The DOI matches its Crossref record; 37 of the 37 references it cites were found in a registry or an index; 14 of the 34 sources its claims cite were checked for support.
DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection
Open accessAbstract
A standardized benchmark for deepfake detection with unified data management, 15 state-of-the-art methods, 9 datasets, and evaluation protocols.
Overview
CoChat Summary
TL;DR
DeepfakeBench is a comprehensive benchmark for deepfake detection that unifies data processing, training, evaluation, and analysis across 15 detectors and 9 datasets. Its evaluations show that naive detectors can be competitive under consistent settings, while cross-domain and cross-manipulation generalization remains a major challenge.
Key findings
- DeepfakeBench includes 15 state-of-the-art detection methods, 9 deepfake datasets, and standardized evaluation metrics: AUC, AP, ACC, EER, Precision, and Recall, with frame-level evaluation adopted for fair comparison.
- In within-domain evaluation trained on FF++ c23, UCF, Xception, EfficientB4, and F3Net achieved average AUCs of 95.37%, 94.50%, 93.89%, and 94.49%, respectively; in cross-domain evaluation, SPSL achieved an average score of 78.75%.
- Naive detectors such as Xception and EfficientB4 performed comparably to more sophisticated methods; adding tricks like data augmentation and pre-training reduced the gap, with Xception improving by 30.56% on average, RECCE by 24.08%, and SRM by 21.13% (Table 10).
- Removing all augmentations improved within-domain FF++ c23 performance by about 2% for Xception and SPSL; Xception and EfficientNet-B4 consistently outperformed ResNet34; adding depthwise separable convolution to ResNet improved average AUC by +5.64% (Table 4).
- Cross-manipulation evaluation showed a significant generalization gap: models trained on specific forgeries struggled on unseen forgeries, with FF-NT especially challenging and FF-DF comparatively easy; adding phase spectrum to FWA (iFWA) improved average cross-data AUC from 73.16% to 80.35% and cross-manipulation average from 0.91 to 0.96.
Methods
- The benchmark provides a modular codebase with three core modules: Data Processing (frame extraction, DLIB face detection, cropping, alignment, mask/landmark extraction, JSON-based arrangement), Training (15 detectors across naive, spatial, and frequency categories), and Evaluation/Analysis (metrics, ROC-AUC, radar charts, Grad-CAM, t-SNE, custom visualizations).
- Main experiments train detectors on FF++ c23 and evaluate on 14 testing datasets (6 within-domain, 8 cross-domain); faces are aligned and resized to 256×256, Adam is used with learning rate 0.0002, batch size 32 (16 for DSP-FWA and Face X-ray), 32 frames sampled per video, ImageNet pre-trained backbones where feasible, and experiments run on NVIDIA A100 GPUs.
- Evaluations include within-domain and cross-domain AUC, cross-manipulation analysis, and ablation-style analyses of data augmentation, backbone architecture, pre-training, and learned representations.
Limitations
- The paper states that DeepfakeBench currently focuses on frame-level algorithms and evaluation; future work will add video-level detectors and metrics, more evaluations for diffusion- and GAN-generated images, and a wider range of detectors and datasets.
- The paper notes an ethical risk that malicious actors might exploit DeepfakeBench to refine deepfakes to evade detection, and mentions considering controlled access as a potential solution.
