Pith. sign in

REVIEW 5 major objections 4 minor 39 references

RoGA: Towards Generalizable Deepfake Detection through Robust Gradient Alignment

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RoGA, a training objective that adds sharpness-aware perturbations to model parameters and aligns each domain's perturbed gradient with its empirical-risk gradient, is claimed to deliver state-of-the-art cross-dataset deepfake detection…

desk verdict Plausible optimizer recipe, broken empirical case: RoGA's own tables contradict its SOTA claim, but the idea and code are real. read the letter →

arxiv 2505.20653 v1 pith:34W5F4DS submitted 2025-05-27 cs.CV cs.AI

Lingyu Qiu , Ke Jiang , Xiaoyang Tan This is my paper · ORCID
classification cs.CVcs.AI
keywords deepfakedetectiondomaingeneralizationsharpness-awareminimizationgradientalignmentfaceforgerycross-datasetevaluationrobustoptimization
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 is trying to establish that cross-domain deepfake detection can be improved without adding modules, data augmentation, or domain-invariant feature extractors. The proposed training objective, Robust Gradient Alignment (RoGA), first nudges the model's weights in whatever direction most increases the loss, then requires each domain's gradient at the nudged weights to stay aligned with that domain's ordinary empirical-risk gradient, as combined in Eq. 6. The claim is that the model then lands in a minimum that is simultaneously flat and consistent across domains, so it stops overfitting to forgery-specific patterns. A sympathetic reader would care because real-world deepfake detectors typically drop sharply on forgeries from unseen generators, and RoGA is offered as a parameter-space fix for exactly that drop.

What carries the argument

The load-bearing object is the RoGA objective (Eq. 6): $$\frac{1}{K}\sum_{i=1}^{K}\big[L(\$\theta$+\epsilon_i; D_i) - \$\alpha$\langle\nabla L(\$\theta$+\epsilon_i; D_i), \nabla L(\$\theta$; D_i)\rangle\big]$$ where $\epsilon_i$ is the sharpness-aware perturbation $\rho\,\nabla L(\theta; D_i)/\|\nabla L(\theta; D_i)\|$ from Eq. 4. The first term is a robust loss that penalizes sharpness and steers the model toward flat minima; the second term is a conservative alignment term that stops each domain's perturbed gradient from deviating from its empirical-risk gradient, which is meant to prevent inter-domain gradient conflicts. The optimization decouples the two: the perturbation is estimated at the current $\theta_t$, and then $\theta$ is updated with ordinary gradient descent.

What would settle it

Reproduce the published protocol (DeepfakeBench splits, ResNet34, SGD at learning rate 0.005, $\alpha=0.0002$, $\rho=0.1$) and test on DFDC and Celeb-DF; if the reported 0.751 and 0.877 AUCs, or the claimed superiority over ConfR and SAGM, do not reproduce under matched backbones and hyperparameters, the central generalization claim fails. A second check is to measure the actual inner maximum of Eq. 3 at $\rho=0.1$ against the approximation in Eq. 4, since a large gap would mean the mechanism the paper invokes is not what the objective actually optimizes.

Watch

Extended reading notes

Core claim

RoGA's central claim is that overfitting to domain-specific artifacts can be countered in the parameter space itself. During each update, the method computes a perturbation $\hat{\epsilon}_i = \rho\,\nabla L(\theta; D_i)/\|\nabla L(\theta; D_i)\|$ for each domain, evaluates the loss at $\theta+\hat{\epsilon}_i$, and subtracts $\alpha\langle\nabla L(\theta+\hat{\epsilon}_i; D_i), \nabla L(\theta; D_i)\rangle$ so that the perturbed update does not drift away from the empirical-risk direction. The resulting objective is claimed to preserve domain-invariant features, to keep domain-specific characteristics under control, and to guide the model toward flatter minima than ERM or plain sharpness-aware minimization. Trained on FF++(c23) with a ResNet34 backbone, the paper reports cross-dataset AUCs of 0.877 on Celeb-DF, 0.858 on CDF-v2, and 0.959 on UADFV, and states that these are the best cross-domain results among the compared methods.

