Pith. sign in

REVIEW 5 major objections 7 minor 59 references

Contrastive Desensitization Learning for Cross Domain Face Forgery Detection

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Training on genuine faces alone can detect forged faces on unseen domains, with a 0.14% false positive rate at 85% true positive rate.

desk verdict Solid empirical package for low-FPR cross-domain deepfake detection, but the central theoretical guarantee is unproved and the proof is not a minor gap. read the letter →

arxiv 2505.20675 v1 pith:DMN74IBO submitted 2025-05-27 cs.CV

Lingyu Qiu , Ke Jiang , Xiaoyang Tan This is my paper · ORCID
classification cs.CV
keywords faceforgerydetectiondeepfakedomaingeneralizationcontrastivelearningdesensitizationdomain-invariantrepresentationfalsepositiveratereconstruction
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

This paper claims that cross-domain face forgery detection can be made robust to unseen forgery methods by learning a representation of genuine faces from real images only, without ever seeing a fake image during the representation-learning stage. The proposal is a conditional-independence criterion ($Z \perp D \mid I$) and a desensitization objective that minimizes the KL divergence between the representation of a face dressed in two different domain styles; the authors argue that its optimum satisfies the criterion. The practical method, a Contrastive Desensitization Network, implements this objective by mixing the feature statistics of two real faces and reconstructing the original image, and the paper's Theorem 2 asserts that this reconstruction loss is an upper bound on the KL objective. If the claim is right, a system can achieve a low false positive rate on unseen domains without needing examples of future forgery techniques; on FaceForensics++(c23) the reported false positive rate is 0.14% at 85% true positive rate, compared with 0.84% for Xception and 0.53% for RECCE.

What carries the argument

The load-bearing object is the desensitization objective together with the reconstruction loss that upper-bounds it. The encoder maps an image to latent statistics $D=(\mu(z),\sigma(z))$ and normalized content $I=(z-\mu(z))/\sigma(z)$ (Eq. (3)); the objective in Eq. (4) is $D_{KL}[P_\theta(z|i_A,d_B)\,\|\,P_\theta(z|i_A,d_A)]$, the gap between representations of the same intrinsic content under two domain styles. The domain transformation (Eq. (8)) is adaptive-instance-normalization-style mixing, $z_{\mathrm{out}} = \sigma_B (z_A-\mu_A)/\sigma_A + \mu_B$, which wraps sample A's content in sample B's style. The denoising reconstruction loss (Eq. (12)), $\|\phi(z_{\mathrm{out}})-x_A\|_2^2$, then forces the decoder to remove the injected style. Theorem 2 is the bridge: it asserts that maximizing the reconstruction likelihood is equivalent to minimizing an upper bound of the KL objective, so the practical loss carries the theoretical guarantee.

What would settle it

On a trained CDN, compute $\Delta = \log P_\theta(z|i_A,d_A) - \log P_\phi(i_A,d_A|z)$ over held-out genuine faces; any sample with $\Delta < 0$ violates the density-dominance assumption and removes the proof that the reconstruction loss upper-bounds the KL objective of Eq. (4).

Watch

Extended reading notes

Core claim

The central claim is that desensitization is a sufficient inductive bias for cross-domain deepfake detection. Decomposing a real face into intrinsic content $I$ and domain-specific statistics $D$, and then training an encoder so that the distribution of its representation $Z$ does not change when $D$ is swapped, yields a representation that is invariant to both known and unseen domains (Definition 1). The paper proves (Theorem 1) that the optimum of the KL objective in Eq. (4) achieves $Z \perp D \mid I$, and proves (Theorem 2) that maximizing the likelihood of the denoising reconstruction in Eq. (12) is equivalent to minimizing an upper bound of that KL objective, under a density-dominance assumption. The implemented CDN mixes the statistics of a source feature with those of a target feature to create the perturbed sample, then reconstructs the source image through a decoder; only genuine images participate in this stage. The learned encoder representation is passed to a downstream classifier, and the experiments report lower false positive rates and improved AUC on FF++, Celeb-DF, WildDeepfake, and DFDC relative to the compared methods.

Load-bearing premise

The load-bearing premise is that in the trained model, the encoder's probability density in latent space is never smaller than the decoder's probability density in image space for the same sample, and that the decoder's density is always below 1 so a key logarithm is negative; continuous probability densities are not guaranteed to behave this way, so the proof's bound collapses if the premise fails.

