Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Finetuning a standard pretrained vision model—a single linear head on a CLIP backbone, trained at a low learning rate with cosine annealing—lifts Deepfake-Eval-2024 accuracy from 63% to 81%, within one point of the leading commercial detect

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Finetuned CLIP-pretrained ConvNeXt-base and ViT-b32 classifiers reach 81% accuracy on Deepfake-Eval-2024, within noise of the leading commercial detector's 82%.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful tuning study on a new in-the-wild benchmark, but the 81% headline is fragile because the benchmark labels mix AI-generated, edited, and source-trust-judged images—and the paper itself admits that. the 3 major comments →

arxiv 2509.04150 v1 pith:CSC267TB submitted 2025-09-04 cs.CV

Revisiting Simple Baselines for In-The-Wild Deepfake Detection

Orlando Castaneda , Kevin So-Tang , Kshitij Gurung This is my paper
classification cs.CV
keywords deepfake detectionin-the-wild benchmarkDeepfake-Eval-2024CLIP pretrainingcosine annealingfine-tuningViT-b32ConvNeXt-base
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that the weak open-source baseline on the Deepfake-Eval-2024 benchmark was not inherently weak—it was undertuned. Taking the same simple recipe of a standard pretrained vision model with a dropout layer and one linear classifier, and finetuning the whole network at a low learning rate with cosine annealing, pushes test accuracy from the previously reported 63% to 81%. That result lands within one point of the leading commercial detector (82%) and beats the previous best open-source model (69%) by 12 points. The authors also show CLIP pretraining matters more than ImageNet pretraining, and that a ViT-b32 model delivers near-top accuracy at a fraction of the compute of ConvNeXt-base. A reader should care because it suggests open-source, inspectable detectors can compete with proprietary ones on realistic social-media data.

Core claim

The central claim is that the 18-point gap between the simple pretrained-features baseline and commercial detectors on Deepfake-Eval-2024 was an artifact of poor hyperparameter choice, not a ceiling of the method. Finetuning all weights of CLIP-pretrained ViT-b32 and ConvNeXt-base models with an initial learning rate of 1e-5 and cosine annealing with warm restarts yields 81% accuracy on the benchmark's fixed 40% test split, close to the 82% reported for the best commercial model. The same tuning raises the method's own prior result from 63% and outperforms all previously reported open-source models. The paper further claims that this gain comes with practical tradeoffs: ResNet-50 is smaller

What carries the argument

The mechanism is a two-part adaptation of standard pretrained vision models: a lightweight classification head (dropout plus a single linear layer) on a CLIP-pretrained backbone, with the entire network unfrozen and finetuned at a low initial learning rate (1e-5) under cosine-annealing warm restarts. The low learning rate and annealing schedule let the pretrained features shift toward deepfake-specific cues rather than being overwritten or stuck in sharp minima; unfreezing the backbone is what enables both the accuracy gain and the GradCAM interpretability. CLIP pretraining on a large 2021 web-scale image-text corpus supplies features better matched to social-media imagery than ImageNet feat

Load-bearing premise

The benchmark's ground-truth labels reliably measure what we mean by deepfake, and the fixed 60/40 split's 789 test images give a stable estimate of accuracy; if labels capture source trustworthiness or any editing rather than AI generation, or if the split is noisy, the 81% claim and the one-point gap to the commercial detector weaken.

