DFCon: Attention-Driven Supervised Contrastive Learning for Robust Deepfake Detection

Abir Muhtasim; Hafijul Hoque Chowdhury; Mahir Labib Dihan; MD Roqunuzzaman Sojib; MD Sadik Hossain Shanto; MD Tanvir Hassan; M. Saifur Rahman; Rakib Ahsan; Riad Ahmed Anonto; Sheikh Azizul Hakim

REVIEW 3 major objections 5 minor 24 references

DFCon: Attention-Driven Supervised Contrastive Learning for Robust Deepfake Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Three pretrained vision models, fine-tuned with supervised contrastive loss and combined by majority voting, detect deepfakes on the DFWild validation set at 95.83% accuracy, outperforming standard backbones and each model alone.

desk verdict A clean competition pipeline whose headline accuracy is likely inflated by tuning on the same validation set; worth a look for practitioners, not a research advance. read the letter →

arxiv 2501.16704 v1 pith:HZGIRGQD submitted 2025-01-28 cs.CV cs.CRcs.LGeess.IVeess.SP

MD Sadik Hossain Shanto , Mahir Labib Dihan , Souvik Ghosh , Riad Ahmed Anonto , Hafijul Hoque Chowdhury , Abir Muhtasim , Rakib Ahsan , MD Tanvir Hassan

show 3 more authors

MD Roqunuzzaman Sojib Sheikh Azizul Hakim M. Saifur Rahman

This is my paper · ORCID

classification cs.CVcs.CRcs.LGeess.IVeess.SP

keywords deepfakedetectionsupervisedcontrastivelearningensemblevisiontransformerMaxViTCoAtNetEVA-02DFWildvalidation

verification ladder T0 review T1 audit T2 compute T3 formal

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 sets out to build a deepfake detector that generalizes across diverse face-manipulation methods, using the DFWild-Cup validation set as its benchmark. Its recipe is two-stage: first fine-tune three complementary pretrained vision models (MaxViT, CoAtNet, EVA-02) with supervised contrastive loss so real and fake embeddings form separate clusters; then freeze the backbones, train small classifiers, and combine their votes by majority. The reported payoff is 95.83% validation accuracy, clearly above every conventional baseline and above each backbone alone. A sympathetic reader should care because the claim is that architecture complementarity plus contrastive embedding separation, not a single new network, drives robustness to unseen deepfake types.

What carries the argument

The central mechanism is supervised contrastive (SupCon) loss computed on backbone embeddings during the first training stage, followed by frozen-backbone MLP classifiers trained with BCE loss and a majority-voting fusion of three models. SupCon normalizes the embeddings, builds pairwise similarity, pulls same-label (real-real, fake-fake) embeddings together, and pushes different-label embeddings apart, producing distinct clusters that make the later classifiers' job easier. The voting step combines MaxViT's local-artifact sensitivity, CoAtNet's multi-scale hybrid features, and EVA-02's global contextual understanding.

What would settle it

Run the trained ensemble on the competition's held-out test set or on deepfakes produced by generation methods absent from the eight training datasets, and compare accuracy with the 95.83% validation number; a substantial drop would show the reported robustness was specific to the validation distribution.

Watch

Extended reading notes

Core claim

The paper's central claim is that a two-stage pipeline—fine-tuning three complementary pretrained vision backbones (MaxViT, CoAtNet, EVA-02) with supervised contrastive loss, freezing them, training lightweight MLP classifiers, and fusing predictions by majority voting—detects deepfakes more reliably than any single component or standard ImageNet backbone. On the DFWild validation set the ensemble reaches 95.83% accuracy, 0.9586 F1, and 0.9807 AUC, against a best baseline of 0.8724 (InceptionResNetV2) and a best single ensemble member of 0.9508 (MaxViT). The authors attribute the gain to complementary feature extraction and to supervised contrastive loss creating well-separated real/fake embedding clusters, visualized with t-SNE showing overlapping distributions before training and separated clusters afterward.