Editorial extensions

If this is right

  • On FaceForensics++(c23) at an 85% true positive rate, the method's false positive rate is 0.14%, versus 0.84% for Xception and 0.53% for RECCE, showing the low-alarm behavior that makes the system usable.
  • Because the representation stage uses only genuine images, generalization to new forgery methods no longer depends on collecting fake examples of those methods; fake data enter only through an optional domain-boundary constraint and the downstream classifier.
  • In cross-dataset tests with a model trained on FF++(c40) and tested on Celeb-DF, WildDeepfake, and DFDC, the reported AUC/EER improve over the compared methods, indicating the representation transfers to entirely unseen data sources.
  • In cross-manipulation tests within FF++, the AUC advantage persists when the manipulation type at test time differs from training; for example, a Face2Face-trained model reaches 85.86 AUC on DeepFakes test data, versus 78.07 for the best compared method.
  • The extra machinery adds modest cost: 1.14 G FLOPs versus RECCE's 2.27 G with similar parameter count, so the low-false-alarm behavior does not require a heavier detector.

Reading between the lines

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

  • The recipe of splitting content from domain statistics, mixing statistics across genuine samples, and reconstructing could transfer to other open-ended spoof-detection settings, such as audio deepfakes or document forgery, where genuine samples are abundant but attack types are unbounded; the paper evaluates faces only, so this is an extrapolation.
  • Theorem 2's bound is testable: on a trained model, estimate $\Delta = \log P_\theta(z|i_A,d_A) - \log P_\phi(i_A,d_A|z)$ from Eq. (30); if $\Delta$ is negative for a non-negligible fraction of held-out genuine faces, the theoretical guarantee is not doing the work and the low FPR would be an empirical property of the trained architecture.
  • The domain-boundary constraint ablation shows a direct trade-off between false negatives and false positives; replacing that contrastive term with a one-class constraint on the real-face manifold would test whether the whole detector can remain entirely fake-free while keeping both rates low.
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

5 major / 7 minor

Summary. The paper proposes Contrastive Desensitization Network (CDN), a cross-domain face forgery detection method that learns a representation from genuine face images by mixing low-level feature statistics across domains (Eq. 8) and then reconstructing the source image through desensitization, intrinsic, and domain-alignment losses (Eqs. 9-12), together with a domain-boundary contrastive loss (Eq. 18). The central claim is that minimizing the desensitization objective Eq. (4) guarantees a domain-invariant representation satisfying Z ⊥⊥ D | I (Definition 1, Theorem 1), and that the implemented reconstruction loss Eq. (12) is an upper-bound surrogate for Eq. (4) (Theorem 2). Experiments on FF++, Celeb-DF, WildDeepfake, and DFDC report improved AUC/ACC and lower false-positive rates at fixed true-positive rates, with ablations and visualizations.

Significance. If the theoretical guarantee were valid, the work would offer a principled genuine-face-only representation stage for zero-shot cross-domain deepfake detection. The empirical results are extensive and directionally consistent: Tables 1-4 place CDN at or near the top on most intra-dataset, cross-dataset, cross-manipulation, and multi-source settings, and Figure 5 reports a large FPR reduction relative to RECCE and Xception. However, the central theoretical claim is not established: the proof of Theorem 2 contains invalid substitutions between P_phi and P_theta and a false assertion about log-densities, and Theorem 1 is largely definitional. Without the theory, the paper is an empirical study with several reporting gaps, including garbled tables and no released code or data. The significance as a theoretical contribution is therefore currently unsupported.

