pith. sign in

arxiv: 2606.01843 · v1 · pith:LAALUDSXnew · submitted 2026-06-01 · 💻 cs.CV · cs.AI

Suppressing Forgery-Specific Shortcuts for Generalizable Deepfake Detection

Yihui Wang , Yonghui Yang , Jilong Liu , Fengbin Zhu , Le Wu , Tat-Seng Chua This is my paper

Pith reviewed 2026-06-28 15:10 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords deepfake detectiongeneralizationshortcut suppressionsubspace modelingSVDforgery artifactsfeature representationscross-method evaluation
0
0 comments X

The pith

The S^3 framework suppresses method-specific shortcut subspaces in deepfake detectors to improve generalization across unseen forgery methods.

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

Deepfake detectors often rely on spurious artifacts tied to specific forgery methods, which do not transfer to new manipulations. The paper introduces the Shortcut Subspace Suppression framework that models these shortcuts by training a linear probe on forgery method classification and extracting dominant directions via SVD. Soft suppression of this subspace during training pushes the model toward more general real-versus-fake cues. A training-free attenuation step at inference provides an additional plug-and-play boost. Experiments across benchmarks show gains in cross-method performance while preserving in-domain accuracy.

Core claim

Variations that distinguish forgery methods serve as a proxy for method-specific shortcuts; extracting their dominant subspace via SVD on a linear probe allows the model to suppress those directions softly in training and attenuate aligned neurons at inference, yielding better cross-method generalization without loss of in-domain performance.

What carries the argument

The shortcut subspace, obtained as the dominant singular vectors from SVD on features of a linear probe trained to classify forgery methods, used as a proxy to identify and suppress method-specific artifacts.

If this is right

  • Training with subspace suppression reduces reliance on forgery-method artifacts for the real/fake decision.
  • The training-free attenuation step enables immediate generalization gains on existing models without retraining.
  • Identified neurons aligned with the shortcut subspace become more interpretable as method-specific rather than general cues.
  • The approach preserves strong performance on seen forgery methods while lifting results on unseen ones.

Where Pith is reading between the lines

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

  • The same probe-plus-SVD approach could be tested on other vision tasks where class-specific shortcuts are suspected.
  • If the subspace directions prove stable across different backbone architectures, the method could serve as a lightweight post-hoc regularizer.
  • Nonlinear extensions of the probe might reveal whether linear SVD misses higher-order method-specific patterns.

Load-bearing premise

The dominant singular vectors from the forgery-method probe capture only method-specific shortcuts and do not overlap with generalizable real-versus-fake cues.

What would settle it

If suppressing the identified subspace directions causes a drop in real-versus-fake accuracy on the original training distribution comparable to the drop on unseen methods, the subspace would be shown to contain essential cues rather than removable shortcuts.

Figures

Figures reproduced from arXiv: 2606.01843 by Fengbin Zhu, Jilong Liu, Le Wu, Tat-Seng Chua, Yihui Wang, Yonghui Yang.

Figure 2
Figure 2. Figure 2: (a) Singular value energy distribution: the first few [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed S3 framework. The upper part illustrates training-time subspace suppression (NSP) with alternating probe training and gradient projection; the lower part shows inference-time neuron activation editing (NAE) with probe training and activation editing rules. 0.0 0.1 0.2 0.3 0.4 0.5 s r/f 0.0 0.5 1.0 1.5 2.0 2.5 s m Median Score (a) 1 6 11 16 21 26 31 Top k indices 0.0 0.2 0.4 0.6 0.8… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Neuron Distribution: Decision contribution [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NAE hyperparameter sensitivity. Cross-domain AUC (trained on FR, tested on EFS) varies with suppression ratio 𝜌 and strength 𝛼. Optimal performance lies at 𝜌 ∈ [1/4, 1/2] and 𝛼 = 1.0. over the baseline, indicating that the suppression mechanism mean￾ingfully influences gradient dynamics. Performance remains robust across a broad range (𝛼 = 0.1 to 0.9), with extreme values causing noticeable degradation. Op… view at source ↗
Figure 7
Figure 7. Figure 7: Neuron-level analysis: decision contribution [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Features from an FR-trained model tested on EFS, colored by real/fake. Baseline (a) shows no separation; NSP (b) establishes a clear boundary; NAE (c) improves but less than NSP. To move beyond fixed artifacts, later efforts expanded training diversity through data synthesis [14, 34, 42] or learned invariant representations via reconstruction [45], frequency constraints [37], identity disentanglement [9], … view at source ↗
Figure 10
Figure 10. Figure 10: Average AUC as a function of CLIP block index. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Complete NSP ablation. Each cell shows the average cross [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Complete NAE ablation. Each subplot shows cross [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
read the original abstract

