Pith. sign in

REVIEW 4 major objections 5 minor 87 references

Trident: Detecting Face Forgeries with Adversarial Triplet Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Trident detects face forgeries by training on triplets that fix identity and scene, so the learned embedding captures manipulation itself, and it reports state-of-the-art or near-state-of-the-art results on FaceForensics++ and several…

desk verdict Plausible combination of triplet learning and domain-adversarial training with a genuinely controlled triplet construction, but the headline cross-dataset SOTA rests on unreported baseline reproductions and should not be taken at face value yet. read the letter →

arxiv 2506.23189 v1 pith:AHO63HO2 submitted 2025-06-29 cs.CV

Mustafa Hakan Kara , Aysegul Dundar , Uğur Güdükbay This is my paper · ORCID
classification cs.CV
keywords faceforgerydetectiondeepfaketripletlearningdomain-adversarialtraininggradientreversallayerdomaingeneralizationparameter-efficientfine-tuningSiamesenetwork
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

Trident is a face-forgery detection method built on a simple idea: train a network to tell real from fake by showing it two frames of the same person that have the same authenticity and one frame of the same person, at the same moment, with the opposite authenticity. Because identity and scene are held fixed, the only consistent cue the network can use is the manipulation itself, so the learned embedding is about forgery rather than about who the person is. Trident adds a domain-adversarial head that tries to guess the forgery type from the embedding, with gradients reversed so the embedding becomes blind to forgery-specific signatures, and it cuts gradient flow from the real/fake classifier back into the embedding. The paper reports that the CNN version reaches an AUC of 0.9793 on FaceForensics++ (HQ), ahead of earlier methods, and that the ViT version beats previous detectors on several cross-dataset benchmarks including Celeb-DF, DFD, DFDC, and UADFV. If these results hold, Trident offers a route toward deepfake detectors that keep working when a new manipulation method appears.

What carries the argument

The load-bearing object is the controlled triplet. For each identity, real frames are split into R1/R2 and fake frames into F1/F2; the network is trained on (R1, R2, F1) and (F1, F2, R1) and their swaps, so the anchor and positive are the same authenticity and the anchor and negative are temporally aligned with opposite authenticity. This is paired with a Gradient Reversal Layer, which reverses the gradient coming from a forgery-type classifier, and with a detached classification head that stops the real/fake loss from backpropagating into the embedding. The ViT variant adds BitFit, a parameter-efficient fine-tuning scheme that updates only bias terms. Together these pieces push the embedding to be discriminative of authenticity but invariant to identity, scene, and forgery type.

What would settle it

Train Trident and its strongest comparators from scratch on FaceForensics++ HQ under a single published protocol that fixes frame sampling, face cropping, and aggregation, then test on DFD and DFDC; if Trident's cross-dataset AUC advantage over the reproduced baselines falls below the paper's reported margins (for example DFD 0.8740 vs 0.8163 for Xception), the generalization claim is not supported. A simpler check: remove the detached head, and if the cross-dataset AUC does not drop, the detachment mechanism is not carrying the claimed load.

Watch

Extended reading notes

Core claim

The paper's central claim is that controlled triplet learning is what makes a forgery detector generalize. Instead of the usual triplet of two faces from different people, Trident builds triplets from one identity only: the anchor and positive share both identity and authenticity, while the negative shares identity and timestamp but flips authenticity, and the anchor/negative are temporally aligned. This forces the embedding to encode authenticity while suppressing identity and scene. Around this embedding, Trident trains a forgery-type discriminator through a gradient reversal layer so that the embedding cannot be used to tell which manipulation produced a fake, and it blocks gradients from the binary real/fake head so classifier-specific shortcuts do not distort the embedding. The authors report that the CNN variant achieves 0.9793 AUC on FF++ (HQ) versus 0.9705 for the previous best method, and the ViT variant with bias-only fine-tuning outperforms previous methods on Celeb-v1, Celeb-v2, DFD, DFDC, and UADFV. Ablations show triplet learning, the adversarial discriminator, and the detached head each raise AUC, with the largest jump from adding the adversarial components.

Load-bearing premise

The load-bearing premise is that every method compared in the experiments, including the reproduced baselines, ran under exactly the same frame sampling, face cropping, and video aggregation rules; if those details were not identical, the reported AUC gaps between Trident and the other detectors could shrink or disappear.

Editorial extensions

If this is right

  • Training on two forgery types (Deepfakes and NeuralTextures) and testing on all four FF++ types, the full Trident configuration reaches 0.9646 AUC, up from 0.7363 for the baseline, showing the components transfer to unseen manipulation families.
  • On cross-dataset tests from FF++ to DFD, Trident's DFD AUC is 0.8740 versus 0.8163 for Xception, so the method carries over to a different dataset and distribution.
  • The ViT variant's dominance on cross-dataset benchmarks implies bias-only fine-tuning of a large pretrained transformer is a useful recipe for forgery detection generalization.
  • The CNN and ViT variants are complementary: CNN is stronger on the in-domain benchmark while ViT generalizes better, so backbone choice should depend on deployment setting.
  • Each component contributes: triplet learning alone raises baseline AUC from 0.9497 to 0.9613 on all forgery types, and adding adversarial training and the detached head pushes it to 0.9793.