What would settle it

Run the same finetuning recipe on a freshly drawn 60/40 split of Deepfake-Eval-2024, or on an independent batch of 2024 social-media images with labels assigned by forensic review rather than source trustworthiness. If accuracy falls back toward the original 63%, or the 81%-versus-82% ordering flips across splits, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On Deepfake-Eval-2024, open-source detectors are competitive with the leading commercial detector: 81% vs 82%, with the previous open-source best at 69%.
  • Hyperparameter choices—initial learning rate, scheduler, and whether the backbone is frozen—are decisive for this baseline; cosine annealing at 1e-5 is worth roughly 18 points over the original report.
  • CLIP-pretrained backbones generalize better than ImageNet-pretrained or randomly initialized ones on this social-media benchmark, so detector performance depends heavily on the pretraining distribution's recency and scale.
  • Deployment tradeoffs separate the models: ViT-b32 gives near-best accuracy with about 8.8 GFLOPs and 16 ms inference, while ConvNeXt-base gives the best operating-characteristic curve but costs about 30.7 GFLOPs; ResNet-50 remains an option for memory-limited settings at 79%.
  • A ConvNeXt-base detector can flag more than 30% of fakes with zero false positives, making it usable as a precision-first prefilter for human review.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because Deepfake-Eval-2024's labels partly reflect source trustworthiness and conventional editing, the 81% figure may be measuring 'inauthentic or untrustworthy image' rather than strictly AI-generated deepfake; retesting on a purely generative corpus would separate those constructs.
  • The single 60/40 split with 789 test images means the 81%-versus-82% gap is within plausible sampling error; the robust conclusion is parity, not superiority, and a multi-split or bootstrapped evaluation would sharpen it.
  • If the parity holds, deployment incentives shift: organizations could run transparent local detectors instead of paid proprietary APIs, and detection benchmarks could emphasize calibration and uncertainty rather than a single accuracy number.
  • The optimization finding suggests other 'simple baselines' on in-the-wild benchmarks may be similarly undertuned; low-rate cosine-annealing finetuning is a cheap check before declaring a method saturated.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper revisits the Ojha et al. baseline for deepfake detection on the Deepfake-Eval-2024 benchmark. Using standard pretrained backbones (ResNet-50, ViT-b32, ConvNeXt-base) with a final dropout + linear layer, the authors tune hyperparameters such as learning rate, scheduler, pretraining source, and finetuning vs. freezing. They report that a CLIP-pretrained ConvNeXt-base or ViT-b32, finetuned with a low learning rate and cosine annealing, reaches 81% test accuracy on the benchmark's single 60/40 split, compared to the 63% originally reported for the same baseline and the 82% of the leading commercial detector. They also report computational costs, ROC AUC, and GradCAM visualizations, and release code.

Significance. If correct, this is a practically useful demonstration that simple open-source detectors can be competitive with proprietary systems on an in-the-wild benchmark, and it highlights the sensitivity of such baselines to hyperparameter choices. The paper's strengths include a clear hyperparameter ablation, evaluation on an external benchmark (not circular), a focus on compute/interpretability tradeoffs, and released code. However, the central comparative claim rests on small accuracy differences on a single test split and on benchmark labels that the paper itself acknowledges are not purely AI-generated; as presented, the evidence does not firmly establish that open-source detectors 'compete' with the commercial detector on the construct of deepfake detection.

major comments (3)
  1. [Section 4.3 / Table 2] The headline comparison (81% vs. 82%, 79% vs. 81%) is reported without confidence intervals, bootstrap estimates, or multiple random seeds. With n=789 test images, the standard error of accuracy is roughly sqrt(0.81*0.19/789) ≈ 1.4 percentage points, so a 1–2 point gap is within sampling noise. The claim that these open-source models 'closely follow' or 'compete with' the commercial detector is not statistically supported. Please report CIs, bootstrap intervals, or at least results over several seeds and random splits.
  2. [Section 4.4] The paper states: 'Not all fake images were AI-generated' and that TrueMedia.org reviewers 'labeled the ground truth of these image based on the trustworthiness of the original source or if the photo appears to be edited in any manner.' This means the 81% accuracy is for a 'fake/inauthentic' class that mixes AI generation, conventional editing, and source-reputation judgments, not purely deepfakes. The central claim about deepfake-detection competitiveness is therefore ambiguous. Please re-analyze results on a subset of images that are clearly AI-generated (or otherwise filter by label rationale), and temper the conclusions accordingly.
  3. [Section 3.1 / Table 1] Hyperparameters were selected using only 116 validation images drawn from the same single split, and the final test result is from that same split. This creates a risk of selection overfitting and does not measure generalization to other splits. Additionally, the paper compares its 81% to the 63% reported in [5] without rerunning the original baseline under the same protocol (augmentation, resolution, epochs, optimizer). To support an '18% improvement' claim, please include a controlled comparison where the Ojha-style baseline is retrained with the same preprocessing and training setup (except the hyperparameters under study), and ideally evaluate across multiple splits.
minor comments (4)
  1. [Section 1 / Abstract] Typo: 'commerial' should be 'commercial.' Also, 'signficant' in Section 4.2 and 'to deployed' in Section 4.3 need correction.
  2. [Section 3.3] The training resolution is 256 pixels after random crop, while pretrained models often expect 224; please clarify how the positional/input embeddings of ViT-b32 handle this discrepancy.
  3. [Section 4.1] 'mean average precision' should be defined as AP averaged over classes, and it is used later as 'mean precision' in Figure 4; please standardize terminology.
  4. [Section 4.4] The last row of Figure 5 is described as 'appears be a real photo' — grammatical fix and a clearer description of what is shown would help.