Load-bearing premise

The DFWild validation set is used to choose the three backbones, to schedule the learning rate, and to report the 95.83% accuracy, so the central claim assumes this validation set faithfully represents performance on truly unseen deepfake data.

Editorial extensions

If this is right

If the ensemble generalizes as claimed, large-scale ImageNet pretraining plus supervised contrastive fine-tuning can yield deepfake detectors that work across generators not seen during training.
The ablation indicates that online augmentation is the most influential single component, since removing it drops CoAtNet accuracy from 0.9447 to 0.8659, so data diversity deserves as much attention as architecture choice.
Majority voting improves on the best individual model by about one percentage point (0.9508 to 0.9583), showing complementary vision transformers can be combined without retraining the backbones.
Replacing SupCon with BCE loss lowers CoAtNet accuracy to 0.9163, suggesting the contrastive objective contributes beyond what the final classifier alone provides.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper selects its three backbones, schedules learning rates, and reports its final number all on the same DFWild validation set, so the 95.83% figure is likely an optimistic estimate of performance on a truly novel deepfake distribution; a held-out competition test set would be needed to quantify the gap.
The same two-stage recipe of contrastive fine-tuning, frozen features, and a lightweight head could transfer to other binary image-forensics tasks such as GAN-image detection or copy-move detection, provided the augmentation diversity covers those artifact types.
A testable extension would be to replace majority voting with soft voting or a learned fusion layer, since the paper's rule only keeps the highest or lowest probability and discards some confidence information from the middle model.
Because all three backbones carry ImageNet-scale pretraining, part of the gain may come from pretraining scale rather than the deepfake-specific objective; ablating SupCon with a randomly initialized small backbone could separate these effects.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Hi [Colleague],

Quick take: this is a well-executed competition report, not a research paper. The headline 95.83% on DFWild validation is probably a few points optimistic because the same validation set was used to pick these three backbones out of five, schedule the learning rate, and choose the ensemble. So the number is not an unbiased estimate of how it will do on unseen data. Nothing here is conceptually new—SupCon, MaxViT/CoAtNet/EVA-02, and majority voting are all published—but the engineering is clean and the ablation study gives the components some support.

What I like: the two-stage freeze-then-classify scheme is described precisely enough to reproduce, and the ablation showing that offline augmentation, online augmentation, and SupCon each matter on CoAtNet is useful evidence. Generating 12,200 extra fake images from seven modern forgery methods is a reasonable move for a competition and nicely documented. The tables are internally consistent.

The soft spots are the usual ones for competition papers. No independent test set, no repeated runs or error bars, no code or data release. The baseline comparison is single-run, so we don't know if the 8-point gap over InceptionResNetV2 is meaningful. Claims about fairness and generalization to 'unseen scenarios' rely on 'we expect' rather than measurements (Section III-D). The ablation only covers CoAtNet, so we don't get a read on whether SupCon helps the ensemble as a whole. The selection itself—five backbones evaluated on the validation set, then top three kept—inflates the reported accuracy, and with only ~3,072 validation images, the winner's curse is likely non-trivial.

Who is this for? Teams entering similar competitions, and people building deepfake detection pipelines who want a set of components that work together. It is not a methodological contribution.

If this lands on my desk, I'd send it to review, because the empirical content is real and the flaws are fixable. But I'd require a proper held-out evaluation or at least confidence intervals, artifact release, and toned-down generalization claims before accepting. As a preprint, it's a useful data point for the competition.

Referee Report

3 major / 5 minor

Summary. This paper, a technical report for the IEEE SP Cup 2025, proposes DFCon, a deepfake detection system that ensembles three pretrained vision backbones (MaxViT, CoAtNet, and EVA-02). The method first fine-tunes each backbone with supervised contrastive loss on a balanced subset of the DFWild training data, then freezes the backbones and trains MLP classifiers with BCE loss, and finally combines the three classifiers by majority voting. On the DFWild validation set, the ensemble is reported to achieve 95.83% accuracy, 0.9586 F1 score, and 0.9807 AUC, outperforming the individual backbones and several ResNet/Inception baselines. The paper also includes an ablation study on CoAtNet and t-SNE visualizations of the learned embeddings.