Reading between the lines

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

  • Because the controlled triplet needs identity-matched real and fake frames from the same timestamp, the method assumes paired training videos; for settings with only unpaired or single-image forgeries, one could synthesize pseudo-pairs with self-blending or augmentation and expect similar gains, but that extension is not tested in the paper.
  • The observed CNN/ViT split suggests a two-backbone ensemble could combine in-domain accuracy with cross-dataset robustness, a direction the paper does not explore.
  • The detachment of the classifier head is a simple regularization with no extra cost; applying the same detachment to other contrastive or metric-based detectors may improve their generalization independently of triplet formation.
  • The reported DFDCP and DFDC gains for the CNN variant imply the method may be especially sensitive to diverse, low-quality forgeries; testing on video-level aggregation rather than frame-level AUC would clarify deployment value.
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.

Referee Report

4 major / 5 minor

Summary. Trident proposes a face forgery detection framework that combines Siamese triplet learning, a domain-adversarial forgery discriminator connected through a gradient reversal layer, and a detached binary classification head. Triplets are curated so that anchor and positive share identity and authenticity while anchor and negative share identity and timestamp but differ in authenticity, with the aim of disentangling manipulation cues from identity and scene. The paper reports two backbone variants, EfficientNet-B4 and a ViT, the latter fine-tuned with BitFit, and evaluates on FF++ for intra-dataset performance and on Celeb-DF, DFD, DFDCP, DFDC, and UADFV for cross-dataset generalization, all under the DeepfakeBench protocol for most baselines. The main claims are that the CNN variant achieves state-of-the-art FF++ (HQ) AUC and that the ViT variant achieves superior cross-dataset generalization, outperforming existing methods on several benchmarks. Ablations and t-SNE visualizations are used to support the contribution of each component.

Significance. If the results are reproducible, Trident would be a useful strong baseline for generalizable deepfake detection: the controlled triplet formation and adversarial forgery discriminator are sensible design choices, and the paper evaluates on a broad set of external benchmarks with a standardized evaluation harness for most baselines. The paper is not circular: the central claim is an empirical generalization result, and the method is a training scheme rather than a fitted prediction. The ablation suite, the use of DeepfakeBench, and the stated intention to release code are positive features. However, the current evidence is not yet sufficient to support the headline cross-dataset claims, because the comparison against the strongest baselines (Effort and LSDA) is made under an undocumented reproduction protocol, the ViT backbone is inconsistently identified, and the reported differences are small, with no error bars or significance testing.

major comments (4)
  1. [Section IV.D.1 and Tables I, III, IV] The backbone used for the ViT variant is internally inconsistent. Section IV.D.1 states that the architectures are EfficientNet-B4 and MARLIN [74] (ViT), whereas the method section, Table I, Table III, and Table IV identify the transformer backbone as CLIP ViT-L/14, and the discussion attributes the ViT results to BitFit applied to CLIP. This is load-bearing because the cross-dataset superiority claim for Trident(ViT) in Table IV depends on which pretrained model was actually adapted. The authors must state the exact backbone, correct the text or the tables, and if MARLIN was used, the experiments must be redone with CLIP ViT-L/14 or the claims must be re-scoped to MARLIN.
  2. [Table IV footnote and Section V.A] The central cross-dataset claim is not yet established because the two strongest baselines, LSDA and Effort, are reproduced outside the DeepfakeBench harness. The footnote says these were reproduced with official implementations because weights are unavailable in DeepfakeBench, while the text claims that all metrics are obtained through DeepfakeBench's controlled evaluation setting. The paper does not report the frame sampling, face cropping, preprocessing, or video-level aggregation used in those reproductions. The ViT variant's margins over Effort are very small (e.g., +0.0007 AUC on UADFV and +0.0074 on Celeb-v1), so protocol differences can flip the comparison. The authors must provide the exact reproduction protocol, ideally run the baselines inside the same harness or release their checkpoints and evaluation scripts so that the comparison can be verified.
  3. [Table VI and Section VI(b)] The ablation evidence for the detached classification head is mixed and is presented as conclusive. In the limited-forgery training setting (Deepfakes and NeuralTextures), TL+GRL+DH achieves an AUC of 0.9646, which is slightly lower than TL+GRL's 0.9652, even though the discussion states that the detached head enhances detection performance. Table V and Table VII do show improvements from adding DH, but the discrepancy in Table VI is not addressed. The authors should report multiple seeds with means and standard deviations, and either explain why DH helps in two settings but not in the third or temper the claim about the detached head's contribution.
  4. [Eq. (3) and Table III] The total loss in Eq. (3) is L_total = L_BCE + alpha * L_triplet + beta * L_forgery, but the hyperparameter table reports values only for the triplet margin, the GRL lambda, and what appears to be beta (labeled 'Forgery loss weight'). No value is given for alpha, the triplet loss weight, so the training objective is not fully specified and the experiments cannot be reproduced without guessing this hyperparameter. The authors should report alpha explicitly, and ideally a short sensitivity analysis for alpha and beta.
minor comments (5)
  1. [Section IV.C] Scenario 3 is described as a cross-dataset evaluation on DFD only, but Table IV reports results on Celeb-v1, Celeb-v2, DFD, DFDCP, DFDC, and UADFV; the text should be updated to describe all cross-dataset benchmarks.
  2. [Section V.B and Figures 2-3] The text says the t-SNE plots show 'five distinct clusters corresponding to the five forgery types present in the FF++ dataset,' but FF++ contains four manipulation types plus the real class; the wording should be corrected.
  3. [Figure 3 caption] The figure caption lists panels (a) TL, (b) TL+GRL, and (c) TL+GRL+DH, but the main text refers to the third panel as TL+GRL+DH and the second as TL+GRL; the figure and caption should be aligned.
  4. [Section IV.D.1] The 'NetworkTree structure' is mentioned as the implementation of the proposed architecture, but it is never described; either give a concise explanation or remove the term.
  5. [Table IV] The per-method rows for Trident(CNN) and Trident(ViT) report the FF++ (HQ) overall AUC as well as per-manipulation AUCs; it would be helpful to state explicitly how the overall FF++ AUC is computed (e.g., pooling all frames of all manipulation types).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Trident's contributions are architectural and empirical, validated on external benchmarks with no self-citation chain or fitted-parameter-as-prediction reduction.