major comments (5)
  1. [Theoretical Justification, Theorem 2] The proof of Theorem 2 does not establish the claimed bound. In Eq. (27) the quantity log(1/P_phi(z|i_A,d_A)) is introduced, but the KL divergence in Eq. (29) requires log(P_theta(z|i_A,d_B)/P_theta(z|i_A,d_A)). The added term in Eq. (28) is E log P_phi(z|i_A,d_B), which is not the term appearing in the KL between the P_theta conditionals. The assumed dominance P_theta(z|i_A,d_A) >= P_phi(i_A,d_A|z) does not justify Eq. (27), both because it does not relate P_phi(i_A,d_A|z) to P_phi(z|i_A,d_A) and because it gives no control on P_theta(z|i_A,d_A). Consequently the chain (25)-(29) is invalid, and Theorem 2, which is the central bridge between the implemented loss Eq. (12) and the desensitization objective Eq. (4), is unsupported.
  2. [Theoretical Justification, Eq. (30)] The assumed dominance condition is not meaningful as stated because P_theta(z|i_A,d_A) is a density on the latent space while P_phi(i_A,d_A|z) is a density on the data space; comparing their pointwise values depends on a choice of reference measure. The Gaussian heuristic in Eq. (30) compares log-densities of different dimensionalities and concludes the variance term is positive whenever sigma_phi^2 > sigma_theta^2, which ignores the normalization constants and allows the bias term to be negative. The separate assertion used for inequality (b), that P_phi(z|i_A,d_B) < 1 so its logarithm is negative, is false for continuous densities: a Gaussian N(0, 0.01) has density about 3.99 at the origin. Removing either assertion collapses the proof, so the theoretical guarantee advertised in the abstract and in the section 'Theoretical Justification for the Proposed Method' is not justified.
  3. [Learning to Desensitize; Eq. (31)] Even if the inequalities in the proof were repaired, the result would be a one-sided upper bound with uncontrolled slack; it would not show that minimizing Eq. (12) minimizes Eq. (4). The actual training loss in Eq. (31) is L_cls + lambda_1 L_d + lambda_2 L_i + lambda_3 L_s and contains no KL term, no bound, and no domain-boundary term L_b from Eq. (18), despite the paper's extensive ablations of the Domain Boundary Constraint. The connection between the theoretical objective and the optimized loss is therefore not established.
  4. [Theorem 1] Theorem 1 is largely definitional. The objective Eq. (4) directly penalizes the KL divergence between P_theta(z|i_A,d_B) and P_theta(z|i_A,d_A); at the optimum these two conditionals coincide for all d_A,d_B, which is essentially the conditional independence Z ⊥⊥ D | I stated in Definition 1. The 'redundancy elimination' step in the proof assumes the conditional independence it is meant to derive: the assertion that P_theta*(z|i_A,d_A,d_B)=P_theta*(z|i_A,d_A) is exactly the claim that, given i_A, the additional domain variable d_B is irrelevant. This circularity undermines the claim of a formal guarantee.
  5. [Experiments, Tables 2 and 11] The experimental reporting is not verifiable in its current form. Table 2 prints CDN results as '70.73±0.634.66±1.871.26±2.135.20±4.370.21±2.735.08±5.8', which is unparsable and does not allow comparison with baselines. Table 11 contains the identical hyperparameter row (lambda_1=0.1, lambda_2=0.1, lambda_3=0.1) twice with different reported AUC/ACC (92.50/84.82 and 91.46/84.12). The real-world evaluation in Table 5 and the DBC ablations in Tables 7-9 lack dataset construction details, standard deviations, and, for Table 5, any description of the manipulation protocols. No code or data are released, so the consistency of the empirical tables cannot be checked.
minor comments (7)
  1. [Abstract and Introduction] There are several language errors, including 'an Contrastive', 'with regard to the its robustness', and 'Constrastive' in the introduction; these should be corrected.
  2. [Implementation Details] References to 'Section a', 'Appendix a', and 'eq.a' point to non-existent sections; the appendix is not included, and these cross-references should either be resolved or removed.
  3. [Intra-dataset Evaluation, Figure 4] The paragraph discussing ROC curves appears to mislabel the subfigures: the text refers to 'Figure 4(a)' and 'Figure 4(b)' in a way that does not match the described intra-evaluation and cross-manipulation results.
  4. [Eq. (8)] The domain transformation is described only for z_A, but the symmetric use for z_B is not defined; the formula should state how the two samples are treated and what happens when the batch contains more than two domains.
  5. [Eq. (18)] The function Dis(x,y) is used in Eq. (18) before it is defined in Eq. (19); the definition should be introduced before first use.
  6. [Table 1] Several entries in Table 1 are missing separators (e.g., '99.29±0.491.54±0.7'), making the table difficult to read; the formatting should be corrected.
  7. [Conclusion] The assumption that domain features have clear boundaries in latent space and that intrinsic features overlap sufficiently is stated only in the conclusion; because this assumption is load-bearing for the desensitization mechanism, it should be stated and discussed earlier, ideally with a test or diagnostic.

Circularity Check

2 steps flagged · score 6.0 of 10

The formal guarantee is partly definitional and self-referential: Theorem 1 restates Eq. (4) as conditional independence, and Theorem 2's bridge from Eq. (12) to Eq. (4) relies on a density-dominance assumption justified by the very reconstruction loss it is meant to validate.

  1. self definitional [Section 'Cross Domain Desensitization', Definition 1, Eq. (4), Theorem 1 and its proof]
    "Definition 1.(Domain-invariant representation) We define a representation Z ... as domain-invariant if it is conditionally independent of the domain-specific information D, i.e., Z⊥⊥D|I ... Theorem 1. The optimal solution of minimizing Eq. (4) guarantees the representation is conditionally independent of the domain information, i.e., Z⊥⊥D|I."

    Eq. (4) is exactly DKL(Pθ(z|iA,dB) || Pθ(z|iA,dA)). Its optimum forces Pθ(z|iA,dB)=Pθ(z|iA,dA), which is literally the conditional-independence condition in Definition 1. The theorem therefore restates the objective as the target property by construction; the promised 'guarantee' is built into the loss, not derived independently.