Deepfake detection suffers from poor generalization across forgery methods, as existing models tend to rely on spurious method-specific shortcuts that fail to transfer to unseen manipulations. While recent approaches attempt to improve generalization, they lack an explicit mechanism to identify and suppress such shortcuts in learned representations. In this work, we propose Shortcut Subspace Suppression (S^3) framework that explicitly characterizes and suppresses method-specific shortcuts via subspace modeling. Our key insight is that variations distinguishing different forgery methods capture method-specific artifacts and thus serve as an effective proxy for method-specific shortcuts. To this end, we train a lightweight linear probe for forgery method classification and perform Singular Value Decomposition (SVD) to extract the dominant shortcut subspace. Building on this formulation, we develop two complementary strategies to reduce shortcut reliance. During training, we softly suppress the shortcut subspace in feature representations, encouraging the model to rely on more generalizable cues for real/fake discrimination. At inference time, we introduce a training-free counterpart that attenuates neurons aligned with the identified shortcut directions, enabling plug-and-play generalization enhancement with improved interpretability. Extensive experiments on multiple benchmarks demonstrate that our method significantly improves cross-method generalization while maintaining strong in-domain performance. The code will be released upon acceptance of the submission.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes the Shortcut Subspace Suppression (S^3) framework for deepfake detection. It trains a linear probe on forgery-method labels, extracts the dominant subspace via SVD as a proxy for method-specific shortcuts, softly suppresses this subspace in features during training to encourage generalizable real/fake cues, and applies a training-free neuron attenuation at inference. The central claim is that this yields significantly better cross-method generalization while preserving strong in-domain performance, supported by extensive experiments on multiple benchmarks.

Significance. If the core assumption holds and the reported gains are reproducible, the explicit subspace-based shortcut suppression mechanism would be a useful addition to the deepfake detection literature, particularly the training-free inference component for plug-and-play use. The promise to release code is a positive factor for reproducibility.

major comments (3)
  1. [Abstract] Abstract: the claim that 'extensive experiments on multiple benchmarks demonstrate significant gains' supplies no quantitative results, baseline comparisons, data-split details, or statistical tests, making it impossible to verify whether the data support the stated improvements in cross-method generalization.
  2. [Method] Method (linear-probe + SVD construction): the central assumption that dominant singular vectors from the method-classification probe form a clean, non-overlapping proxy for shortcuts (rather than also capturing generalizable real/fake cues) is not verified; no analysis shows orthogonality to the real/fake decision boundary or that suppression preserves in-domain accuracy for the claimed reason rather than incidental effects.
  3. [Experiments] Experiments section: absence of ablation on the number of retained singular vectors, sensitivity of the soft-suppression hyperparameter, or direct measurement of subspace overlap with real/fake features leaves the load-bearing mechanism untested.
minor comments (2)
  1. [Method] Notation for the shortcut subspace and the attenuation operation at inference should be defined with explicit equations rather than prose descriptions.
  2. [Abstract] The abstract mentions 'multiple benchmarks' without naming them; the experiments section should list the exact datasets and forgery methods used for cross-method evaluation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments. We address each major comment below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'extensive experiments on multiple benchmarks demonstrate significant gains' supplies no quantitative results, baseline comparisons, data-split details, or statistical tests, making it impossible to verify whether the data support the stated improvements in cross-method generalization.