Significance. The engineering recipe is sensible and the paper is transparent about many training details, including model variants, hyperparameters, per-epoch training times, and an ablation study. If the reported validation accuracy were confirmed on an independent test set, the system would be a competitive solution for the DFWild-Cup. However, as presented, the evidence supports only a descriptive statement about this particular validation partition. The absence of a held-out test set, the use of the validation set for model selection and learning-rate scheduling, and the lack of uncertainty estimates mean that the paper's broader generalization and robustness claims are not yet substantiated. The strengths of the paper are its clear pipeline description, the comparison of five backbones, and the explicit reporting of the validation-based selection process; these are useful for practitioners who want to reproduce the result.

major comments (3)

[III-B, III-C, II-E] The validation set is used for model selection (Table V: choosing MaxViT, CoAtNet, and EVA-02 from five candidates), for learning-rate scheduling (ReduceLROnPlateau in Section II-E), and for the final accuracy claim (Table VI). Consequently, the reported 95.83% accuracy is an in-sample, selection-biased estimate and cannot be interpreted as a measure of generalization to unseen deepfakes. The abstract and conclusion nevertheless claim generalization to 'unseen scenarios' and 'real-world conditions', but no independent test set, cross-validation, or repeated-seed results are provided. This overclaim is load-bearing because the reported performance is the paper's main contribution.
[III-D] The paragraph on diverse scenarios states that the authors 'expect that these strategies will make our model both generalizable and fair across diverse scenarios' and contains no quantitative evaluation on age, ethnicity, or other scenario shifts. This is an explicit limitation in the text, yet the conclusion asserts that ensembling 'has improved accuracy, ensuring better performance in diverse and unseen datasets.' The paper needs either actual evaluations on held-out scenario splits or a substantial moderation of the generalization claims.
[II-F and Table II] The supervised contrastive loss is described only in prose; the exact loss formula, the temperature parameter, and any projection head are not specified. Since the SupCon temperature is a known critical hyperparameter and is not listed in Table II, the reported results are not reproducible from the manuscript. This is a technical omission in the core training procedure and should be corrected with the loss equation and the temperature value.

minor comments (5)

[II-A and II-G] The dataset description is inconsistent: Section II-A reports 42,690 real and 219,470 fake training images, while Section II-G says three subsets of 73,157 fake images are sampled (which sums to 219,471) and refers to the 12,200 generated fake images as 'from the provided dataset' rather than from the secondary generation described in Section II-B. Please reconcile these numbers and the wording.
[III-B, Table V] The validation set contains only 3,072 images, so accuracy differences of less than about one percentage point may not be statistically meaningful; no confidence intervals or repeated-seed variability are reported, making the ranking of backbones less definitive than the table suggests.
[Figure 1 and II-I] The labels 'CoatNet' and 'EV A-02' in Figure 1 are inconsistent with the text's 'CoAtNet' and 'EVA-02', and 'V oting' in Section II-I should be 'Voting'.
[Figure 2 caption] The t-SNE visualizations are generated from 2,000 real and 2,000 fake images from the training dataset; the caption should state this explicitly and the paper should avoid implying that the plots show validation-set behavior.
[III-A] The phrase 'robustness and generalization capability' overstates what a single validation-set accuracy can demonstrate; 'performance on the DFWild validation set' would be a more accurate description of the reported baseline numbers.

Circularity Check

1 steps flagged · score 3.0 of 10

The headline 95.83% accuracy is measured on the same DFWild validation set used for backbone selection, ensemble choice, and learning-rate scheduling, so it is a selection-tuned score rather than an unbiased estimate of generalization to unseen scenarios.

fitted input called prediction [Section III-B/III-C (Model Selection Results; Ensembling Results); also Section II-E (Hyperparameters)]
"We evaluated advanced backbone architectures, including MaxViT, CoAtNet, EV A-02, ConvNeXtV2, and EfficientNetV2, based on accuracy, AUC, and parameter efficiency, as summarized in Table V. ... These findings confirmed the choice of MaxViT, CoAtNet, and EV A-02 as the backbone models for our proposed framework due to their complementary strengths and superior metrics. ... As shown in Table VI, the ensemble outperformed individual models, achieving an F1 score of 0.9586, an accuracy of 0.9583, and an AUC of 0.9807."