  2. other [Section 'Theoretical Justification for the Proposed Method', Theorem 2, Eqs. (26)-(30) and the justification after Eq. (30)]
    "Theorem 2. Under the assumption that the probability density of the hidden space is commonly larger than the original sample space, i.e., ∀z,θ, Pθ(z|iA,dA)≥Pϕ(iA,dA|z). Then maximizing the denoising reconstruction term in Eq.(20) ... is equivalent to minimizing the upper bound of the following objective, DKL(Pθ(z|iA,dB) || Pθ(z|iA,dA)) ... In the training of VAE, the reconstruction loss like Eq.(12) tends to minimize ∥x−µϕ∥²/2σ²ϕ, so enhancing the σ²ϕ, hence ... we can assert that the assumption holds."

    The claimed reduction of the implemented loss Eq. (12) to the desensitization objective Eq. (4) is conditional on the dominance assumption Pθ(z|iA,dA)≥Pϕ(iA,dA|z). That assumption is defended by appealing to the behavior of the very reconstruction loss Eq. (12): 'the reconstruction loss like Eq.(12) tends to minimize ... so enhancing the σ²ϕ ... we can assert that the assumption holds.' Thus the conclusion of Theorem 2 is used to justify the premise from which it is derived. In addition, the final step silently substitutes Pφ densities in Eq. (28) for the Pθ densities in the KL of Eq. (29), so the linkage is a renaming plus a self-supported premise rather than an independent derivation.

full rationale

The empirical part of the paper is largely self-contained: it is evaluated on external benchmarks (FF++, Celeb-DF, WildDeepfake, DFDC) against published baselines, and the one prior work from the same group (Qiu et al. 2024) is used only as a comparison baseline, not as a load-bearing justification. No self-citation chain forces the empirical results. The circularity is in the theoretical claim. Eq. (4) is defined as the KL divergence between two conditionals whose equality is, by Definition 1, precisely the target property, so Theorem 1 is a tautological restatement of the objective. The key bridge, Theorem 2, is supposed to show that the practical reconstruction loss Eq. (12) minimizes the theoretical objective Eq. (4), but its proof depends on a density-dominance assumption that the paper justifies on the basis of Eq. (12) itself, making the derivation self-referential; the proof also swaps Pθ and Pφ when identifying Eq. (28) with the KL in Eq. (29). The advertised formal guarantee therefore reduces, at key points, to the construction and to an assumption supported by the very loss being justified, even though the experimental comparisons retain independent content.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on a feature-statistics decomposition of domain information, an isotropic Gaussian likelihood, and two density assumptions that the proof does not establish. Hyperparameters lambda_1, lambda_2, lambda_3, and alpha are tuned on validation performance. No new physical or external entity is introduced.

free parameters (4)
  • lambda_1 = 0.1
    Weight for the denoising reconstruction loss L_d in Eq. (31), chosen by cross-validation; Table 11 explores values around 0.1.
  • lambda_2 = 0.1
    Weight for the intrinsic alignment loss L_i in Eq. (31), chosen by cross-validation.
  • lambda_3 = 0.1
    Weight for the domain alignment loss L_s in Eq. (31), chosen by cross-validation.
  • alpha = 0.3
    Proportion of samples undergoing domain transformation during training, tuned to 0.3 as best on FS-to-DF in Table 12.
assumptions (5)
  • domain assumption Every face image can be decomposed into intrinsic features I and domain-specific information D via channel-wise mean and standard deviation, as in Eq. (3).
    This decomposition is the foundation of the desensitization objective; if domain information is not captured by feature statistics, the method has no mechanism to remove it.
  • domain assumption The likelihood P_phi(i_A, d_A | z) = P_phi(x_A | z) is isotropic Gaussian with covariance lambda I.
    This assumption converts the reconstruction objective Eq. (12) into a negative log-likelihood in Eqs. (15) through (17).
  • ad hoc to paper P_theta(z | i_A, d_A) >= P_phi(i_A, d_A | z) for all z and theta.
    This density dominance assumption is needed for inequality (a) in Theorem 2 and is justified only by a dimensional heuristic in the text, not by a proof or measurement.
  • ad hoc to paper P_phi(z | i_A, d_B) < 1 so that E log P_phi(z | i_A, d_B) < 0.
    Used for inequality (b) in Theorem 2; probability densities can exceed 1, so the log can be positive, making the step invalid for general continuous densities.
  • ad hoc to paper In the latent space, different domains are far enough apart and intrinsic features overlap sufficiently.
    Stated in the conclusion as a limitation of CDN: if domain noise is too complex, intrinsic and domain features may overlap too much and discrimination fails.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Desensitization Learning for Cross Domain Face Forgery Detection." pith.science (2026). https://pith.science/paper/DMN74IBO