    Authors: We agree that the abstract would benefit from including key quantitative highlights to support the generalization claims. In the revised version, we will update the abstract to report representative cross-method AUC improvements (e.g., average gains over baselines), the specific benchmarks used, and a brief note on the evaluation protocol. revision: yes

  2. Referee: [Method] Method (linear-probe + SVD construction): the central assumption that dominant singular vectors from the method-classification probe form a clean, non-overlapping proxy for shortcuts (rather than also capturing generalizable real/fake cues) is not verified; no analysis shows orthogonality to the real/fake decision boundary or that suppression preserves in-domain accuracy for the claimed reason rather than incidental effects.

    Authors: The linear probe is trained solely on forgery-method labels, providing a natural separation from the real/fake task. However, we acknowledge the need for explicit verification. We will add analysis in the revised manuscript, including cosine similarity between the shortcut subspace and real/fake classifier weights, as well as in-domain accuracy curves under varying suppression levels, to confirm the subspace primarily captures shortcuts. revision: yes

  3. Referee: [Experiments] Experiments section: absence of ablation on the number of retained singular vectors, sensitivity of the soft-suppression hyperparameter, or direct measurement of subspace overlap with real/fake features leaves the load-bearing mechanism untested.

    Authors: We agree these ablations would better substantiate the mechanism. In the revised Experiments section, we will add: (i) results varying the number of retained singular vectors, (ii) sensitivity plots for the soft-suppression hyperparameter, and (iii) direct overlap measurements such as feature projections onto the shortcut subspace and correlation with real/fake decision boundaries. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained via external operations

full rationale

The paper defines S^3 by training a linear probe on forgery-method labels followed by SVD to extract a subspace, then applies soft suppression in training and neuron attenuation at inference. These steps are procedural definitions using standard linear algebra and probing techniques applied to learned features; no equation reduces the reported cross-method generalization gain to a quantity fitted directly to the target real/fake metric or to a self-citation chain. The central modeling choice (that the method-classification subspace proxies shortcuts) is an assumption whose empirical consequences are tested on external benchmarks rather than enforced by construction. No self-citation load-bearing steps, ansatz smuggling, or renaming of known results appear in the provided derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Only the abstract is available; no explicit free parameters, background axioms, or new physical entities are described. The shortcut subspace is introduced as a modeling construct without independent evidence supplied.

invented entities (1)
  • shortcut subspace no independent evidence
    purpose: proxy for method-specific artifacts in feature space
    Defined via SVD on the method-classification probe; no external validation or falsifiable prediction is mentioned in the abstract.

pith-pipeline@v0.9.1-grok · 5763 in / 1215 out tokens · 28967 ms · 2026-06-28T15:10:40.696493+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

58 extracted references · 9 canonical work pages · 3 internal anchors

  1. [1]

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. 2018. Mesonet: a compact facial video forgery detection network. In2018 IEEE in- ternational workshop on information forensics and security (WIFS). IEEE, 1–7

    2018

  2. [2]

    Guillaume Alain and Yoshua Bengio. 2016. Understanding intermediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644(2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  3. [3]

    Inzamamul Alam, Md Tanvir Islam, and Simon S. Woo. 2025. SpecXNet: A Dual- Domain Convolutional Network for Robust Deepfake Detection. InProceedings of the 33rd ACM International Conference on Multimedia(Dublin, Ireland)(MM ’25). Association for Computing Machinery, New York, NY, USA, 11667–11676. doi:10.1145/3746027.3755707

    work page doi:10.1145/3746027.3755707 2025

  4. [4]

    David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2017. Network dissection: Quantifying interpretability of deep visual representations. InProceedings of the IEEE conference on computer vision and pattern recognition. 6541–6549

    2017

  5. [5]

    Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and ad- vances.Computational Linguistics48, 1 (2022), 207–219

    2022

  6. [6]

    Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. 2022. End-to-End Reconstruction-Classification Learning for Face Forgery Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 4113–4122