The DFWild validation set is used both for model selection and for the final accuracy claim. In Section III-B, five backbone architectures are compared on the validation set and the top three (MaxViT, CoAtNet, EV A-02) are chosen by their validation accuracy and AUC; in Section III-C, the ensemble is selected and its 0.9583 accuracy is reported on the same validation partition; in Section II-E, ReduceLROnPlateau adjusts the learning rate using validation loss. Thus the headline number is the score of the configuration that was explicitly chosen to maximize performance on that very set.

full rationale

The paper's method is not definitionally circular: the supervised contrastive loss, frozen-backbone classifier, and majority-voting ensemble are all described independently, and the reported accuracy is an empirical measurement rather than a quantity defined as its own input. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation; the backbones and losses come from external sources with their own independent benchmarks. The only circularity concern is in the evaluation protocol: the same DFWild validation set is used for backbone selection (Section III-B), ensemble choice (Section III-C), learning-rate scheduling (Section II-E), and final accuracy reporting. This means the 95.83% headline is positively biased relative to a truly held-out estimate, and claims about 'unseen scenarios' and 'real-world conditions' are not supported by an independent test measurement. This is a real but limited circularity in the performance claim, not in the derivation of the method, so the score is moderate rather than high.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper contains no derivation; the central claim is a measured validation accuracy. The free parameters are the hand-set hyperparameters and design choices that the result depends on. The axioms are domain assumptions about dataset representativeness and transferability that are not validated independently. No invented entities.

free parameters (5)

Backbone learning rate = 3e-5
Set by hand for all three backbones; the reported accuracy depends on this choice.
Classifier learning rate = 5e-5
Chosen by hand for the MLP classifier heads.
Weight decay = 1e-2
AdamW weight decay for backbones, chosen by hand.
Epochs per model = Backbone: 4, 2, 6; Classifier: 8, 8, 7
Selected based on validation performance; the result is sensitive to early stopping.
SupCon temperature = not reported
The supervised contrastive loss temperature is never specified, but it controls feature separation and downstream accuracy.

assumptions (3)

domain assumption The DFWild validation set is representative of unseen deepfake scenarios.
The paper claims generalization to unseen scenarios based on validation accuracy, but this set is also used for model selection and early stopping, so it is not an unbiased holdout.
domain assumption Pretrained ImageNet backbones transfer to deepfake detection.
The approach relies on backbones pretrained on ImageNet-21k/22k/12k without additional domain-specific pretraining, an assumption that is standard but untested here.
ad hoc to paper Adding fake images from seven external generation methods improves robustness.
The paper generates 12,200 extra fake images but provides no ablation showing their contribution; this is an ad hoc design choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DFCon: Attention-Driven Supervised Contrastive Learning for Robust Deepfake Detection." pith.science (2026). https://pith.science/paper/HZGIRGQD

@misc{pith2026250116704,
  author       = {Pith},
  title        = {Pith review of: DFCon: Attention-Driven Supervised Contrastive Learning for Robust Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZGIRGQD}},
  note         = {Machine review of arXiv:2501.16704}
}

read the original abstract

This report presents our approach for the IEEE SP Cup 2025: Deepfake Face Detection in the Wild (DFWild-Cup), focusing on detecting deepfakes across diverse datasets. Our methodology employs advanced backbone models, including MaxViT, CoAtNet, and EVA-02, fine-tuned using supervised contrastive loss to enhance feature separation. These models were specifically chosen for their complementary strengths. Integration of convolution layers and strided attention in MaxViT is well-suited for detecting local features. In contrast, hybrid use of convolution and attention mechanisms in CoAtNet effectively captures multi-scale features. Robust pretraining with masked image modeling of EVA-02 excels at capturing global features. After training, we freeze the parameters of these models and train the classification heads. Finally, a majority voting ensemble is employed to combine the predictions from these models, improving robustness and generalization to unseen scenarios. The proposed system addresses the challenges of detecting deepfakes in real-world conditions and achieves a commendable accuracy of 95.83% on the validation dataset.

