pith. sign in

arxiv: 2606.26384 · v1 · pith:TKFFUWTVnew · submitted 2026-06-24 · 💻 cs.CV

What Do Deepfake Benchmarks Measure? An Audit Using Frozen Self-Supervised Representations

Samuel Pagon , Yixuan Shen , Vishal Asnani , Feng Liu This is my paper

Pith reviewed 2026-06-26 01:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords deepfake detectionbenchmark auditself-supervised representationslinear probesforensic understandingrepresentation geometrymultimodal benchmarks
0
0 comments X

The pith

Deepfake benchmarks largely reward general modality understanding rather than forensic skills.

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

The paper audits deepfake benchmarks in video, image, and audio by testing a deliberately simple diagnostic: whether a linear probe on frozen general-purpose self-supervised representations can approach the accuracy of purpose-built detectors. If the probe succeeds, the benchmark is mostly measuring broad understanding of the data modality instead of the specific ability to spot manipulations. A reader would care because this pattern implies that benchmark scores may reflect progress on an easier problem than real-world detection, where detectors must handle unseen generators and distribution shifts. The audit further ties differences in generator difficulty to geometric properties of the same representation space.

Core claim

Across three modalities, linear probes trained on frozen self-supervised representations closely approach the performance of bespoke deepfake detectors on standard benchmarks. Generator-level difficulty rankings are partly explained by Frechet geometry distances computed in the identical representation space. These observations indicate that the benchmarks are largely solved by general-purpose representations and therefore measure modality understanding more than forensic understanding of fakes.

What carries the argument

Linear probe on frozen general-purpose self-supervised representations, serving as a diagnostic that isolates how much of benchmark performance is already captured without task-specific forensic training.

If this is right

  • High scores on existing benchmarks cannot be read as direct evidence of forensic understanding.
  • Detector development may be optimizing for signals already present in general representations.
  • Benchmark design should incorporate controls that separate general modality features from manipulation-specific cues.
  • Generator difficulty can be predicted in advance using geometry in a fixed representation space.

Where Pith is reading between the lines

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

  • The same audit could be run on other detection or classification benchmarks to check whether they also collapse to general representations.
  • Future benchmarks might deliberately include examples where general representations fail, forcing models to learn forensic features.
  • Deployment settings where distribution shift is large may still require forensic-specific training even if benchmarks appear solved.

Load-bearing premise

That a linear probe matching bespoke detector performance on the benchmark means the benchmark is testing general modality understanding rather than forensic understanding.

What would settle it

A bespoke detector that substantially outperforms the frozen linear probe on the benchmark while also showing markedly better generalization to real-world unseen deepfakes would falsify the audit conclusion.

Figures

Figures reproduced from arXiv: 2606.26384 by Feng Liu, Samuel Pagon, Vishal Asnani, Yixuan Shen.