@misc{pith2026250520675,
  author       = {Pith},
  title        = {Pith review of: Contrastive Desensitization Learning for Cross Domain Face Forgery Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMN74IBO}},
  note         = {Machine review of arXiv:2505.20675}
}
read the original abstract

In this paper, we propose a new cross-domain face forgery detection method that is insensitive to different and possibly unseen forgery methods while ensuring an acceptable low false positive rate. Although existing face forgery detection methods are applicable to multiple domains to some degree, they often come with a high false positive rate, which can greatly disrupt the usability of the system. To address this issue, we propose an Contrastive Desensitization Network (CDN) based on a robust desensitization algorithm, which captures the essential domain characteristics through learning them from domain transformation over pairs of genuine face images. One advantage of CDN lies in that the learnt face representation is theoretical justified with regard to the its robustness against the domain changes. Extensive experiments over large-scale benchmark datasets demonstrate that our method achieves a much lower false alarm rate with improved detection accuracy compared to several state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2505.20675 by the authors.

Figure 1
Figure 1. The diagram (left) of the domain shift problem, shows that the diver￾gence between the source and target data distribution would potentially lead to a high false alarm rate. We also perform reconstruction (right) over cross-domain samples, and observe that the distribution of real face images reconstructed from the target dataset (WildDeepfake) differs significantly from those from the source domain (Celeb-DF) while… view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed CDN for face forgery detection. To learn domain-invariant representations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Diagram of the domain-invariant objective. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: False Alarm Rate(FPR) (↓) when cross-dataset testing among dataset FF++, Celeb-DF(CDF), WildDeepfake(WDF). The left two are trained on FF++, and the right two are on CDF. Cross-Domain Evaluation Cross-Dataset Evaluation. To explore the generalization of our method on u…
Figure 6
Figure 6. Figure 6: The representation space differences between our CDN and RECCE methods are illustrated through reconstruction and residual images on the FaceForen [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Ablation studies in terms of ACC (%), AUC (%) and FPR (%) including [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 10
Figure 10. Figure 10: The GradCAMSelvaraju et al. (2017) visualizations of our proposed [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 48 canonical work pages

  1. [1]

    , author Wang, X

    author Bai, N. , author Wang, X. , author Han, R. , author Hou, J. , author Wang, Q. , author Pang, S. , year 2024 . title Towards generalizable face forgery detection via mitigating spurious correlation . journal Neural Networks , pages 106909

    work page 2024

  2. [2]

    , author Kumar, N

    author Bitouk, D. , author Kumar, N. , author Dhillon, S. , author Belhumeur, P. , author Nayar, S.K. , year 2008 . title Face swapping: automatically replacing faces in photographs . journal ACM Transactions on Graphics , pages 1–8 http://dx.doi.org/10.1145/1360612.1360638, :10.1145/1360612.1360638

    arXiv 2008

  3. [3]

    , author Ma, C

    author Cao, J. , author Ma, C. , author Yao, T. , author Chen, S. , author Ding, S. , author Yang, X. , year 2022 . title End-to-end reconstruction-classification learning for face forgery detection , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. pages 4113--4122

    work page 2022

  4. [4]

    , author Tan, S

    author Chen, B. , author Tan, S. , year 2021 . title Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection . journal Security and Communication Networks volume 2021 , pages 1--8

    work page 2021

  5. [5]

    , author Zhang, Y

    author Chen, L. , author Zhang, Y. , author Song, Y. , author Liu, L. , author Wang, J. , year 2022 . title Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection , in: booktitle 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . http://dx.doi.org/10.1109/cvpr52688.2022.01815, :10.110...

    arXiv 2022

  6. [6]

    , author Xu, Z

    author Chen, M. , author Xu, Z. , author Weinberger, K. , author Sha, F. , year 2012 . title Marginalized denoising autoencoders for domain adaptation . journal arXiv preprint arXiv:1206.4683

    arXiv 2012

  7. [7]

    , year 2017

    author Chollet, F. , year 2017 . title Xception: Deep learning with depthwise separable convolutions , in: booktitle Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2017

  8. [8]

    , author Howes, R

    author Dolhansky, B. , author Howes, R. , author Pflaum, B. , author Baram, N. , author Ferrer, C.C. , year 2019 . title The deepfake detection challenge (dfdc) preview dataset . arXiv:1910.08854 http://arxiv.org/abs/1910.08854

    arXiv 2019

Show all 59 references
  1. [9]

    , author Zhen, C

    author Guo, Y. , author Zhen, C. , author Yan, P. , year 2023 . title Controllable guide-space for generalizable face forgery detection , in: booktitle Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pp. pages 20818--20827

    2023

  2. [10]

    , author Mira, R

    author Haliassos, A. , author Mira, R. , author Petridis, S. , author Pantic, M. , year 2022 . title Leveraging real talking faces via self-supervision for robust forgery detection , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2022

  3. [11]

    , author Fan, H

    author He, K. , author Fan, H. , author Wu, Y. , author Xie, S. , author Girshick, R. , year 2020 . title Momentum contrast for unsupervised visual representation learning , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. p...