Figures

Figures reproduced from arXiv: 2501.16704 by the authors.

**Figure 2.** t-SNE visualization of feature embeddings before and after training for MaxViT, EVA-02, and CoAtNet. The plots [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 12 canonical work pages

[1]

Maxvit: Multi-axis vision transformer,

Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y . Li, “Maxvit: Multi-axis vision transformer,” 2022. [Online]. Available: https://arxiv.org/abs/2204.01697

arXiv 2022
[2]

Coatnet: Marrying convolution and attention for all data sizes,

Z. Dai, H. Liu, Q. V . Le, and M. Tan, “Coatnet: Marrying convolution and attention for all data sizes,” 2021. [Online]. Available: https://arxiv.org/abs/2106.04803

arXiv 2021
[3]

Eva-02: A visual representation for neon genesis,

Y . Fang, Q. Sun, X. Wang, T. Huang, X. Wang, and Y . Cao, “Eva-02: A visual representation for neon genesis,” Image and Vision Computing, vol. 149, p. 105171, Sep. 2024. [Online]. Available: http://dx.doi.org/10.1016/j.imavis.2024.105171

work page arXiv 2024
[4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

arXiv 2021
[5]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” 2021. [Online]. Available: https://arxiv.org/abs/2004.11362

arXiv 2021
[6]

Deepfakebench: A comprehensive benchmark of deepfake detection,

Z. Yan, Y . Zhang, X. Yuan, S. Lyu, and B. Wu, “Deepfakebench: A comprehensive benchmark of deepfake detection,” inAdvances in Neural Information Processing Systems, vol. 36. Curran Associates, Inc., 2023, pp. 4534–4565

work page 2023
[7]

Celeb-df: A large- scale challenging dataset for deepfake forensics,

Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-df: A large- scale challenging dataset for deepfake forensics,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3207–3216

2020
[8]

Faceforensics++: Learning to detect manipulated facial images,

A. R ¨ossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “Faceforensics++: Learning to detect manipulated facial images,” 2019. [Online]. Available: https://arxiv.org/abs/1901.08971

arXiv 2019

Show all 24 references

[9]

Faceshifter: Towards high fidelity and occlusion aware face swapping,

L. Li, J. Bao, H. Yang, D. Chen, and F. Wen, “Faceshifter: Towards high fidelity and occlusion aware face swapping,” 2020. [Online]. Available: https://arxiv.org/abs/1912.13457

2020 arXiv
[10]

The deepfake detection challenge (dfdc) preview dataset,

B. Dolhansky, R. Howes, B. Pflaum, N. Baram, and C. C. Ferrer, “The deepfake detection challenge (dfdc) preview dataset,” 2019. [Online]. Available: https://arxiv.org/abs/1910.08854

2019 arXiv
[11]

Fine-grained face swapping via regional gan inversion,

Z. Liu, M. Li, Y . Zhang, C. Wang, Q. Zhang, J. Wang, and Y . Nie, “Fine-grained face swapping via regional gan inversion,” 2023. [Online]. Available: https://arxiv.org/abs/2211.14068

2023 arXiv
[12]

Facedancer: Pose- and occlusion-aware high fidelity face swapping,

F. Rosberg, E. E. Aksoy, F. Alonso-Fernandez, and C. Englund, “Facedancer: Pose- and occlusion-aware high fidelity face swapping,”
[13]

Blendface: Re-designing identity encoders for face-swapping,

K. Shiohara, X. Yang, and T. Taketomi, “Blendface: Re-designing identity encoders for face-swapping,” 2023. [Online]. Available: https://arxiv.org/abs/2307.10854