Circularity Check

0 steps flagged

No significant circularity: the 81% test accuracy is an empirical result on a fixed external benchmark split, not a fitted value, a renamed known result, or a self-citation chain.

full rationale

The paper's central claim is empirical, not definitional. Section 3.1 states: 'we adhere to the same finetuning split prescribed by Deepfake-Eval-2024 which designates 60% (1161 images) for training and 40% (789 images) for testing,' and additionally: 'From the training data, we further reserve 10% (116 images) of training samples for validation during our hyperparameter search.' Hyperparameters are selected on the validation subset, and the reported 81% is measured on the held-out test split. No test accuracy is used as a training signal; no fitted parameter is renamed as a prediction. The comparison to the previously reported 63% baseline is a comparison to an externally published number obtained under a different training protocol, which is a measurement-comparability/fairness concern, not circularity. The model family (simple linear classifier over pretrained features) is attributed to Ojha et al. [17], and the benchmark to Chandra et al. [5]; no load-bearing self-citation occurs in the method or evaluation chain. Section 4.4 explicitly flags the benchmark's label-validity limits ('Not all fake images were AI-generated', 'TrueMedia.org reviewers often labeled the ground truth of these image based on the trustworthiness of the original source or if the photo appears to be edited in any manner'), which is an external-validity caveat, not a circular step. The 81% claim therefore does not reduce by construction to the paper's inputs; it is a genuine held-out evaluation result.

Axiom & Free-Parameter Ledger

7 free parameters · 3 axioms · 0 invented entities

The contribution is almost entirely empirical: the 81% number is produced by a tuned training configuration evaluated on a single small benchmark split. No new math, measurements, or entities are introduced. The result leans on the benchmark's label quality, the third-party pretrained checkpoints, and validation-based hyperparameter selection.

free parameters (7)
  • initial learning rate = 1e-5 (best for CLIP ViT-b32 and ConvNeXt-base), from grid {1e-3, 1e-4, 1e-5}
    Tuned on the 116-image validation set; the abstract's 81% claim depends on this low learning rate.
  • learning rate scheduler = cosine annealing (vs step decay)
    Chosen by validation accuracy; Table 1 shows scheduler choice shifts accuracy by up to ~10 points for large models.
  • backbone pretraining source = CLIP on LAION-400M (vs ImageNet, random init)
    Best performing option selected from the Table 1 grid; the paper attributes gains to CLIP exposure to modern internet images.
  • backbone finetuning vs frozen = finetune all weights
    Preliminary experiments (Section 4.2) showed frozen backbones, as in the original Ojha linear-probe setup, perform worse; the 81% result requires finetuning, a method change from the baseline being revisited.
  • dropout rate and L2 penalty = moderate levels, exact values not reported
    Chosen by hand; values are absent from the paper, hurting reproducibility.
  • early stopping patience = 5 epochs
    Fixed choice affecting training length and final model state.
  • random crop scale and resolution = 50-100% crop, resize to 256
    Data augmentation choices made at training time; not swept.
axioms (3)
  • domain assumption Deepfake-Eval-2024 ground-truth labels are valid and correct
    The paper inherits labels from forensic analysts and TrueMedia.org (Sections 2, 4.4); Section 4.4 notes labels reflect source trustworthiness and include non-AI edits, so the construct may be inauthenticity rather than AI generation.
  • domain assumption The single 60/40 split gives a stable accuracy estimate
    Section 3.1 uses the benchmark's prescribed split; 789 test images, no cross-validation, no confidence intervals, so reported accuracies have roughly +/-3% sampling error.
  • domain assumption Third-party pretrained checkpoints (ImageNet, CLIP/OpenCLIP) are valid and their training exposure is as described
    The performance claim depends on the quality and provenance of checkpoints referenced through OpenCLIP profiles [10]; any leakage or mismatch would shift results.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Simple Baselines for In-The-Wild Deepfake Detection." pith.science (2026). https://pith.science/paper/CSC267TB