- The text makes evident additional limitations: temporal-based detectors are not yet incorporated, traditional detectors such as Headpose are excluded due to limited scalability, some methods are excluded because they lack open-source code or are difficult to reproduce, and the validation set is not used—models are selected based on highest test-set performance.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 4 Not valid · 9 Not checked · 99 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“While these benchmarking methodologies have made significant contributions, they specifically focus on their own datasets, without offering a standardized way to handle data across different datasets”
What the sources say
- Partly supportedcites [33] [33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF Conference on International Conference on Computer Vision , pages 1-11, 2019. 1, 2, 3, 4, 6, 7, 15, 17, 18, 19, 20, 21, 29 - sourceExperimental — accuracy not yet measured
Section 5 (Benchmark; page unavailable) describes a benchmark built from additional videos collected and manipulated for FaceForensics++, supporting the point that this benchmark uses its own data. It does not establish that existing benchmarks generally focus only on their own datasets or lack a standardized way to handle data across datasets; its stated goal is standardized comparison of follow-up work. (LLM, conf=0.82)
- Partly supportedcites [10] [10] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge dataset. arXiv preprint arXiv:2006.07397 , 2020. 3, 17, 20, 21 - sourceExperimental — accuracy not yet measured
The Introduction (p. 2) describes a benchmark competition using the DFDC dataset, and Large scale benchmarking (p. 9) describes evaluation on its public and private test sets. That supports the dataset-specific aspect of the claim. The paper does not assess whether existing benchmarks offer a standardized way to handle data across different datasets. (LLM, conf=0.81)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, these two benchmarks still lack a modular, extensible, and comprehensive codebase that includes data preprocessing, unified settings, training modules, evaluations, and a series of analytical tools.”
What the sources say
- Partly supportedcites [25] [25] Chenhao Lin, Jingyi Deng, Pengbin Hu, Chao Shen, Qian Wang, and Qi Li. Towards benchmarking and evaluating deepfake detection. arXiv preprint arXiv:2203.02115 , 2022. 3 - sourceExperimental — accuracy not yet measured
Section 2.3, ‘Evaluation Metrics and Forensic Benchmarks,’ says most existing benchmarks do not re-implement methods with the same training data, creating evaluation bias. The Introduction likewise identifies inconsistent training data as a problem. Our copy does not establish that the two benchmarks lack a modular, extensible codebase with all the listed components. (LLM, conf=0.82)
- Not matchedcites [18] [18] Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Yabin Wang, Mohamad Shahbazi, Xiaopeng Hong, and Luc Van Gool. A continual deepfake detection benchmark: Dataset, methods, and essentials. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1339-1349, 2023. 3 -
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“9 deepfake datasets”
- Also stated in: § 3.1 Datasets and Detectors · p. 3 · plain
“Our benchmark currently incorporates a collection of 9 widely recognized and extensively used datasets in the realm of deepfake detection”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 3.1 Datasets and Detectors · p. 3 · plain
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“our benchmark offers a collection of well-processed and systematically organized datasets”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“as well as comprehensive evaluations”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
52 more claims
Create a free account to see all 120 claims and to verify the 3 references still open.
- Probably not a substantial claim. 47
- It may state two separate claims as one. 28
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“As detectors matured, research increasingly separated approaches into naïve end-to-end classifiers versus methods that explicitly model forgery artifacts such as blending boundaries, warping, or inconsistent face boundaries [8].”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Nonetheless, standardized reporting exposes the magnitude of the generalization gap that dominates practical risk [8].”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“First, content-based deepfake detectors can achieve strong in-domain performance, but their reliability drops under cross-dataset shift, unseen generators, compression, and adversarial post-processing.”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [8]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“robustness improvements frequently come from engineering discipline (unified preprocessing, augmentation, and evaluation) rather than from a single novel architecture [8].”
From A Review of Tools and Technologies to Combat Deepfakes (2026) · cites this paper as [8]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“We surprisingly observe a substantial improvement over the baseline methods within the deepfake benchmark [55].”
From Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (2024) · cites this paper as [55]
5 more claims about this paper
Create a free account to see every claim other papers make about this one.
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.1 Experimental Setup · p. 5
“The batch size is fixed at 32 for all experiments.”
§ A.3 Training Details and Full Experimental Results · p. 21
“However, for the DSPFWA [22] and Face X-ray [20] detectors, the batch size is adjusted to 16 due to the input data being pairs.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ A.4 Other Analysis Results · p. 32
“Table 11: The results of our main experiments (shown in Table. 2 and Figure. 2 of the manuscript) are generated using the following settings.”