    2022

  7. [7]

    Francois Chollet. 2017. Xception: Deep Learning With Depthwise Separable Convolutions. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  8. [8]

    Davide Cozzolino, Andreas Rossler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. 2024. Raising the bar of ai-generated image detection with clip.arXiv preprint arXiv:2405.13122(2024)

    work page arXiv 2024

  9. [9]

    Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge

  10. [10]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Implicit identity leakage: The stumbling block to improving deepfake detection generalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3994–4004

  11. [11]

    Ricard Durall, Margret Keuper, and Janis Keuper. 2020. Watch your up- convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7890–7899

    2020

  12. [12]

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets.Advances in neural information processing systems27 (2014)

    2014

  13. [13]

    Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pan- tic. 2021. Lips don’t lie: A generalisable and robust approach to face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). 5039–5049

    2021

  14. [14]

    Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. 2020. Ganspace: Discovering interpretable gan controls.Advances in neural information processing systems33 (2020), 9841–9850

    2020

  15. [15]

    Ahmed Abul Hasanaath, Hamzah Luqman, Raed Katib, and Saeed Anwar. 2024. FSBI: Deepfakes Detection with Frequency Enhanced Self-Blended Images.ArXiv abs/2406.08625 (2024). https://api.semanticscholar.org/CorpusID:270440586

    work page arXiv 2024

  16. [16]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  17. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  18. [18]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models.arXiv preprint arXiv:2106.09685(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  19. [19]

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al . 2018. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). InInternational conference on machine learning. PMLR, 2668–2677

    2018

  20. [20]

    Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides

  21. [21]

    InProceedings of the IEEE/CVF international conference on computer vision (ICCV)

    Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. InProceedings of the IEEE/CVF international conference on computer vision (ICCV). 21011–21021

  22. [22]

    HyunJae Lee, Hyo-Eun Kim, and Hyeonseob Nam. 2019. SRM: A Style-Based Recalibration Module for Convolutional Neural Networks. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2019

  23. [23]

    Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Bain- ing Guo. 2020. Face x-ray for more general face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). 5001–5010

    2020

  24. [24]

    Yuezun Li and Siwei Lyu. 2019. Exposing deepfake videos by detecting face warping artifacts. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (CVPRW). 46–52

    2019

  25. [25]

    Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. 2021. Spatial-phase shallow learning: rethinking face forgery detection in frequency domain. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 772–781

    2021

  26. [26]

    Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Guru- datt, and Wael AbdAlmageed. 2020. Two-branch recurrent network for isolating deepfakes in videos. InEuropean conference on computer vision. Springer, 667–684

    2020

  27. [27]

    Yisroel Mirsky and Wenke Lee. 2021. The creation and detection of deepfakes: A survey.ACM computing surveys (CSUR)54, 1 (2021), 1–41

    2021

  28. [28]

    Yunsheng Ni, Depu Meng, Changqian Yu, Chengbin Quan, Dongchun Ren, and Youjian Zhao. 2022. Core: Consistent representation learning for face forgery detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (CVPRW). 12–21

    2022

  29. [29]

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. 2023. Towards universal fake image detectors that generalize across generative models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24480–24489

    2023

  30. [30]

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. 2020. Think- ing in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision. Springer, 86–103

    2020

  31. [31]

    Tong Qiao, Shichuang Xie, Yizhi Chen, Florent Retraint, and Xiangyang Luo. 2024. Fully unsupervised deepfake video detection via enhanced contrastive learning. InProceedings of the IEEE/CVF winter conference on applications of computer vision (W ACV). 4691–4700

    2024

  32. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings of the 38th Inter- national Conference on Machine Learning (Proceedings of Machi...