Figure 1
Figure 1. Figure 1: AIGVDBench. (a) Macro-average per-generator AUC on the full test split as a function of V-JEPA2 layer for logistic regression (LR) and ridge probes. (b) Reported overall average AUC across 20 open-source and 11 closed-source generators for the top-10 detectors in the AIGVDBench paper and our top LR/Ridge probes. 4.1 Linear probes reveal strong benchmark separability 4.1.1 Video: AIGVDBench The strongest fr… view at source ↗
Figure 2
Figure 2. Figure 2: Celeb-DF++. (a) Layer-wise macro-average per-generator AUC on the Celeb-DF++ GF-eval test split for logistic regression (LR) and ridge probes on DINOv3 representations. (b) Reported frame-level average AUC from Celeb-DF++ paper for detectors trained on Celeb-DF and evaluated on Celeb-DF++; and our top LR/Ridge probes. 4.1.3 Audio: ASVspoof2019 LA and English MLAAD The audio setting requires a slightly diff… view at source ↗
Figure 3
Figure 3. Figure 3: Audio. (a) Layer-wise macro-average per-generator spoof accuracy on the MLAAD v9 English split (84 generators) for logistic regression (LR) and ridge probes on XLS-R representations. (b) Top-5 single systems on ASVspoof2019 LA by EER, alongside the top-5 of our probes on the same LA evaluation split. Transfer to English MLAAD. On the current English MLAAD split (84 target generators), logistic regression p… view at source ↗
Figure 4
Figure 4. Figure 4: AIGVDBench: layer-wise Pearson and Spearman correlations between per-generator AUC and the three Fréchet-distance summaries dreal, dspoof, ∆ for logistic regression (LR) and ridge probes on V-JEPA2 representations. Across most layers, dspoof is negatively correlated with performance and ∆ is positively correlated with performance, while dreal is weaker and less stable ( [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 5
Figure 5. Figure 5: Celeb-DF++: layer-wise Pearson and Spearman correlations between per-generator AUC and the three Fréchet-distance summaries for logistic regression (LR) and ridge probes on DINOv3 representations. r = 0.630 at layer 18 and the strongest Spearman is ρ = 0.689 at layer 15; for ridge, r = 0.629 at layer 18 and ρ = 0.728 at layer 14. The absolute distance to source real is informative but weaker, peaking at r … view at source ↗
Figure 6
Figure 6. Figure 6: English MLAAD: layer-wise Pearson and Spearman correlations between per-generator spoof accuracy and the three Fréchet-distance summaries for logistic regression (LR) and ridge probes on XLS-R representations. On MLAAD ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: AIGVDBench: per-generator AUC at the strongest probe configuration (V-JEPA2 layer 22, ridge). Macro-average 88.51, median 92.50, range 63.00–100.00. Strong performance is broadly distributed; a small number of generators (wan, pika, Luma) remain meaningfully harder. Open-Sora and Opensora are distinct AIGVDBench generator labels: Open-Sora denotes the open-source generator used for source training, while O… view at source ↗
Figure 8
Figure 8. Figure 8: Celeb-DF++: per-generator AUC at the strongest probe configuration (DINOv3 layer 6, logistic regression). Macro-average 79.72, median 82.22, range 60.45–94.81 across face-swap, face-reenactment, and talking-face methods. Compute resources. All experiments used frozen pretrained backbones; we did not fine-tune V￾JEPA2, DINOv3, or XLS-R. The main computational cost was one-time feature extraction, followed b… view at source ↗
Figure 9
Figure 9. Figure 9: English MLAAD: per-generator spoof accuracy at the strongest probe configuration (XLS-R layer 18, ridge). Macro-average 88.84%, median 93.1%, range 39.2–100.0% across 84 TTS systems. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
read the original abstract

As deepfake generators approach perceptual indistinguishability, reliable detection becomes critical. Yet, detectors that score well on benchmarks routinely fail in the wild. A concerning feedback loop has emerged: benchmarks drive increasingly complex, engineered detectors, yet if those benchmarks do not reflect real-world deepfakes, this complexity may be solving the wrong problem entirely. This raises a prior question: what are these benchmarks actually measuring? We conduct an audit of video, image, and audio deepfake benchmarks using a deliberately simple diagnostic. If a linear probe on frozen, general-purpose self-supervised representations can approximate the performance of a bespoke detector, the benchmark is largely rewarding general modality understanding rather than forensic understanding. This has two implications: the benchmark may not reflect realistic threat models, and it raises the question of whether the bespoke detectors the probe approaches are truly learning forensic understanding. We observe, across three modalities, linear probes on general-purpose self-supervised representations closely approach the performance of bespoke detectors. We further show that generator-level difficulty is partly explained by Frechet geometry in the same representation space. Together, these results support a benchmark-audit view of deepfake detection: before high scores are read as evidence of forensic understanding, it is worth asking how much of the benchmark is already solved by general-purpose representations.

