to provide more human-understandable detection results.
  • Evident from the text: the ablation analysis is conducted on Celeb-DF(v2) only, and cross-dataset generalization results are notably lower than intra-dataset results, e.g., training on Celeb-DF and testing on DFDC gives AUC 70.14%, while training on DFDC and testing on Celeb-DF gives AUC 73.60%.

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

Integrity signals

Registry

Crossref · 10.1145/3503161.3547913

Cites

54 of 54 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY · From the PDF
Image 2
CC-BY · From the PDF
Image 3
CC-BY · From the PDF
Image 4
CC-BY · From the PDF
Image 5
CC-BY · From the PDF
Image 6
CC-BY · From the PDF
Figure 2 The framework of Spatiotemporal Dropout Transformer for deepfake video detection. We first employ frame extraction, face detection and alignment to get facial image sequence F = { x 𝑖 } 𝑛 𝑖 = 1 . Then each facial image x 𝑖 ∈ F is processed by STD operation to get facial patch set 𝑃 𝑖 ′ = { p 𝑖 ′ , 𝑗 } and further reorganized as a bag of patches instance P which is then fed into our vision transformer backbone to learn the discriminative representation and detect.
Figure 2 The framework of Spatiotemporal Dropout Transformer for deepfake video detection. We first employ frame extraction, face detection and alignment to get facial image sequence F = { x 𝑖 } 𝑛 𝑖 = 1 . Then each facial image x 𝑖 ∈ F is processed by STD operation to get facial patch set 𝑃 𝑖 ′ = { p 𝑖 ′ , 𝑗 } and further reorganized as a bag of patches instance P which is then fed into our vision transformer backbone to learn the discriminative representation and detect.CC-BY · From the PDF
Image 8
CC-BY · From the PDF
Figure 4 Two video examples from Celeb-DF(v2) with five augmentation methods and the predicted possibilities of being fake.
Figure 4 Two video examples from Celeb-DF(v2) with five augmentation methods and the predicted possibilities of being fake.CC-BY · From the PDF
Figure 5 Visualization of representations on Celeb-DF(v2), DFDC and FF++ (five subsets) via t-SNE. Red: Real, Blue: Fake.
Figure 5 Visualization of representations on Celeb-DF(v2), DFDC and FF++ (five subsets) via t-SNE. Red: Real, Blue: Fake.CC-BY · From the PDF
Figure 6 Ablation analysis on temporal dropout rate 𝛼 on Celeb-DF(v2). We set 𝛼 to 1/2, 1/4, 1/8, and 1/16 with other hyper-parameters fixed.
Figure 6 Ablation analysis on temporal dropout rate 𝛼 on Celeb-DF(v2). We set 𝛼 to 1/2, 1/4, 1/8, and 1/16 with other hyper-parameters fixed.CC-BY · From the PDF
Figure 7 Ablation analysis on Spatial Dropout rate 𝛽 ′ = 1 -𝛽 on Celeb-DF(v2). We set 𝛽 ′ to 1/6, 1/9, 1/18, and 1/36 with other hyper-parameters fixed.
Figure 7 Ablation analysis on Spatial Dropout rate 𝛽 ′ = 1 -𝛽 on Celeb-DF(v2). We set 𝛽 ′ to 1/6, 1/9, 1/18, and 1/36 with other hyper-parameters fixed.CC-BY · From the PDF

Verification

Claims the paper makes

6 Valid · 1 Not valid · 0 Not checked · 51 Own results and interpretations

26 more claims

Create a free account to see all 58 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 19
  • The quote only partly states this claim. 2
  • Probably not a substantial claim. 6

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.

1 of 1

Citations

Cites

Show all 36 references

Research works printed but not linked: 4

  1. Ian J. Goodfellow Jean Pouget-Abadie Mehdi Mirza Bing Xu David Warde-Farley Sherjil Ozair Aaron C. Courville and Yoshua Bengio. 2014. Generative Adversarial Nets. In NeurIPS. 2672--2680. Ian J. Goodfellow Jean Pouget-Abadie Mehdi Mirza Bing Xu David Warde-Farley Sherjil Ozair Aaron C. Courville and Yoshua Bengio. 2014. Generative Adversarial Nets. In NeurIPS. 2672--2680.
  2. Kai Han Yunhe Wang Hanting Chen Xinghao Chen Jianyuan Guo Zhenhua Liu Yehui Tang An Xiao Chunjing Xu Yixing Xu etal 2020. A survey on visual transformer. arXiv preprint (2020). Kai Han Yunhe Wang Hanting Chen Xinghao Chen Jianyuan Guo Zhenhua Liu Yehui Tang An Xiao Chunjing Xu Yixing Xu et al. 2020. A survey on visual transformer. arXiv preprint (2020).
  3. Yuezun Li Xin Yang Pu Sun Honggang Qi and Siwei Lyu. 2020c. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics. In CVPR. 3204--3213. Yuezun Li Xin Yang Pu Sun Honggang Qi and Siwei Lyu. 2020c. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics. In CVPR. 3204--3213.
  4. Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Llion Jones etal 2017. Attention is all you need. In NeurIPS. 5998--6008. Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Llion Jones et al. 2017. Attention is all you need. In NeurIPS. 5998--6008.

Cited by

See all 50

Metadata

Authors
Zhang, D., Lin, F., Hua, Y., Wang, P., Zeng, D., & Ge, S.
Affiliations
  • Chinese Academy of Sciences (CN)
  • Shanghai University (CN)
Venue
Proceedings of the 30th ACM International Conference on Multimedia
Publisher
ACM
Published
2022-10-10
Pages
5833-5841
OpenAlex
W4304015019
Cited by
50
References
36
Citations by year
2026: 7 · 2025: 19 · 2024: 16 · 2023: 7 · 2022: 1
FWCI
3.90
Type
conference paper
Language
en
Open access
Locations
  • ACM International Conference on Multimedia (ACM MM) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Key Research and Development Program of China (Grant No. 2020AAA0140001)
  • National Natural Science Foundation of China (61772513)
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, vision transformer, data augmentation, facial region

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

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

Cited sources are checked with a language model’s help (gpt-6-sol and 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: Deepfake Video Detection with Spatiotemporal Dropout Transformer. https://app.cochat.ai/papers/10.1145/3503161.3547913
The paper
Zhang, D., Lin, F., Hua, Y., Wang, P., Zeng, D., & Ge, S. (2022). Deepfake Video Detection with Spatiotemporal Dropout Transformer. Proceedings of the 30th ACM International Conference on Multimedia. https://doi.org/10.1145/3503161.3547913

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.