full rationale

The paper's central claim is empirical: a training objective L_total = L_BCE + α·L_triplet + β·L_forgery is trained on FF++ (HQ) and evaluated on held-out datasets, with ablations isolating the contributions of triplet learning, the gradient reversal layer, adversarial loss, and detached classification head. No equation in the paper defines a predicted quantity in terms of a quantity fitted to the same data; the triplet margin, GRL scaling, and loss weights are standard hyperparameters, not fitted constants renamed as predictions. The method builds on external prior work (FaceNet triplet loss, Ganin and Lempitsky's GRL, BitFit) rather than on the authors' own prior results, and the paper contains no self-citations by Kara, Dundar, or Güdükbay. The only notable weakness is that the LSDA and Effort baselines in Table IV are marked as reproduced with official implementations because weights are unavailable on DeepfakeBench, and the reproduction details (frame sampling, cropping, aggregation) are not fully specified. That is a reproducibility and fairness concern that could affect the magnitude of reported gains, but it is not circularity: the comparison is against external methods, and Trident's own numbers come from a model trained and evaluated under the stated protocol. Accordingly, the derivation chain is self-contained and the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. The free parameters are standard hyperparameters of the loss and training, but the omission of α and β from the text is a notable gap. The axioms are domain assumptions about the data and the transferability of the learned invariance.

free parameters (4)
  • α (triplet loss weight) = not reported
    Defined in Eq. 3 as the weight of L_triplet but never specified in the text or tables.
  • β (forgery discriminator loss weight) = 1.0 (CNN), 0.5 (ViT)
    Reported in Table III as 'Forgery loss weight'; chosen without sensitivity analysis.
  • Triplet margin m = 1.0
    Set in Table III for both backbones; no ablation or justification.
  • GRL scaling λ = 1.0
    Set in Table III for both backbones; no ablation or sensitivity study.
assumptions (3)
  • domain assumption Temporal alignment of real and fake frames in FF++ ensures anchor and negative share the same timestamp.
    The triplet construction relies on FF++ providing paired real/fake frames from the same source video; this is assumed but not verified for all identities.
  • domain assumption Forgery type is a meaningful domain for adversarial training, and removing type information from embeddings improves generalization to unseen types.
    The method assumes that making embeddings invariant to the five training forgery types transfers to novel types; no theoretical or empirical proof beyond the reported results.
  • domain assumption BitFit preserves CLIP's pretrained generalization while adapting to forgery detection.
    The paper assumes bias-only fine-tuning avoids catastrophic forgetting; this is based on prior work, not demonstrated here beyond the main results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trident: Detecting Face Forgeries with Adversarial Triplet Learning." pith.science (2026). https://pith.science/paper/AHO63HO2

@misc{pith2026250623189,
  author       = {Pith},
  title        = {Pith review of: Trident: Detecting Face Forgeries with Adversarial Triplet Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHO63HO2}},
  note         = {Machine review of arXiv:2506.23189}
}
read the original abstract

As face forgeries generated by deep neural networks become increasingly sophisticated, detecting face manipulations in digital media has posed a significant challenge, underscoring the importance of maintaining digital media integrity and combating visual disinformation. Current detection models, predominantly based on supervised training with domain-specific data, often falter against forgeries generated by unencountered techniques. In response to this challenge, we introduce \textit{Trident}, a face forgery detection framework that employs triplet learning with a Siamese network architecture for enhanced adaptability across diverse forgery methods. \textit{Trident} is trained on curated triplets to isolate nuanced differences of forgeries, capturing fine-grained features that distinguish pristine samples from manipulated ones while controlling for other variables. To further enhance generalizability, we incorporate domain-adversarial training with a forgery discriminator. This adversarial component guides our embedding model towards forgery-agnostic representations, improving its robustness to unseen manipulations. In addition, we prevent gradient flow from the classifier head to the embedding model, avoiding overfitting induced by artifacts peculiar to certain forgeries. Comprehensive evaluations across multiple benchmarks and ablation studies demonstrate the effectiveness of our framework. We will release our code in a GitHub repository.