Load-bearing premise

The method's success rests on the unproven premise that adding controlled noise to the model's weights, then keeping each domain's noisy update pointing in the same direction as its ordinary update, genuinely pushes the model toward a flatter, more domain-neutral solution, and that the shortcut used to compute the noise (a first-order approximation) is accurate enough at the chosen noise size.

Editorial extensions

If this is right

  • If RoGA's claim holds, any existing deepfake detector can be retrained on the same data with this objective and gain cross-dataset robustness without adding parameters or inference cost.
  • Cross-manipulation generalization improves substantially, with reported AUCs of 90.08 on GID-DF and 98.08 on GID-FS when training on the other three FF++ forgeries.
  • The method is backbone-agnostic, with consistent AUC gains over baselines on ResNet34, Xception, and EfficientNetB4, and no extra parameters added.
  • The perturbation term and the alignment term are complementary: ablation results show each contributes independently, and the full RoGA objective gives the best in-domain AUC of 99.30 on the FF++(c23) DeepFakes test set.

Reading between the lines

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

  • The paper's statement that RoGA achieves the best DFDC AUC of 0.751 is not supported by its own Table I, which lists ConfR at 0.803; the cross-dataset comparison should be read with that discrepancy in mind.
  • Because RoGA is a training objective rather than an architecture change, it would likely transfer to other domain-generalization tasks beyond forgeries, such as medical imaging or autonomous driving, but that transfer is untested by the paper.
  • The closest prior optimizer, SAGM, also performs sharpness-aware gradient matching across domains, so the marginal contribution of RoGA's per-domain ERM alignment over SAGM needs a matched-backbone and matched-hyperparameter head-to-head to be isolated.
  • The Taylor-approximation justification can be tested directly: at the chosen $\rho=0.1$, one could compare the true perturbation maximizing Eq. 3 against the normalized-gradient approximation in Eq. 4 to check whether the claimed ascent direction is actually what is optimized.
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 / 4 minor

Summary. The paper proposes RoGA, a domain-generalization objective for deepfake detection that combines sharpness-aware parameter perturbations (after SAM) with a per-domain gradient-alignment term. The claimed contribution is a regularization-free training objective that preserves domain-invariant features and improves cross-dataset and cross-manipulation generalization. The authors report results on FF++(c23) as source and Celeb-DF, DFDC, DFDCP, and UADFV as targets, plus multi-source leave-one-out experiments on FF++. The central empirical claim is that RoGA achieves state-of-the-art cross-domain performance, with the code made publicly available.

Significance. The idea of aligning perturbed gradients across domains is a plausible extension of sharpness-aware minimization and could be useful if its improvements were demonstrated cleanly. The paper also provides open code, which is a strength. However, as submitted, the evidence for the central claim is internally inconsistent: Table I contradicts the headline DFDC result, Table II contradicts the claim of consistent improvements, and the hyperparameter settings and selection protocol are not coherent. Since the paper's contribution is empirical, these issues are load-bearing rather than cosmetic.