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

2 major / 1 minor

Summary. The paper audits deepfake benchmarks across video, image, and audio modalities using a diagnostic based on linear probes trained on frozen general-purpose self-supervised representations. It claims that if such probes closely approach the performance of bespoke detectors, the benchmarks largely reward general modality understanding rather than forensic understanding. The authors report that this holds in their experiments and that generator-level difficulty is partly explained by Fréchet geometry in the same representation spaces.

Significance. If the central diagnostic and its interpretation hold after clarification, the work supplies a lightweight, reproducible method for auditing whether deepfake benchmarks align with intended forensic goals. It could shift how benchmark results are interpreted and encourage more realistic threat models. The reliance on public SSL models and the geometric analysis of difficulty are concrete strengths that make the approach falsifiable and extensible.

major comments (2)
  1. [Abstract] Abstract (diagnostic premise): The claim that linear-probe success means the benchmark is 'largely rewarding general modality understanding rather than forensic understanding' requires evidence that forensic cues (blending boundaries, spectral artifacts, etc.) are not linearly separable in the SSL embedding space. If they are linearly accessible, the probe is still performing forensic detection and the closeness to bespoke detectors only shows that those detectors also use linearly accessible signals. This assumption is load-bearing for the audit's central conclusion.
  2. [Results] Results and methods description: The statement that probes 'closely approach' bespoke performance is central to the claim, yet the provided text gives no quantitative metrics, statistical tests, confidence intervals, or controls for confounds (e.g., dataset overlap, probe regularization). Without these, it is impossible to verify whether the observed closeness supports the diagnostic or could be explained by other factors.
minor comments (1)
  1. [Results] The Fréchet-geometry analysis is presented as downstream support, but the manuscript should clarify how the geometry is computed (e.g., which layers, covariance estimation) and whether it is independent of the linear-probe results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our audit of deepfake benchmarks. The comments help clarify the scope of our diagnostic and the need for stronger quantitative support. We address each major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract (diagnostic premise): The claim that linear-probe success means the benchmark is 'largely rewarding general modality understanding rather than forensic understanding' requires evidence that forensic cues (blending boundaries, spectral artifacts, etc.) are not linearly separable in the SSL embedding space. If they are linearly accessible, the probe is still performing forensic detection and the closeness to bespoke detectors only shows that those detectors also use linearly accessible signals. This assumption is load-bearing for the audit's central conclusion.

    Authors: We agree this distinction is important and that the original abstract wording risks overstating the separation between 'general' and 'forensic' signals. Our intended claim is narrower: because the SSL representations were trained without any forensic supervision, their ability to linearly approximate bespoke detector performance indicates that the benchmark does not require learning representations specialized for the detection task. This still supports questioning whether high benchmark scores reflect capabilities that would transfer beyond the distribution captured by general-purpose models. We will revise the abstract and introduction to make this scope explicit and to note that linear separability of forensic cues within SSL space remains an open question not resolved by the current experiments. revision: partial

  2. Referee: [Results] Results and methods description: The statement that probes 'closely approach' bespoke performance is central to the claim, yet the provided text gives no quantitative metrics, statistical tests, confidence intervals, or controls for confounds (e.g., dataset overlap, probe regularization). Without these, it is impossible to verify whether the observed closeness supports the diagnostic or could be explained by other factors.

    Authors: The full manuscript contains tables with per-modality accuracy/F1 numbers comparing linear probes to published bespoke detectors, but we acknowledge the absence of formal statistical comparisons and confound controls in the version seen by the referee. In revision we will add: (1) bootstrap confidence intervals on all reported metrics, (2) paired statistical tests between probe and detector performance, (3) explicit checks for train/test overlap between the SSL pretraining corpora and the deepfake benchmarks, and (4) ablation results on probe regularization strength and hidden-layer choice. These additions will be placed in a new 'Quantitative validation' subsection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external models and detectors

full rationale

The paper conducts an empirical audit by applying linear probes to publicly available frozen self-supervised representations (e.g., standard SSL models) and comparing performance to published bespoke detectors. No equations, fitted parameters, or results reduce by construction to quantities defined or optimized inside the paper. The central diagnostic premise is an interpretive inference from these external comparisons rather than a self-referential definition or self-citation chain. The Frechet geometry observation is likewise computed in the same external representation space without internal fitting that forces the outcome. This is a standard non-circular empirical audit against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper introduces no new free parameters or invented entities. It relies on one domain assumption about the diagnostic power of linear probes.