Figures

Figures reproduced from arXiv: 2506.23189 by the authors.

Figure 1
Figure 1. Overview of our high-level architecture. We leverage a Siamese network with a triplet learning setup to generate [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. t-SNE embeddings for models without TL. GRL+DH [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. t-SNE embeddings for TL-based models. TL promotes [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

87 extracted references · 72 canonical work pages

  1. [74]

    Marlin: Masked autoencoder for facial video representation learning,

    Z. Cai, S. Ghosh, K. Stefanov, A. Dhall, J. Cai, H. Rezatofighi, R. Haffari, and M. Hayat, “Marlin: Masked autoencoder for facial video representation learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’23, 2023, pp. 1493–1504. 6

    work page 2023

  2. [1]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing Systems , ser. NIPS ’14, 2014, pp. 2672–2680. 1

    2014

  3. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems , ser. NIPS ’20, vol. 33, 2020, pp. 6840–6851. 1

    2020

  4. [3]

    CNN- generated images are surprisingly easy to spot... for now,

    S.-Y . Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, “CNN- generated images are surprisingly easy to spot... for now,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, ser. CVPR ’20, 2020, pp. 8695–8704. 1, 7

    2020

  5. [4]

    FaceNet: A unified embedding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “FaceNet: A unified embedding for face recognition and clustering,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , ser. CVPR ’15, 2015, pp. 815–823. 1, 3

    2015

  6. [5]

    Unsupervised domain adaptation by backpropagation,

    Y . Ganin and V . Lempitsky, “Unsupervised domain adaptation by backpropagation,” in Proceedings of the 32nd International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, F. Bach and D. Blei, Eds., vol. 37. Lille, France: PMLR, 07–09 Jul 2015, pp. 1180–1189. [Online]. Available: https://proceedings.mlr.press/v37/ganin15.htm...

    2015

  7. [6]

    BitFit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,

    E. Ben Zaken, Y . Goldberg, and S. Ravfogel, “BitFit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , S. Muresan, P. Nakov, and A. Villavicencio, Eds. Dublin, Ireland: Association for Computational Li...

    2022

  8. [7]

    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, ser. NIPS ’23, vol. 36, 2023, pp. 4534–

    2023

Show all 87 references
  1. [8]

    Xception: Deep learning with depthwise separable convolu- tions,

    F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, ser. CVPR ’17, 2017, pp. 1251–1258. 2, 7

    2017

  2. [9]

    EfficientNet: Rethinking model scaling for convolu- tional neural networks,

    M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolu- tional neural networks,” in Proceedings of the International Conference on Machine Learning , ser. ICML ’19, 2019, pp. 6105–6114. 2, 4, 6, 7

    2019

  3. [10]

    FaceForensics++: Learning to detect manipulated facial images,

    A. R ¨ossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Niess- ner, “FaceForensics++: Learning to detect manipulated facial images,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, ser. ICCV ’19, 2019, pp. 1–11. 2, 5, 6, 7

    2019

  4. [11]

    Face X-ray for more general face forgery detection,

    L. Li, J. Bao, T. Zhang, H. Yang, D. Chen, F. Wen, and B. Guo, “Face X-ray for more general face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’20, 2020, pp. 5001–5010. 2, 7

    2020

  5. [12]

    What makes fake images detectable? understanding properties that generalize,

    L. Chai, D. Bau, S.-N. Lim, and P. Isola, “What makes fake images detectable? understanding properties that generalize,” in Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVI . Berlin, Heidelberg: Springer-Verlag, 202...

    2020 doi

  6. [13]

    Artifacts-disentangled adversarial learning for deepfake detection,

    X. Li, R. Ni, P. Yang, Z. Fu, and Y . Zhao, “Artifacts-disentangled adversarial learning for deepfake detection,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 33, no. 4, pp. 1658– 1670, 2023. 2

    2023

  7. [14]

    LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,

    D. Nguyen, N. Mejri, I. P. Singh, P. Kuleshova, M. Astrid, A. Kacem, E. Ghorbel, and D. Aouada, “LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2024

  8. [15]

    Masked relation learning for deepfake detection,

    Z. Yang, J. Liang, Y . Xu, X.-Y . Zhang, and R. He, “Masked relation learning for deepfake detection,” IEEE Transactions on Information Forensics and Security, vol. 18, pp. 1696–1708, 2023. 2

    2023

  9. [16]

    Shallowfake and deepfake image manipulation localization using noise and RGB-based dual branch method,

    D. Dagar and D. K. Vishwakarma, “Shallowfake and deepfake image manipulation localization using noise and RGB-based dual branch method,” Signal, Image and Video Processing , vol. 18, pp. 7065–7077,

  10. [17]