major comments (5)
  1. [Section V-B, Table I] The text states that RoGA achieves "the highest AUC of 0.877 on Celeb-DF and 0.751 on DFDC" and outperforms state-of-the-art methods, but Table I lists ConfR at 0.803 on DFDC and 0.828 on DFDCP, both above RoGA's 0.751 and 0.753. Because ConfR is discussed as related work in Section II, this is an internal contradiction in the paper's own reported numbers, not an artifact of an external benchmark.
  2. [Section V-B, Table II] The accompanying paragraph claims that Table II results "consistently outperform competitive baselines," but the table shows DisGRL with 72.8 AUC on GID-NT versus RoGA's 70.52. Since the multi-source cross-manipulation evaluation is one of the two main evaluation protocols, this contradiction undermines the claimed generality of the method.
  3. [Section V-A and Section V-C4] The training details report alpha = 0.0002 and rho = 0.1, while the hyperparameter sensitivity study states that the optimal values are alpha = 0.001 and rho = 0.1. The manuscript does not reconcile which setting produced Tables I and II. Moreover, Table IV reports UADFV AUC 95.95 under the stated setting, while Table VI reports 94.23 for the supposedly optimal alpha = 0.001, rho = 0.1, adding another unreconciled mismatch among the paper's own results.
  4. [Section V-C4, Table VI] The hyperparameters alpha and rho are tuned using cross-dataset AUC values on UADFV and CelebDF, which are the same target benchmarks used in the main generalization evaluation. Selecting hyperparameters on the target test sets makes the reported cross-domain numbers in Tables I and IV partially selected rather than independent, so the claim of generalizing to unseen domains is evaluator-circular.
  5. [Section IV and Section V-C2] SAGM [36], which is closely related to the proposed sharpness-aware gradient alignment, appears as a baseline in Table IV but is never discussed in Sections II or IV. The paper provides neither a theoretical comparison with SAGM nor an ablation that isolates the alignment term from SAGM's objective, leaving the novelty and mechanism of Eq. (6) relative to that baseline unestablished.
minor comments (4)
  1. [Section V-B] The heading "Muli-source Cross-Manipulation Evaluation" contains a typo and should be "Multi-source".
  2. [Section VI vs. Section IV-B] The conclusion describes the core idea as "gradient orthogonal alignment," but Eq. (6) uses an inner-product alignment term that does not enforce orthogonality; this terminology should be made consistent.
  3. [References] Reference [24] duplicates the DFDC preview dataset reference [22] and has garbled title and author text, and reference [28] cites Pollard's book on stochastic processes rather than the original SGD optimizer.
  4. [Abstract and Eq. (6)] The abstract says the method works "without introducing additional regularization," but Eq. (6) contains an explicit penalty term weighted by alpha; the paper should clarify why this term is not considered a regularization term.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; mild target-set hyperparameter selection and internal inconsistencies compromise the generalization claim.

  1. fitted input called prediction [Section V-C-4 (Hyperparameter Sensitivity), Table VI; Section V-A (Training Details)]
    "The hyperparameter α is set as 0.0002, while ρ = 0.1. ... The optimal values, α = 0.001 and ρ = 0.1, consistently yield superior results under both evaluations. Table VI presents the AUC(%) results across intra- and cross-dataset settings on FF++."

    The paper presents cross-dataset AUCs on UADFV and CelebDF as evidence of generalization, but the hyperparameter sensitivity analysis selects α and ρ by inspecting AUC on those same target domains (Table VI includes UADFV and CelebDF columns) and calls the chosen values optimal. If the final configuration is selected on the target benchmarks, the reported target-domain numbers are partially fitted rather than independent predictions. The contradiction between the training details (α=0.0002) and the claimed optimal (α=0.001) further obscures which configuration produced the headline results. This is a mild evaluator circularity; it does not make Eq. 6 itself circular.

full rationale

The core derivation chain is not circular. Equations (3)-(6) construct a SAM-style perturbed objective plus a gradient-alignment penalty; the alignment term is an explicit design choice, not a quantity that is later used to define its own input. The paper does not rely on a load-bearing self-citation chain: the only self-citation, reference [2], is background context about threats, not a justification of the method or a uniqueness theorem. The objective has independent content and is compared against external baselines. The main concerns are not circularity: Table I lists ConfR at 0.803 on DFDC and 0.828 on DFDCP, above RoGA's 0.751 and 0.753, contradicting the text's 'highest AUC' claim for DFDC; Table II shows DisGRL outperforming RoGA on GID-NT; and the reported α values differ between Section V-A and Section V-C-4. These are correctness and reproducibility issues, not self-referential reductions. The only mild circularity is the selection of α and ρ using target-domain test sets in the hyperparameter sensitivity study, which weakens the claim that the reported cross-dataset AUCs are purely out-of-sample predictions. Overall, the central derivation is not equivalent to its inputs, so the circularity score is low.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or architectural entities; it is an optimization scheme with two free hyperparameters, alpha and rho, which are tuned on target-domain data. The main axioms are the Taylor approximation, the epsilon decoupling, and the unproven alignment benefit.