    2021

  33. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  34. [34]

    Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. 2019. FaceForensics++: Learning to Detect Manipulated Facial Images. InInternational Conference on Computer Vision (ICCV)

    2019

  35. [35]

    Minenko, Dmitrii I

    Ayush Roy, Sk Mohiuddin, Maxim V. Minenko, Dmitrii I. Kaplun, and Ram Sarkar

  36. [36]

    InVISIGRAPP : VISAPP

    DeepSpace: Navigating the Frontier of Deepfake Identification Using Attention-Driven Xception and a Task-Specific Subspace. InVISIGRAPP : VISAPP. https://api.semanticscholar.org/CorpusID:276760616

  37. [37]

    Kaede Shiohara and Toshihiko Yamasaki. 2022. Detecting deepfakes with self- blended images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 18720–18729

    2022

  38. [38]

    Stefan Smeu, Elisabeta Oneata, and Dan Oneata. 2025. DeCLIP: Decoding CLIP representations for deepfake localization. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 149–159

    2025

  39. [39]

    Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. 2021. Improving the efficiency and robustness of deepfakes detection through precise geometric features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). 3609–3618

    2021

  40. [40]

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. 2024. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 5052–5060. Conference’17, July 2017, Washington, DC, USA Yihui Wang et al

    2024

  41. [41]

    Mingxing Tan and Quoc Le. 2019. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. InProceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97), Kamalika Chaudhuri and Ruslan Salakhutdinov (Eds.). PMLR, 6105–

    2019

  42. [42]

    https://proceedings.mlr.press/v97/tan19a.html

  43. [43]

    Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. 2020. Deepfakes and beyond: A survey of face manipulation and fake detection.Information fusion64 (2020), 131–148

    2020

  44. [44]

    Chengrui Wang and Weihong Deng. 2021. Representative forgery mining for fake face detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). 14923–14932

    2021

  45. [45]

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. 2023. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 22445– 22455

    2023

  46. [46]

    Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. 2024. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8984–8994

    2024

  47. [47]

    Zhiyuan Yan, Jiangming Wang, Peng Jin, Ke-Yue Zhang, Chengchun Liu, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. 2024. Orthogonal subspace decomposition for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [48]

    Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, and Li Yuan

  49. [49]

    InAdvances in Neural Information Processing Systems, A

    DF40: Toward Next-Generation Deepfake Detection. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37. Curran Associates, Inc., 29387–29434. doi:10.52202/079017-0925

    work page doi:10.52202/079017-0925

  50. [50]

    Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. 2023. Ucf: Uncovering common features for generalizable deepfake detection. InProceedings of the IEEE/CVF international conference on computer vision. 22412–22423

    2023

  51. [51]

    Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. 2023. Deepfakebench: A comprehensive benchmark of deepfake detection.arXiv preprint arXiv:2307.01426(2023)

    work page arXiv 2023

  52. [52]

    Kelu Yao, Jin Wang, Boyu Diao, and Chao Li. 2023. Towards understanding the generalization of deepfake detectors from a game-theoretical view. InProceedings of the IEEE/CVF international conference on computer vision. 2031–2041

    2023

  53. [53]

    Andrii Yermakov, Jan Cech, and Jiri Matas. 2025. Unlocking the hidden potential of CLIP in generalizable deepfake detection.arXiv preprint arXiv:2503.19683 (2025)

    work page arXiv 2025

  54. [54]

    Zixin Yin, Jiakai Wang, Yisong Xiao, Hanqing Zhao, Tianlin Li, Wenbo Zhou, Aishan Liu, and Xianglong Liu. 2024. Improving deepfake detection generalization by invariant risk minimization.IEEE Transactions on Multimedia26 (2024), 6785– 6798

    2024

  55. [55]

    Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia

  56. [56]

    InProceedings of the IEEE/CVF international conference on computer vision (ICCV)

    Learning self-consistency for deepfake detection. InProceedings of the IEEE/CVF international conference on computer vision (ICCV). 15023–15033

  57. [57]

    Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. 2021. Ex- ploring temporal coherence for more general video face forgery detection. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV). 15044–15054

    2021

  58. [58]

    training on FS

    Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. 2022. Uia-vit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. InEuropean conference on computer vision (ECCV). Springer, 391–407. Suppressing Forgery-Specific Shortcuts for Generalizable Deepfake Detecti...

    2022