    Aw-msa: Adaptively weighted multi-scale attentional features for deepfake detection,

    A. Yadav and D. K. Vishwakarma, “Aw-msa: Adaptively weighted multi-scale attentional features for deepfake detection,” Engineering Applications of Artificial Intelligence , vol. 127, p. 107443, 01

  11. [18]

    Lgdf-net: Local and global feature-based dual-branch fusion networks for deepfake detection,

    M. Long, Z. Liu, L.-B. Zhang, and F. Peng, “Lgdf-net: Local and global feature-based dual-branch fusion networks for deepfake detection,”IEEE Transactions on Circuits and Systems for Video Technology , vol. 35, no. 6, pp. 5489–5500, 2025. 2

    2025

  12. [19]

    Exposing deep fakes using inconsistent head poses,

    X. Yang, Y . Li, and S. Lyu, “Exposing deep fakes using inconsistent head poses,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing , ser. ICASSP ’19, 2019, pp. 8261–8265. 2, 6

    2019

  13. [20]

    Lips don’t lie: A generalisable and robust approach to face forgery detection,

    A. Haliassos, K. V ougioukas, S. Petridis, and M. Pantic, “Lips don’t lie: A generalisable and robust approach to face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’21, 2021, pp. 5039–5049. 2 KARA ET AL.: D...

    2021

  14. [21]

    Spatiotemporal inconsistency learning for deepfake video detection,

    Z. Gu, Y . Chen, T. Yao, S. Ding, J. Li, F. Huang, and L. Ma, “Spatiotemporal inconsistency learning for deepfake video detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’21, 2021, pp. 10 343–10 352. 2

    2021

  15. [22]

    Deepfake video detection through optical flow based CNN,

    I. Amerini, R. Caldelli, and F. Picchioni, “Deepfake video detection through optical flow based CNN,” in Proceedings of the IEEE Inter- national Conference on Computer Vision Workshops , ser. ICCVW ’19, 2019, pp. 1205–1207. 2

    2019

  16. [23]

    Exploiting inconsistencies in stylegan latent space for deepfake detection,

    I. Choi, Y . Kim, and S. Hwang, “Exploiting inconsistencies in stylegan latent space for deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’24,

  17. [24]

    V oice–Face homogeneity tells deepfake,

    H. Cheng, Y . Guo, T. Wang, Q. Li, X. Chang, and L. Nie, “V oice–Face homogeneity tells deepfake,” ACM Transactions on Multimedia Com- puting, Communications and Applications , vol. 20, no. 3, 2024. 2

    2024

  18. [25]

    Augmented multi-scale spatiotemporal inconsistency magnifier for generalized deep- fake detection,

    Y . Yu, X. Zhao, R. Ni, S. Yang, Y . Zhao, and A. C. Kot, “Augmented multi-scale spatiotemporal inconsistency magnifier for generalized deep- fake detection,” IEEE Transactions on Multimedia , vol. 25, pp. 8487– 8498, 2023. 2

    2023

  19. [26]

    Detection of deepfake videos using long-distance attention,

    W. Lu, L. Liu, B. Zhang, J. Luo, X. Zhao, Y . Zhou, and J. Huang, “Detection of deepfake videos using long-distance attention,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 7, pp. 9366–9379, 2024. 2

    2024

  20. [27]

    Tall: Thumbnail layout for deepfake video detection,

    Y . Xu, J. Liang, G. Jia, Z. Yang, Y . Zhang, and R. He, “Tall: Thumbnail layout for deepfake video detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2023, pp. 22 658–22 668. 2

    2023

  21. [28]

    Deepfake detection using robust spatial and temporal features from facial landmarks,

    M. Li, B. Liu, Y . Hu, L. Zhang, and S. Wang, “Deepfake detection using robust spatial and temporal features from facial landmarks,” in Proc. Int. Workshop on Biometrics and Forensics (IWBF) , 2021, pp. 1–6. 2

    2021

  22. [29]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues,

    Y . Qian, G. Yin, L. Sheng, Z. Chen, and J. Shao, “Thinking in frequency: Face forgery detection by mining frequency-aware clues,” in Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XII . Berlin, Heidelberg: Springer-Ve...

    2020 doi

  23. [30]

    Frequency-aware discrim- inative feature learning supervised by single-center loss for face forgery detection,

    J. Li, H. Xie, J. Li, Z. Wang, and Y . Zhang, “Frequency-aware discrim- inative feature learning supervised by single-center loss for face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’21, 2021, pp. 6458–6467. 2

    2021

  24. [31]

    Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain,

    H. Liu, X. Li, W. Zhou, Y . Chen, Y . He, H. Xue, W. Zhang, and N. Yu, “Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’21, 2021, pp. 772–781. 2, 7

    2021

  25. [32]

    Leveraging frequency analysis for deep fake image recogni- tion,

    J. Frank, T. Eisenhofer, L. Sch ¨onherr, A. Fischer, D. Kolossa, and T. Holz, “Leveraging frequency analysis for deep fake image recogni- tion,” in Proceedings of the 37th International Conference on Machine Learning, ser. ICML ’20. JMLR.org, 2020, pp. 3247–3258, article no

    2020

  26. [33]

    Rethinking the up- sampling operations in CNN-based generative network for generalizable deepfake detection,

    C. Tan, Y . Zhao, S. Wei, G. Gu, P. Liu, and Y . Wei, “Rethinking the up- sampling operations in CNN-based generative network for generalizable deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’24, June 2024,...