free parameters (2)
  • alpha (balance coefficient) = 0.0002 in training details, 0.001 in Table VI (inconsistent)
    Controls the strength of the gradient alignment term in Eq. 6. Selected via sensitivity analysis using target-domain performance, effectively fitting to the benchmarks.
  • rho (perturbation radius) = 0.1
    Defines the L2 ball for the SAM-style perturbation in Eq. 4. Chosen in Table VI based on target-domain performance.
assumptions (4)
  • standard math The inner maximization in Eq. 3 can be linearized via Taylor expansion (Eq. 4), assuming small rho and differentiability.
    Standard first-order approximation used by SAM, but assumed valid at rho=0.1 across all domains without verification.
  • ad hoc to paper Treating epsilon as fixed when updating theta (decoupling) is a valid approximation of the joint objective in Eq. 6.
    The paper states this decoupling without proof; it is a pragmatic gradient estimation choice, not a derived equivalence.
  • ad hoc to paper The per-domain gradient alignment term preserves domain-invariant features and does not conflict with ERM optimization.
    The central claim of the method relies on this unproven behavioral assumption about the loss landscape.
  • domain assumption A latent domain distribution P(D) underlies the data partition and the multi-domain objective.
    Section III assumes domains are sampled from a distribution; the method uses domain labels to compute per-domain gradients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoGA: Towards Generalizable Deepfake Detection through Robust Gradient Alignment." pith.science (2026). https://pith.science/paper/34W5F4DS

@misc{pith2026250520653,
  author       = {Pith},
  title        = {Pith review of: RoGA: Towards Generalizable Deepfake Detection through Robust Gradient Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/34W5F4DS}},
  note         = {Machine review of arXiv:2505.20653}
}
read the original abstract

Recent advancements in domain generalization for deepfake detection have attracted significant attention, with previous methods often incorporating additional modules to prevent overfitting to domain-specific patterns. However, such regularization can hinder the optimization of the empirical risk minimization (ERM) objective, ultimately degrading model performance. In this paper, we propose a novel learning objective that aligns generalization gradient updates with ERM gradient updates. The key innovation is the application of perturbations to model parameters, aligning the ascending points across domains, which specifically enhances the robustness of deepfake detection models to domain shifts. This approach effectively preserves domain-invariant features while managing domain-specific characteristics, without introducing additional regularization. Experimental results on multiple challenging deepfake detection datasets demonstrate that our gradient alignment strategy outperforms state-of-the-art domain generalization techniques, confirming the efficacy of our method. The code is available at https://github.com/Lynn0925/RoGA.

Figures

Figures reproduced from arXiv: 2505.20653 by the authors.

Figure 1
Figure 1. Comparison among previous cross-domain deepfake detection based [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The left part illustrates the framework of RoGA. The right part demonstrates the basic idea of RoGA, where it aligns the gradients with different [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The GradCAM visualizations [37] comparing SRM [39] baseline and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: t-SNE [38] visualization of latent space [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 30 canonical work pages

  1. [36]

    Sharpness- aware gradient matching for domain generalization,

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang, “Sharpness- aware gradient matching for domain generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3769–3778

    work page 2023

  2. [1]

    Deepfake detection: Current challenges and next steps,

    Siwei Lyu, “Deepfake detection: Current challenges and next steps,” in 2020 IEEE international conference on multimedia & expo workshops (ICMEW). IEEE, 2020, pp. 1–6

    work page 2020

  3. [2]

    Multi-level distributional discrepancy enhancement for cross domain face forgery detection,