    2020

  4. [12]

    , author Peng, C

    author He, Q. , author Peng, C. , author Liu, D. , author Wang, N. , author Gao, X. , year 2024 . title Gazeforensics: Deepfake detection via gaze-guided spatial inconsistency learning . journal Neural Networks volume 180 , pages 106636

    2024

  5. [13]

    , author Yu, N

    author He, Y. , author Yu, N. , author Keuper, M. , author Fritz, M. , year 2021 . title Beyond the spectrum: Detecting deepfakes via re-synthesis , in: booktitle Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence . http://dx.doi.org/10.2496...

    2021 doi

  6. [14]

    , author Roweis, S.T

    author Hinton, G.E. , author Roweis, S.T. , year 2002 . title Stochastic neighbor embedding , in: editor Becker, S. , editor Thrun, S. , editor Obermayer, K. (Eds.), booktitle Advances in Neural Information Processing Systems 15 [Neural Information Processing Systems, NIPS 200...

    2002

  7. [15]

    , year 2013

    author Hoffman, J. , year 2013 . title Efficient learning of domain-invariant image representations . journal Computer Science

    2013

  8. [16]

    , author Belongie, S

    author Huang, X. , author Belongie, S. , year 2017 . title Arbitrary style transfer in real-time with adaptive instance normalization , in: booktitle Proceedings of the IEEE international conference on computer vision , pp. pages 1501--1510

    2017

  9. [17]

    , author Wang, L

    author Ke, J. , author Wang, L. , year 2023 . title Df-udetector: An effective method towards robust deepfake detection via feature restoration . journal Neural Networks volume 160 , pages 216--226

    2023

  10. [18]

    , author Ba, J

    author Kingma, D.P. , author Ba, J. , year 2014 . title Adam: A method for stochastic optimization . journal arXiv preprint arXiv:1412.6980

    2014 arXiv

  11. [19]

    , author Yang, Y

    author Li, D. , author Yang, Y. , author Song, Y.Z. , author Hospedales, T. , year 2018 . title Learning to generalize: Meta-learning for domain generalization , in: booktitle Proceedings of the AAAI conference on artificial intelligence

    2018

  12. [20]

    , author Wang, S

    author Li, H. , author Wang, S. , author Wan, R. , author Kot, A.C. , year 2020 a. title Gmfad: Towards generalized visual recognition via multi-layer feature alignment and disentanglement . journal IEEE Transactions on Pattern Analysis and Machine Intelligence volume PP

    2020

  13. [21]

    , author Li, Y

    author Li, J. , author Li, Y. , author Tan, J. , author Liu, C. , year 2024 . title It takes two: Dual branch augmentation module for domain generalization . journal Neural Networks volume 172 , pages 106094

    2024

  14. [22]

    , author Xie, H

    author Li, J. , author Xie, H. , author Li, J. , author Wang, Z. , author Zhang, Y. , year 2021 . title Frequency-aware discriminative feature learning supervised by single-center loss for face forgery detection , in: booktitle 2021 IEEE/CVF Conference on Computer Vision and P...

    2021

  15. [23]

    , author Wang, N

    author Li, Y. , author Wang, N. , author Liu, J. , author Hou, X. , year 2017 . title Demystifying neural style transfer . journal arXiv preprint arXiv:1701.01036

    2017 arXiv

  16. [24]

    , author Yang, X

    author Li, Y. , author Yang, X. , author Sun, P. , author Qi, H. , author Lyu, S. , year 2020 b. title Celeb-df: A large-scale challenging dataset for deepfake forensics , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2020

  17. [25]

    , author Wang, Z

    author Liang, B. , author Wang, Z. , author Huang, B. , author Zou, Q. , author Wang, Q. , author Liang, J. , year 2023 . title Depth map guided triplet network for deepfake face detection . journal Neural Networks volume 159 , pages 34--42

    2023

  18. [26]

    , author Li, X

    author Liu, H. , author Li, X. , author Zhou, W. , author Chen, Y. , author He, Y. , author Xue, H. , author Zhang, W. , author Yu, N. , year 2021 . title Spatial-phase shallow learning: rethinking face forgery detection in frequency domain , in: booktitle Proceedings of the I...