§ 2 Related Work · p. 3
“Table 2: Comprehensive comparison of our benchmark with existing benchmark [25].”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ A.4 Other Analysis Results · p. 32
“| Weight Initialization | ImageNet Pre-trained | ImageNet Pre-trained |”
§ A.3 Training Details and Full Experimental Results · p. 21
“However, Meso4 [1] and MesoIncep [1] do not have pre-training weights in ImageNet, so pre-training is not utilized for them.”
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 13
- [5][5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1251-1258, 2017. 2, 18, 19 -
- [16][16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 770-778, 2016. 2, 17, 18 -
- [18][18] Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Yabin Wang, Mohamad Shahbazi, Xiaopeng Hong, and Luc Van Gool. A continual deepfake detection benchmark: Dataset, methods, and essentials. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1339-1349, 2023. 3 -
- [19][19] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Advancing high fidelity identity swapping for forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5074-5083, 2020. 3, 20, 21 -
- [20][20] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020. 2, 3, 5, 6, 15, 18, 19, 21, 31 -
- [21][21] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security , 2018. 2, 3, 17, 20, 21 -
- [23][23] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A new dataset for deepfake forensics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020. 2, 3, 17, 20, 21 -
- [27][27] Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with highfrequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021. 2, 3, 6, 19, 31 -
- [30][30] Yunsheng Ni, Depu Meng, Changqian Yu, Chengbin Quan, Dongchun Ren, and Youjian Zhao. Core: Consistent representation learning for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop , pages 12-21, 2022. 2, 6, 18 -
- [31][31] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Umé, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: A simple, flexible and extensible face swapping framework. arXiv preprint arXiv:2005.05535 , 2020. 2 - · arXiv
- [44][44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research , 2008. 5, 9 -
- [49][49] Mika Westerlund. The emergence of deepfake technology: A review. Technology Innovation Management Review , 9(11), 2019. 3 -
- [52][52] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multiattentional deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021. 1, 18 -
Cited by
Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection
Yan, Z., Luo, Y., Lyu, S., Liu, Q., & Wu, B. (2024) · 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52733.2024.00858 · Cited by 166
Self-citationFake News Detection Using Deep Learning: A Systematic Literature Review
Alnabhan, M. Q., & Branco, P. (2024) · IEEE Access · doi:10.1109/access.2024.3435497 · Cited by 83
Yan, Z., Zhao, Y., Shen, C., Guo, M., Fu, X., Yao, T., Ding, S., Wu, Y., & Li, Y. (2025) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52734.2025.01177 · Cited by 116
Self-citationDeepfake video detection methods, approaches, and challenges
Alrashoud, M. (2025) · Alexandria Engineering Journal · doi:10.1016/j.aej.2025.04.007 · Cited by 49
X2-DFD: A framework for explainable and extendable Deepfake Detection
Chen, Y., Yan, Z., Cheng, G., Zhao, K., Lyu, S., & Wu, B. (2025) · neural information processing systems · doi:10.52202/085713-2809 · Cited by 56
Self-citation
Metadata
- Authors
- Yan, Z., Zhang, Y., Yuan, X., Lyu, S., & Wu, B.
- Venue
- Advances in Neural Information Processing Systems 36
- Published
- 2023
- Pages
- 4534-4565
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W7133207408
- Cited by
- 98
- References
- 0
- Citations by year
- 2026: 66 · 2025: 16 · 2024: 15 · 2023: 1
- FWCI
- 10.45
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
- Physical Sciences › Computer Science › Signal Processing › Biometric Identification and Security
- Keywords
- deepfake detection, benchmark, synthetic media, video manipulation, evaluation metrics
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 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (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: DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection. https://app.cochat.ai/papers/10.52202/075280-0201
- The paper
- Yan, Z., Zhang, Y., Yuan, X., Lyu, S., & Wu, B. (2023). DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection. Advances in Neural Information Processing Systems 36. https://doi.org/10.52202/075280-0201
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.