    Lingyu Qiu, Ke Jiang, Sinan Liu, and Xiaoyang Tan, “Multi-level distributional discrepancy enhancement for cross domain face forgery detection,” in Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 2024, pp. 508–522

    work page 2024

  4. [3]

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

    Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang, “End-to-end reconstruction-classification learning for face forgery detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 4113–4122

    work page 2022

  5. [4]

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

    HuyH. Nguyen, Junichi Yamagishi, and Isao Echizen, “Capsule- forensics: Using capsule networks to detect forged images and videos,” Cornell University - arXiv,Cornell University - arXiv , Oct 2018

    work page 2018

  6. [5]

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

    Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai 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, 2021, pp. 772–781

    work page 2021

  7. [6]

    Exploiting style latent flows for generalizing deepfake video detection,

    Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi, “Exploiting style latent flows for generalizing deepfake video detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 1133–1143

    work page 2024

  8. [7]

    mixup: Beyond empirical risk minimization,

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez- Paz, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412, 2017

    arXiv 2017

Show all 39 references
  1. [8]

    How does mixup help with robustness and generalization?,

    Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Amirata Ghorbani, and James Zou, “How does mixup help with robustness and generalization?,” arXiv preprint arXiv:2010.04819 , 2020

    2010 arXiv

  2. [9]

    Sharpness-aware minimization for efficiently improving generaliza- tion,

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur, “Sharpness-aware minimization for efficiently improving generaliza- tion,” arXiv preprint arXiv:2010.01412 , 2020

    2010 arXiv

  3. [10]

    Core: Consistent representation learning for face forgery detection,

    Yunsheng Ni, Depu Meng, Changqian Yu, Chengbin Quan, Dongchun Ren, and Youjian Zhao, “Core: Consistent representation learning for face forgery detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 12–21

    2022

  4. [11]

    Generalizing face forgery detection with high-frequency features,

    Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu, “Generalizing face forgery detection with high-frequency features,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 16317–16326

    2021

  5. [12]

    Ucf: Uncovering common features for generalizable deepfake detection,

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu, “Ucf: Uncovering common features for generalizable deepfake detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22412–22423

    2023

  6. [13]

    Adaforensics: Learning a characteristic-aware adaptive deepfake detector,

    Xiaoke Yang, Haixu Song, Xiangyu Lu, Shao-Lun Huang, and Yueqi Duan, “Adaforensics: Learning a characteristic-aware adaptive deepfake detector,” in 2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2024, pp. 1–6

    2024

  7. [14]

    Self-supervised learning of adversarial example: Towards good gener- alizations for deepfake detection,

    Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang, “Self-supervised learning of adversarial example: Towards good gener- alizations for deepfake detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 18710–18719

    2022

  8. [15]

    Detecting deepfakes with self-blended images,

    Kaede Shiohara and Toshihiko Yamasaki, “Detecting deepfakes with self-blended images,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 18720–18729

    2022

  9. [16]

    Domain general face forgery detection by learning to weight,

    Ke Sun, Hong Liu, Qixiang Ye, Yue Gao, Jianzhuang Liu, Ling Shao, and Rongrong Ji, “Domain general face forgery detection by learning to weight,” in Proceedings of the AAAI conference on artificial intelligence, 2021, vol. 35, pp. 2638–2646

    2021

  10. [17]

    Learning to generalize: Meta-learning for domain generalization,

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales, “Learning to generalize: Meta-learning for domain generalization,” in Proceedings of the AAAI conference on artificial intelligence , 2018, vol. 32

    2018

  11. [18]

    Multi-task learning for detecting and segmenting manipulated facial images and videos,

    Huy H Nguyen, Fuming Fang, Junichi Yamagishi, and Isao Echizen, “Multi-task learning for detecting and segmenting manipulated facial images and videos,” in 2019 IEEE 10th international conference on biometrics theory, applications and systems (BTAS) . IEEE, 2019, pp. 1–8

    2019

  12. [19]