    2024

  27. [34]

    Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning,

    ——, “Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, ser. AAAI ’24, vol. 38, no. 5, 2024, pp. 5052–5060. 2

    2024

  28. [35]

    A timely survey on vision transformer for deepfake detection,

    Z. Wang, Z. Cheng, J. Xiong, X. Xu, T. Li, B. Veeravalli, and X. Yang, “A timely survey on vision transformer for deepfake detection,” 2024. [Online]. Available: https://arxiv.org/abs/2405.08463 2

    2024 arXiv

  29. [36]

    ISTVT: Interpretable spatial-temporal video transformer for deepfake detection,

    C. Zhao et al., “ISTVT: Interpretable spatial-temporal video transformer for deepfake detection,” IEEE Transactions on Information Forensics and Security, vol. 18, pp. 1335–1348, 2023. 2

    2023

  30. [37]

    M2TR: Multi-modal multi-scale transformers for deepfake detection,

    J. Wang, Z. Wu, J. Chen, and Y . Jiang, “M2TR: Multi-modal multi-scale transformers for deepfake detection,” CoRR, vol. abs/2104.09770, 2021. [Online]. Available: https://arxiv.org/abs/2104.09770 2

    2021 arXiv

  31. [38]

    Protecting celebrities from deepfake with identity consistency transformer,

    X. Dong, J. Bao, D. Chen, T. Zhang, W. Zhang, N. Yu, D. Chen, F. Wen, and B. Guo, “Protecting celebrities from deepfake with identity consistency transformer,” 2022. [Online]. Available: https://arxiv.org/abs/2203.01318 2

    2022 arXiv

  32. [39]

    UIA-ViT: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection,

    W. Zhuang, Q. Chu, Z. Tan, Q. Liu, H. Yuan, C. Miao, Z. Luo, and N. Yu, “UIA-ViT: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection,” 2022. [Online]. Available: https://arxiv.org/abs/2210.12752 2

    2022 arXiv

  33. [40]

    Video transformer for deepfake detection with incremental learning,

    S. A. Khan and H. Dai, “Video transformer for deepfake detection with incremental learning,” CoRR, vol. abs/2108.05307, 2021. [Online]. Available: https://arxiv.org/abs/2108.05307 2

    2021 arXiv

  34. [41]

    Deepfake video detection using generative convolutional vision transformer,

    D. Wodajo, S. Atnafu, and Z. Akhtar, “Deepfake video detection using generative convolutional vision transformer,” 2023. [Online]. Available: https://arxiv.org/abs/2307.07036 2

    2023 arXiv

  35. [42]

    Forgery-aware adaptive learning with vision transformer for generalized face forgery detection,

    A. Luo, R. Cai, C. Kong, Y . Ju, X. Kang, J. Huang, and A. C. Kot, “Forgery-aware adaptive learning with vision transformer for generalized face forgery detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4116–4129, 2025. 2

    2025

  36. [43]

    Tex-ViT: A generalizable, ro- bust, texture-based dual-branch cross-attention deepfake detector,

    D. Dagar and D. K. Vishwakarma, “Tex-ViT: A generalizable, ro- bust, texture-based dual-branch cross-attention deepfake detector,” arXiv preprint arXiv:2408.16892, 2024. 2

    2024

  37. [44]

    Attributing fake images to GANs: learning and analyzing gan fingerprints,

    N. Yu, L. S. Davis, and M. Fritz, “Attributing fake images to GANs: learning and analyzing gan fingerprints,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, ser. ICCV ’19, 2019, pp. 7556–7566. 2

    2019

  38. [45]

    CORE: Consis- tent representation learning for face forgery detection,

    Y . Ni, D. Meng, C. Yu, C. Quan, D. Ren, and Y . Zhao, “CORE: Consis- tent representation learning for face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, ser. CVPRW ’22, 2022, pp. 12–21. 2, 7

    2022

  39. [46]

    Learning self- consistency for deepfake detection,

    T. Zhao, X. Xu, M. Xu, H. Ding, Y . Xiong, and W. Xia, “Learning self- consistency for deepfake detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , ser. ICCV ’21, 2021, pp. 15 023–15 033. 2

    2021

  40. [47]

    Implicit identity leakage: The stumbling block to improving deepfake detection generalization,

    X. Huang, F. Xue, B. Fan, L. Zhong, Y . Fu, and Q. Tian, “Implicit identity leakage: The stumbling block to improving deepfake detection generalization,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , ser. CVPR ’23, 2023, pp. 3994–

    2023

  41. [48]

    UCF: Uncovering common features for generalizable deepfake detection,

    Z. Yan, Y . Zhang, Y . Fan, and B. Wu, “UCF: Uncovering common features for generalizable deepfake detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, ser. ICCV ’23, 2023, pp. 22 355–22 366. 2, 7