2023 arXiv
[14]

inswapper,

H. Wang, “inswapper,” https://github.com/haofanwang/inswapper, 2023, accessed: January 12, 2025

2023
[15]

Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces,

S. Bounareli, C. Tzelepis, V . Argyriou, I. Patras, and G. Tzimiropoulos, “Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces,” in Proceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV) , 2023

2023
[16]

Designing an encoder for stylegan image manipulation,

O. Tov, Y . Alaluf, Y . Nitzan, O. Patashnik, and D. Cohen-Or, “Designing an encoder for stylegan image manipulation,” 2021. [Online]. Available: https://arxiv.org/abs/2102.02766

2021 arXiv
[17]

Styleclip: Text-driven manipulation of stylegan imagery,

O. Patashnik, Z. Wu, E. Shechtman, D. Cohen-Or, and D. Lischinski, “Styleclip: Text-driven manipulation of stylegan imagery,” 2021. [Online]. Available: https://arxiv.org/abs/2103.17249

2021 arXiv
[18]

Huggingface: State-of-the-art natural language process- ing,

HuggingFace, “Huggingface: State-of-the-art natural language process- ing,” 2020, available at: https://huggingface.co/

2020
[19]

Pytorch image models (timm),

R. Wightman, “Pytorch image models (timm),” 2019, available at: https: //github.com/rwightman/pytorch-image-models

2019
[20]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

2009
[21]

Supcontrast: Supervised contrastive learning,

H. Long, “Supcontrast: Supervised contrastive learning,” https://github. com/HobbitLong/SupContrast, accessed: 2025-01-12

2025
[22]

Kaggle: Your home for data science,

“Kaggle: Your home for data science,” 2023, accessed: 2024-12-12. [Online]. Available: https://www.kaggle.com

2023
[23]

Signal processing cup,

IEEE Signal Processing Society, “Signal processing cup,” https://signalprocessingsociety.org/community-involvement/ signal-processing-cup, accessed: 2025-01-12

2025
[2022]

Available: https://arxiv.org/abs/2210.10473

[Online]. Available: https://arxiv.org/abs/2210.10473

arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Maxvit: Multi-axis vision transformer,

Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y . Li, “Maxvit: Multi-axis vision transformer,” 2022. [Online]. Available: https://arxiv.org/abs/2204.01697

arXiv 2022

[2] [2]

Coatnet: Marrying convolution and attention for all data sizes,

Z. Dai, H. Liu, Q. V . Le, and M. Tan, “Coatnet: Marrying convolution and attention for all data sizes,” 2021. [Online]. Available: https://arxiv.org/abs/2106.04803

arXiv 2021

[3] [3]

Eva-02: A visual representation for neon genesis,

Y . Fang, Q. Sun, X. Wang, T. Huang, X. Wang, and Y . Cao, “Eva-02: A visual representation for neon genesis,” Image and Vision Computing, vol. 149, p. 105171, Sep. 2024. [Online]. Available: http://dx.doi.org/10.1016/j.imavis.2024.105171

work page arXiv 2024

[4] [4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

arXiv 2021

[5] [5]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” 2021. [Online]. Available: https://arxiv.org/abs/2004.11362

arXiv 2021

[6] [6]

Deepfakebench: A comprehensive benchmark of deepfake detection,

Z. Yan, Y . Zhang, X. Yuan, S. Lyu, and B. Wu, “Deepfakebench: A comprehensive benchmark of deepfake detection,” inAdvances in Neural Information Processing Systems, vol. 36. Curran Associates, Inc., 2023, pp. 4534–4565

work page 2023

[7] [7]

Celeb-df: A large- scale challenging dataset for deepfake forensics,

Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-df: A large- scale challenging dataset for deepfake forensics,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3207–3216

2020

[8] [8]

Faceforensics++: Learning to detect manipulated facial images,

A. R ¨ossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “Faceforensics++: Learning to detect manipulated facial images,” 2019. [Online]. Available: https://arxiv.org/abs/1901.08971

arXiv 2019