    2021

  19. [27]

    , author Zhang, Y

    author Luo, Y. , author Zhang, Y. , author Yan, J. , author Liu, W. , year 2021 . title Generalizing face forgery detection with high-frequency features , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 16317--16326

    2021

  20. [28]

    , author Li, Y

    author Lv, Q. , author Li, Y. , author Dong, J. , author Chen, S. , author Yu, H. , author Zhou, H. , author Zhang, S. , year 2024 . title Domainforensics: Exposing face forgery across domains via bi-directional adaptation . journal IEEE Transactions on Information Forensics a...

    2024

  21. [29]

    , year 2020

    author Lyu, S. , year 2020 . title Deepfake detection: Current challenges and next steps , in: booktitle 2020 IEEE international conference on multimedia & expo workshops (ICMEW) , organization IEEE . pp. pages 1--6

    2020

  22. [30]

    title Faceswap

    author MarekKowalski , year 2018 . title Faceswap . howpublished https://github.com/MarekKowalski/FaceSwap . note Accessed:

    2018

  23. [31]

    , author Sun, F

    author Peng, C. , author Sun, F. , author Liu, D. , author Wang, N. , author Gao, X. , year 2024 . title Local artifacts amplification for deepfakes augmentation . journal Neural Networks volume 180 , pages 106692

    2024

  24. [32]

    , author Yin, G

    author Qian, Y. , author Yin, G. , author Sheng, L. , author Chen, Z. , author Shao, J. , year 2020 . title Thinking in frequency: Face forgery detection by mining frequency-aware clues , in: booktitle European conference on computer vision , organization Springer . pp. pages 86--103

    2020

  25. [33]

    , author Jiang, K

    author Qiu, L. , author Jiang, K. , author Liu, S. , author Tan, X. , year 2024 . title Multi-level distributional discrepancy enhancement for cross domain face forgery detection , in: booktitle Chinese Conference on Pattern Recognition and Computer Vision (PRCV) , organizatio...

    2024

  26. [34]

    , author Cozzolino, D

    author Rossler, A. , author Cozzolino, D. , author Verdoliva, L. , author Riess, C. , author Thies, J. , author Niessner, M. , year 2019 . title Faceforensics++: Learning to detect manipulated facial images , in: booktitle Proceedings of the IEEE/CVF International Conference o...

    2019

  27. [35]

    , author Cogswell, M

    author Selvaraju, R.R. , author Cogswell, M. , author Das, A. , author Vedantam, R. , author Parikh, D. , author Batra, D. , year 2017 . title Grad-cam: Visual explanations from deep networks via gradient-based localization , in: booktitle Proceedings of the IEEE international...

    2017

  28. [36]

    , author Zhang, J

    author Shi, L. , author Zhang, J. , author Shan, S. , year 2023 a. title Real face foundation representation learning for generalized deepfake detection . journal arXiv preprint arXiv:2303.08439

    2023 arXiv

  29. [37]

    , author Chen, H

    author Shi, Z. , author Chen, H. , author Chen, L. , author Zhang, D. , year 2023 b. title Discrepancy-guided reconstruction learning for image forgery detection , in: booktitle Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence , pp. pa...

    2023

  30. [38]

    , author Yamasaki, T

    author Shiohara, K. , author Yamasaki, T. , year 2022 . title Detecting deepfakes with self-blended images , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. pages 18720--18729

    2022

  31. [39]

    , author Zhong, J

    author Shuai, C. , author Zhong, J. , author Wu, S. , author Lin, F. , author Wang, Z. , author Ba, Z. , author Liu, Z. , author Cavallaro, L. , author Ren, K. , year 2023 . title Locate and verify: A two-stream network for improved deepfake detection , in: booktitle Proceedin...

    2023

  32. [40]

    , author Fang, Z

    author Song, L. , author Fang, Z. , author Li, X. , author Dong, X. , author Jin, Z. , author Chen, Y. , author Lyu, S. , year 2022 . title Adaptive face forgery detection in cross domain , in: booktitle European conference on computer vision , organization Springer . pp. page...

    2022

  33. [41]

    , author Liu, H

    author Sun, K. , author Liu, H. , author Yao, T. , author Sun, X. , author Chen, S. , author Ding, S. , author Ji, R. , year 2022 . title An information theoretic approach for attention-driven face forgery detection , in: booktitle European Conference on Computer Vision , orga...

    2022

  34. [42]

    , author Liu, H

    author Sun, K. , author Liu, H. , author Ye, Q. , author Gao, Y. , author Liu, J. , author Shao, L. , author Ji, R. , year 2021 . title Domain general face forgery detection by learning to weight , in: booktitle Proceedings of the AAAI conference on artificial intelligence , p...