@misc{pith2026250904150,
  author       = {Pith},
  title        = {Pith review of: Revisiting Simple Baselines for In-The-Wild Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSC267TB}},
  note         = {Machine review of arXiv:2509.04150}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The widespread adoption of synthetic media demands accessible deepfake detectors and realistic benchmarks. While most existing research evaluates deepfake detectors on highly controlled datasets, we focus on the recently released "in-the-wild" benchmark, Deepfake-Eval-2024. Initial reporting on Deepfake-Eval-2024 showed that three finetuned open-source models achieve accuracies between 61% and 69%, significantly lagging behind the leading commercial deepfake detector with 82% accuracy. Our work revisits one of these baseline approaches, originally introduced by Ojha et al., which adapts standard pretrained vision backbones to produce generalizable deepfake detectors. We demonstrate that with better-tuned hyperparameters, this simple approach actually yields much higher performance -- 81% accuracy on Deepfake-Eval-2024 -- surpassing the previously reported accuracy of this baseline approach by 18% and competing with commercial deepfake detectors. We discuss tradeoffs in accuracy, computational costs, and interpretability, focusing on how practical these deepfake detectors might be when deployed in real-world settings. Our code can be found at https://github.com/Deepfake-Detection-KKO/deepfake-detection.

Figures

Figures reproduced from arXiv: 2509.04150 by Kevin So-Tang, Kshitij Gurung, Orlando Castaneda.

Figure 1
Figure 1. Figure 1: Real and fake image samples from Deepfake [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows the precision, detection, and false pos￾itive rates of our most accurate ConvNeXt-base, ResNet￾50, and ViT-b32 models on the test dataset. ConvNeXt performs slightly better at most operating points which are summarized by its higher ROC AUC values of 0.89 and mean precision of 0.94. The precision-recall curve specif￾ically shows that our ConvNeXt-base model allows detec￾tion of more than 30% of deepf… view at source ↗
Figure 5
Figure 5. Figure 5: GradCAM Visualizations (important regions in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    Unmasking deepfakes: A systematic review of deepfake detection and generation techniques using artificial intelligence

    Fakhar Abbas and Araz Taeihagh. Unmasking deepfakes: A systematic review of deepfake detection and generation techniques using artificial intelligence. Expert Systems with Applications, 252:124260, 2024. 2

    work page 2024

  2. [2]

    MesoNet: a Compact Facial Video Forgery Detec- tion Network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. MesoNet: a Compact Facial Video Forgery Detec- tion Network. pages 1–7, Dec. 2018. arXiv:1809.00888 [cs]. 2, 7

    Pith/arXiv arXiv 2018

  3. [3]

    Socialdf: Benchmark dataset and detection model for mitigating harmful deepfake content on social media platforms, 2025

    Arnesh Batra, Anushk Kumar, Jashn Khemani, Arush Gum- ber, Arhan Jain, and Somil Gupta. Socialdf: Benchmark dataset and detection model for mitigating harmful deepfake content on social media platforms, 2025. 2

    work page 2025

  4. [4]

    Real-time deepfake detection in the real-world, 2024

    Bar Cavia, Eliahu Horwitz, Tal Reiss, and Yedid Hoshen. Real-time deepfake detection in the real-world, 2024. 2

    work page 2024

  5. [5]

    Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circu- lated in 2024, May 2025

    Nuria Alina Chandra, Ryan Murtfeldt, Lin Qiu, Arnab Kar- makar, Hannah Lee, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Sejin Paik, Changyeon Lee, Jongwook Choi, Aerin Kim, and Oren Etzioni. Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circu- lated in 2024, May 2025. arXiv:2503.02857 [cs]. 1, 2, 3, 5, 6

    Pith/arXiv arXiv 2024

  6. [6]

    Reproducible scal- ing laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scal- ing laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 2818–2829. IEEE, June 2023. 4

    work page 2023

  7. [7]

    Le, Jiwon Kim, Simon Woo, Shahroz Tariq, Alsharif Abuadbba, and Kristen Moore

    Beomsang Cho, Binh M. Le, Jiwon Kim, Simon Woo, Shahroz Tariq, Alsharif Abuadbba, and Kristen Moore. To- wards understanding of deepfake videos in the wild. In Pro- ceedings of the 32nd ACM International Conference on In- formation and Knowledge Management , CIKM ’23, page 4530–4537, New York, NY , USA, 2023. Association for Computing Machinery. 2

    work page 2023

  8. [8]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021. arXiv:2010.11929 [cs]. 2

    Pith/arXiv arXiv 2021

  9. [9]

    Unmasking deepfakes with simple features,

    Ricard Durall, Margret Keuper, Franz-Josef Pfreundt, and Janis Keuper. Unmasking deepfakes with simple features,

  10. [10]

    OpenCLIP Model Profiles

    ML Foundations. OpenCLIP Model Profiles. https://github.com/mlfoundations/open clip/blob/main/ docs/model profile.csv, 2023. [Accessed 25-07-2025]. 4, 5, 6

    work page 2023

  11. [11]