    Wilddeepfake: A challenging real-world dataset for deepfake detection,

    Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang, “Wilddeepfake: A challenging real-world dataset for deepfake detection,” in Proceedings of the 28th ACM International Conference on Multimedia, Oct 2020

    2020

  13. [20]

    Confr: Conflict resolving for generaliz- able deepfake detection,

    Jin Chen, Jiahe Tian, Cai Yu, Xi Wang, Zhaoxing Li, Yesheng Chai, Jiao Dai, and Jizhong Han, “Confr: Conflict resolving for generaliz- able deepfake detection,” in 2024 IEEE International Conference on Multimedia and Expo (ICME) . IEEE, 2024, pp. 1–6

    2024

  14. [21]

    Faceforensics++: Learning to detect manipulated facial images,

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner, “Faceforensics++: Learning to detect manipulated facial images,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2019

    2019

  15. [22]

    The deepfake detection challenge (dfdc) preview dataset,

    Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cris- tian Canton Ferrer, “The deepfake detection challenge (dfdc) preview dataset,” 2019

    2019

  16. [23]

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

    Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu, “Celeb-df: A large-scale challenging dataset for deepfake forensics,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2020

    2020

  17. [24]

    The dee pfake detection challenge (dfdc) pre view dataset,

    B Dolhansky, “The dee pfake detection challenge (dfdc) pre view dataset,” arXiv preprint arXiv:1910.08854 , 2019

    1910 arXiv

  18. [25]

    Exposingaicreated fakevideos- bydetectingeyeblinking,

    Chang M Liy and LYUS InIctuOculi, “Exposingaicreated fakevideos- bydetectingeyeblinking,” in 2018IEEEInterG national Workshop on Information Forensics and Security (WIFS). IEEE , 2018

    2018

  19. [26]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770– 778

    2016

  20. [27]

    Imagenet: A large-scale hierarchical image database,

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li 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. [28]

    David Pollard, Convergence of stochastic processes , Springer Science & Business Media, 2012

    2012

  22. [29]

    Deepfakebench: a comprehensive benchmark of deepfake detection,

    Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu, “Deepfakebench: a comprehensive benchmark of deepfake detection,” in Proceedings of the 37th International Conference on Neural Information Processing Systems, 2023, pp. 4534–4565

    2023

  23. [30]

    Mesonet: a compact facial video forgery detection network,

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen, “Mesonet: a compact facial video forgery detection network,” in 2018 IEEE international workshop on information forensics and security (WIFS). IEEE, 2018, pp. 1–7

    2018

  24. [31]

    Xception: Deep learning with depthwise separable convolutions,

    Francois Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    2017

  25. [32]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    Mingxing Tan and Quoc Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” in International conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  26. [33]

    Discrepancy- guided reconstruction learning for image forgery detection,

    Zenan Shi, Haipeng Chen, Long Chen, and Dong Zhang, “Discrepancy- guided reconstruction learning for image forgery detection,” arXiv preprint arXiv:2304.13349, 2023

    2023 arXiv

  27. [34]

    Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks,

    Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi, “Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks,” in International Conference on Machine Learning. PMLR, 2021, pp. 5905–5914

    2021

  28. [35]

    Adam: A method for stochastic optimization,

    Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  29. [37]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakr- ishna Vedantam, Devi Parikh, and Dhruv Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 618–626

    2017

  30. [38]

    Stochastic neighbor em- bedding,

    Geoffrey E. Hinton and Sam T. Roweis, “Stochastic neighbor em- bedding,” in Advances in Neural Information Processing Systems 15 [Neural Information Processing Systems, NIPS 2002, December 9-14, 2002, Vancouver, British Columbia, Canada], Suzanna Becker, Sebastian Thrun, and K...

    2002

  31. [39]

    Beyond the spectrum: Detecting deepfakes via re-synthesis,

    Yang He, Ning Yu, Margret Keuper, and Mario Fritz, “Beyond the spectrum: Detecting deepfakes via re-synthesis,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence , Aug 2021

    2021

Pith tools

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