    2021

  35. [43]

    , author Seitz, S.M

    author Suwajanakorn, S. , author Seitz, S.M. , author Kemelmacher-Shlizerman, I. , year 2017 . title Synthesizing obama: learning lip sync from audio . journal ACM Transactions on Graphics , pages 1–13 http://dx.doi.org/10.1145/3072959.3073640, :10.1145/3072959.3073640

    2017

  36. [44]

    , author Zollhofer, M

    author Thies, J. , author Zollhofer, M. , author Stamminger, M. , author Theobalt, C. , author Nie ner, M. , year 2016 . title Face2face: Real-time face capture and reenactment of rgb videos , in: booktitle Proceedings of the IEEE conference on computer vision and pattern reco...

    2016

  37. [45]

    , author Zollhöfer, M

    author Thies, J. , author Zollhöfer, M. , author Nießner, M. , year 2019 . title Deferred neural rendering: Image synthesis using neural textures . journal arXiv: Computer Vision and Pattern Recognition,arXiv: Computer Vision and Pattern Recognition

    2019

  38. [46]

    , author Vera-Rodriguez, R

    author Tolosana, R. , author Vera-Rodriguez, R. , author Fierrez, J. , author Morales, A. , author Ortega-Garcia, J. , year 2020 . title Deepfakes and beyond: A survey of face manipulation and fake detection . journal Information Fusion volume 64 , pages 131--148

    2020

  39. [47]

    title Deepfakes

    author torzdf , year 2018 . title Deepfakes . howpublished https://github.com/deepfakes/faceswap . note Accessed:

    2018

  40. [48]

    , author Lebedev, V

    author Ulyanov, D. , author Lebedev, V. , author Vedaldi, A. , author Lempitsky, V. , year 2016 . title Texture networks: Feed-forward synthesis of textures and stylized images . journal arXiv preprint arXiv:1603.03417

    2016 arXiv

  41. [49]

    , author Vedaldi, A

    author Ulyanov, D. , author Vedaldi, A. , author Lempitsky, V. , year 2017 . title Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis , in: booktitle Proceedings of the IEEE conference on computer vision and pattern re...

    2017

  42. [50]

    , author Deng, W

    author Wang, C. , author Deng, W. , year 2021 . title Representative forgery mining for fake face detection , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 14923--14932

    2021

  43. [51]

    , author Bao, J

    author Wang, Z. , author Bao, J. , author Zhou, W. , author Wang, W. , author Li, H. , year 2023 . title Altfreezing for more general video face forgery detection , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 4129--4138

    2023

  44. [52]

    , author Zhang, Y

    author Wu, W. , author Zhang, Y. , author Li, C. , author Qian, C. , author Loy, C.C. , year 2018 . title Reenactgan: Learning to reenact faces via boundary transfer , in: booktitle Proceedings of the European conference on computer vision (ECCV) , pp. pages 603--619

    2018

  45. [53]

    , author Ni, R

    author Yu, Y. , author Ni, R. , author Yang, S. , author Zhao, Y. , author Kot, A.C. , year 2023 . title Narrowing domain gaps with bridging samples for generalized face forgery detection . journal IEEE Transactions on Multimedia , pages 1--13 :10.1109/TMM.2023.3310341

    2023

  46. [54]

    , author Tang, J

    author Zhang, D. , author Tang, J. , author Cheng, K.T. , year 2022 . title Graph reasoning transformer for image parsing , in: booktitle Proceedings of the 30th ACM International Conference on Multimedia , pp. pages 2380--2389

    2022

  47. [55]

    , author Xiao, Z

    author Zhang, D. , author Xiao, Z. , author Li, J. , author Ge, S. , year 2023 . title Self-supervised transformer with domain adaptive reconstruction for general face forgery video detection . journal arXiv preprint arXiv:2309.04795

    2023 arXiv

  48. [56]

    , author Zhou, W

    author Zhao, H. , author Zhou, W. , author Chen, D. , author Wei, T. , author Zhang, W. , author Yu, N. , year 2021 . title Multi-attentional deepfake detection , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 2185--2194

    2021

  49. [57]

    , author Yang, Y

    author Zhou, K. , author Yang, Y. , author Qiao, Y. , author Xiang, T. , year 2023 . title Mixstyle neural networks for domain generalization and adaptation . journal International Journal of Computer Vision , pages 1--15

    2023

  50. [58]

    , author Han, X

    author Zhou, P. , author Han, X. , author Morariu, V.I. , author Davis, L.S. , year 2017 . title Two-stream neural networks for tampered face detection , in: booktitle 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) . http://dx.doi.org/10.1109...

    2017 doi

  51. [59]

    , author Chang, M

    author Zi, B. , author Chang, M. , author Chen, J. , author Ma, X. , author Jiang, Y.G. , year 2020 . title Wilddeepfake: A challenging real-world dataset for deepfake detection , in: booktitle Proceedings of the 28th ACM International Conference on Multimedia . http://dx.doi....

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.