    Deep Residual Learning for Image Recognition, Dec

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, Dec. 2015. arXiv:1512.03385 [cs]. 2

    Pith/arXiv arXiv 2015

  12. [12]

    Deepfake detection using deep learning meth- ods: A systematic and comprehensive review

    Arash Heidari, Nima Jafari Navimipour, Hasan Dag, and Mehmet Unal. Deepfake detection using deep learning meth- ods: A systematic and comprehensive review. WIREs Data Mining and Knowledge Discovery, 14(2):e1520, 2024. 2

    work page 2024

  13. [13]

    Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild

    Xuechen Liu, Xin Wang, Md Sahidullah, Jose Patino, H´ector Delgado, Tomi Kinnunen, Massimiliano Todisco, Junichi Yamagishi, Nicholas Evans, Andreas Nautsch, and Kong Aik Lee. Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2507–2522, 2023. 2

    work page 2021

  14. [14]

    A ConvNet for the 2020s, Mar

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, Mar. 2022. arXiv:2201.03545 [cs]. 2

    Pith/arXiv arXiv 2022

  15. [15]

    The creation and detection of deepfakes: A survey

    Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM Comput. Surv. , 54(1), Jan

    work page

  16. [16]

    A survey on the detection and impacts of deepfakes in visual, audio, and tex- tual formats

    Rami Mubarak, Tariq Alsboui, Omar Alshaikh, Isa Inuwa- Dutse, Saad Khan, and Simon Parkinson. A survey on the detection and impacts of deepfakes in visual, audio, and tex- tual formats. IEEE Access, 11:144497–144529, 2023. 2

    work page 2023

  17. [17]

    Towards Uni- versal Fake Image Detectors that Generalize Across Gen- erative Models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards Uni- versal Fake Image Detectors that Generalize Across Gen- erative Models. In 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 24480– 24489, Vancouver, BC, Canada, June 2023. IEEE. 1, 2

    work page 2023

  18. [18]

    Evaluating deep- fake detectors in the wild, 2025

    Viacheslav Pirogov and Maksim Artemev. Evaluating deep- fake detectors in the wild, 2025. 2

    work page 2025

  19. [19]

    Deepfake videos in the wild: Analysis and detection

    Jiameng Pu, Neal Mangaokar, Lauren Kelly, Parantapa Bhat- tacharya, Kavya Sundaram, Mobin Javed, Bolun Wang, and Bimal Viswanath. Deepfake videos in the wild: Analysis and detection. In Proceedings of the Web Conference 2021, WWW ’21, page 981–992, New York, NY , USA, 2021. As- sociation for Computing Machinery. 2

    work page 2021

  20. [20]

    Do Vision Trans- formers See Like Convolutional Neural Networks?, Mar

    Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do Vision Trans- formers See Like Convolutional Neural Networks?, Mar

    work page

  21. [21]

    Md Shohel Rana, Mohammad Nur Nobi, Beddhu Murali, and Andrew H. Sung. Deepfake Detection: A Systematic Literature Review. IEEE Access, 10:25494–25513, 2022. 2

    work page 2022

  22. [22]

    Con- tinuous fake media detection: Adapting deepfake detectors to new generative techniques

    Francesco Tassone, Luca Maiano, and Irene Amerini. Con- tinuous fake media detection: Adapting deepfake detectors to new generative techniques. Computer Vision and Image Understanding, 249:104143, 2024. 2

    work page 2024

  23. [23]

    Deepfake Video De- tection Using Convolutional Vision Transformer, Mar

    Deressa Wodajo and Solomon Atnafu. Deepfake Video De- tection Using Convolutional Vision Transformer, Mar. 2021. arXiv:2102.11126 [cs]. 7

    Pith/arXiv arXiv 2021

  24. [24]

    T. Zhang. Deepfake generation and detection, a survey. Multimedia Tools and Applications , 81:6259–6276, Febru- ary 2022. 2

    work page 2022

  25. [25]

    Face Forensics in the Wild

    Tianfei Zhou, Wenguan Wang, Zhiyuan Liang, and Jianbing Shen. Face forensics in the wild. CoRR, abs/2103.16076,

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Wilddeepfake: A challenging real-world dataset for deepfake detection

    Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM International Conference on Multimedia , MM ’20, page 2382–2390, New York, NY , USA, 2020. Association for Computing Machinery. 2

    work page 2020

  27. [2022]

    arXiv:2108.08810 [cs]. 5, 7

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.