    2023

  42. [49]

    Towards universal fake image detec- tors that generalize across generative models,

    U. Ojha, Y . Li, and Y . J. Lee, “Towards universal fake image detec- tors that generalize across generative models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’23, 2023, pp. 24 480–24 489. 2

    2023

  43. [50]

    Detecting deepfakes with self-blended images,

    K. Shiohara and T. Yamasaki, “Detecting deepfakes with self-blended images,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’22, 2022, pp. 18 720–18 729. 2

    2022

  44. [51]

    End-to- end reconstruction-classification learning for face forgery detection,

    J. Cao, C. Ma, T. Yao, S. Chen, S. Ding, and X. Yang, “End-to- end reconstruction-classification learning for face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’22, 2022, pp. 4113–4122. 2, 7

    2022

  45. [52]

    Boosting deepfake detection generalizability via expansive learning and confi- dence judgement,

    K. Zhang, Z. Hou, Z. Hua, Y . Zheng, and L. Y . Zhang, “Boosting deepfake detection generalizability via expansive learning and confi- dence judgement,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 1, pp. 953–966, 2025. 2

    2025

  46. [53]

    PV ASS–MDD: Predictive visual–audio alignment self-supervision for multimodal deep- fake detection,

    Y . Yu, X. Liu, R. Ni, S. Yang, Y . Zhao, and A. C. Kot, “PV ASS–MDD: Predictive visual–audio alignment self-supervision for multimodal deep- fake detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 1, pp. 17–30, 2024. 2

    2024

  47. [54]

    MCL: Multimodal contrastive learning for deepfake detection,

    X. Liu, Y . Yu, X. Li, and Y . Zhao, “MCL: Multimodal contrastive learning for deepfake detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2803–2813, 2024. 2

    2024

  48. [56]

    Supervised contrastive learning for generalizable and explainable deepfakes detection,

    Y . Xu, K. Raja, and M. Pedersen, “Supervised contrastive learning for generalizable and explainable deepfakes detection,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, ser. W ACVW ’22, 2022, pp. 379–389. 2

    2022

  49. [57]

    Hierarchical contrastive inconsistency learning for deepfake video detection,

    Z. Gu, T. Yao, Y . Chen, S. Ding, and L. Ma, “Hierarchical contrastive inconsistency learning for deepfake video detection,” in Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XII . Berlin, Heidelberg: Springer-Ve...

    2022 doi

  50. [58]

    Depth map guided triplet network for deepfake face detection,

    B. Liang, Z. Wang, B. Huang, Q. Zou, Q. Wang, and J. Liang, “Depth map guided triplet network for deepfake face detection,” Neural Networks, vol. 159, pp. 34–42, 2023. 2

    2023

  51. [59]

    Detecting deepfakes with metric learning,

    A. Kumar, A. Bhavsar, and R. Verma, “Detecting deepfakes with metric learning,” in Proceedings of the 8th International Workshop on Biometrics and Forensics, ser. IWBF ’20, 2020, pp. 1–6. 2 KARA ET AL.: DETECTING FACE FORGERIES WITH DOMAIN-ADVERSARIAL TRIPLET LEARNING 11

    2020

  52. [60]

    EANN: Event adversarial neural networks for multi-modal fake news detection,

    Y . Wang, F. Ma, Z. Jin, Y . Yuan, G. Xun, K. Jha, L. Su, and J. Gao, “EANN: Event adversarial neural networks for multi-modal fake news detection,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , ser. KDD ’18, 2018, pp. 849–857. 2

    2018

  53. [61]

    Distance metric learning for large margin nearest neighbor classification,

    K. Q. Weinberger and L. K. Saul, “Distance metric learning for large margin nearest neighbor classification,” Journal of Machine Learning Research, vol. 10, pp. 207–244, Jun. 2009. 3

    2009

  54. [62]

    Learning a similarity metric discriminatively, with application to face verification,

    S. Chopra, R. Hadsell, and Y . LeCun, “Learning a similarity metric discriminatively, with application to face verification,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, ser. CVPR ’05, vol. 1. IEEE, 2005, pp. 539–546. 3

    2005

  55. [63]

    Deep metric learning via lifted structured feature embedding,

    H. Oh Song, Y . Xiang, S. Jegelka, and S. Savarese, “Deep metric learning via lifted structured feature embedding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , ser. CVPR ’16, 2016, pp. 4004–4012. 3

    2016

  56. [64]

    Signature verification using a “Siamese

    J. Bromley, I. Guyon, Y . LeCun, E. S ¨ackinger, and R. Shah, “Signature verification using a “Siamese” time delay neural network,” in Advances in Neural Information Processing Systems , ser. NIPS ’93, vol. 6, 1993. 3

    1993

  57. [65]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in Proceedings of the International Conference on Machine Learning , ser. ICML ’21, 202...

    2021

  58. [66]

    DeepFake detection dataset,

    Google and Jigsaw, “DeepFake detection dataset,” 2019, https://ai. googleblog.com/2019/09/contributing-data-to-deepfake-detection.html, Accessed: 2025-02-16. 5, 6

    2019

  59. [67]

    Deepfakes,

    Deepfakes Community, “Deepfakes,” https://github.com/deepfakes/ faceswap, 2018, accessed: 2018-10-29. 6

    2018

  60. [68]

    Face2Face: Real-time face capture and reenactment of RGB videos,

    J. Thies, M. Zollh ¨ofer, M. Stamminger, C. Theobalt, and M. Nießner, “Face2Face: Real-time face capture and reenactment of RGB videos,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, ser. CVPR ’16, 2016, pp. 2387–2395. 6

    2016

  61. [69]

    FaceSwap,

    M. Kowalski, “FaceSwap,” https://github.com/MarekKowalski/ FaceSwap/, 2018, accessed: 2018-10-29. 6

    2018

  62. [70]

    Deferred neural rendering: Image synthesis using neural textures,

    J. Thies, M. Zollh ¨ofer, and M. Nießner, “Deferred neural rendering: Image synthesis using neural textures,” ACM Transactions on Graphics, vol. 38, no. 4, 2019. 6

    2019

  63. [71]

    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 , ser. CVPR ’20, 2020, pp. 3207–3216. 6

    2020

  64. [72]

    The deepfake detection challenge (DFDC) dataset,

    B. Dolhansky, J. Bitton, B. Pflaum, J. Lu, R. Howes, M. Wang, and C. C. Ferrer, “The deepfake detection challenge (DFDC) dataset,” arXiv preprint arXiv:2006.07397, 2020. 6

    2006 arXiv

  65. [73]

    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,” arXiv preprint arXiv:1910.08854, 2019. 6

    1910 arXiv

  66. [75]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proceedings of the 3rd International Conference for Learning Representations , ser. ICLR ’15, 2015. [Online]. Available: https://arxiv.org/abs/1412.6980 6

    2015 arXiv

  67. [76]

    MesoNet: A compact facial video forgery detection network,

    D. Afchar, V . Nozick, J. Yamagishi, and I. Echizen, “MesoNet: A compact facial video forgery detection network,” in Proceedings of the IEEE International Workshop on Information Forensics and Security , ser. WIFS ’18. IEEE, 2018, pp. 1–7. 7

    2018

  68. [77]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, ser. CVPR ’16, 2016, pp. 770–778. 7

    2016

  69. [78]

    Generalizing face forgery de- tection with high-frequency features,

    Y . Luo, Y . Zhang, J. Yan, and W. Liu, “Generalizing face forgery de- tection with high-frequency features,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’21, 2021, pp. 16 317–16 326. 7

    2021

  70. [79]

    Capsule-Forensics: Using capsule networks to detect forged images and videos,

    H. H. Nguyen, J. Yamagishi, and I. Echizen, “Capsule-Forensics: Using capsule networks to detect forged images and videos,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, ser. ICASSP ’19, 2019, pp. 2307–2311. 7

    2019

  71. [80]

    Dynamic routing between capsules,

    S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” in Advances in Neural Information Processing Systems , ser. NIPS ’17, vol. 30, 2017, pp. 3859–3869. 7

    2017

  72. [81]

    Exposing deepfake videos by detecting face warping artifacts,

    Y . Li and S. Lyu, “Exposing deepfake videos by detecting face warping artifacts,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , ser. CVPRW ’19, June 2019, pp. 46–52. 7

    2019

  73. [82]

    Deep high-resolution representation learning for visual recognition,

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang et al. , “Deep high-resolution representation learning for visual recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 10, pp. 3349–3364, 2020. 7

    2020

  74. [83]

    On the detection of digital face manipulation,

    H. Dang, F. Liu, J. Stehouwer, X. Liu, and A. K. Jain, “On the detection of digital face manipulation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ser. CVPR ’20, 2020, pp. 5781–5790. 7

    2020

  75. [84]

    Transcending forgery specificity with latent space augmentation for generalizable deepfake detection,

    Z. Yan, Y . Luo, S. Lyu, Q. Liu, and B. Wu, “Transcending forgery specificity with latent space augmentation for generalizable deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ser. CVPR ’24, 2024, pp. 8984–8994. 7

    2024

  76. [85]

    Orthogonal subspace decomposition for generalizable AI-generated image detection,

    Z. Yan, J. Wang, P. Jin, K.-Y . Zhang, C. Liu, S. Chen, T. Yao, S. Ding, B. Wu, and L. Yuan, “Orthogonal subspace decomposition for generalizable AI-generated image detection,” in Proceedings of the International Conference on Machine Learning , ser. ICML ’25, 2025. 7

    2025

  77. [86]

    Visualizing data using t-SNE,

    L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” Journal of Machine Learning Research , vol. 9, pp. 2579–2605, Nov

  78. [2021]

    Available: https://arxiv.org/abs/2104.11507 2

    [Online]. Available: https://arxiv.org/abs/2104.11507 2

    arXiv

  79. [2024]

    Available: https://linkinghub.elsevier.com/retrieve/pii/ S0952197623016275 2

    [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/ S0952197623016275 2

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.