axioms (1)
  • domain assumption A linear probe on frozen general-purpose self-supervised representations is a valid diagnostic for whether a benchmark measures general modality understanding versus forensic understanding.
    This premise underpins the entire audit and the interpretation of results.

pith-pipeline@v0.9.1-grok · 5764 in / 1112 out tokens · 23252 ms · 2026-06-26T01:22:55.340285+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

47 extracted references · 13 canonical work pages · 8 internal anchors

  1. [1]

    A superb-style benchmark of self-supervised speech models for audio deepfake detection

    Hashim Ali, Nithin Sai Adupa, Surya Subramani, and Hafiz Malik. A superb-style benchmark of self-supervised speech models for audio deepfake detection. InICASSP, 2026

    2026

  2. [2]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    AI deepfakes blur reality in 2026 US midterm campaigns

    Joseph Ax and Helen Coster. AI deepfakes blur reality in 2026 US midterm campaigns. Reuters, March 2026

    2026

  4. [4]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296, 2021

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick V on Platen, Yatharth Saraf, Juan Pino, et al. Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296, 2021

    work page arXiv 2021

  5. [5]

    Is space-time attention all you need for video understanding? InICML, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, 2021

    2021

  6. [6]

    Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024

    Nuria Alina Chandra, Ryan Murtfeldt, Lin Qiu, Arnab Karmakar, Hannah Lee, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Sejin Paik, Changyeon Lee, et al. Deepfake-eval- 2024: A multi-modal in-the-wild benchmark of deepfakes circulated in 2024.arXiv preprint arXiv:2503.02857, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Forgelens: Data-efficient forgery focus for general- izable forgery image detection

    Yingjian Chen, Lei Zhang, and Yakun Niu. Forgelens: Data-efficient forgery focus for general- izable forgery image detection. InICCV, 2025

    2025

  8. [8]

    Can we leave deepfake data behind in training deepfake detector? InNeurIPS, 2024

    Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? InNeurIPS, 2024

    2024

  9. [9]

    Xception: Deep learning with depthwise separable convolutions

    François Chollet. Xception: Deep learning with depthwise separable convolutions. InCVPR, 2017

    2017

  10. [10]

    Raising the bar of ai-generated image detection with clip

    Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. InCVPR, 2024

    2024

  11. [11]

    Forensics adapter: Adapting clip for generalizable face forgery detection

    Xinjie Cui, Yuezun Li, Ao Luo, Jiaran Zhou, and Junyu Dong. Forensics adapter: Adapting clip for generalizable face forgery detection. InCVPR, 2025

    2025

  12. [12]

    The DeepFake Detection Challenge (DFDC) Dataset

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset.arXiv preprint arXiv:2006.07397, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  13. [13]

    X3d: Expanding architectures for efficient video recognition

    Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR, 2020

    2020

  14. [14]

    On the content bias in fréchet video distance

    Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, Jun-Yan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. InCVPR, 2024

    2024

  15. [15]

    A kernel two-sample test.The Journal of Machine Learning Research, 13(1):723–773, 2012

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The Journal of Machine Learning Research, 13(1):723–773, 2012

    2012

  16. [16]

    Leveraging real talking faces via self-supervision for robust forgery detection

    Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. InCVPR, 2022

    2022

  17. [17]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InNeurIPS, 2017

    2017

  18. [18]

    Implicit identity driven deepfake face swapping detection

    Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. InCVPR, 2023. 10

    2023

  19. [19]

    Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection

    Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. InCVPR, 2020

    2020

  20. [20]

    Fr\'echet Audio Distance: A Metric for Evaluating Music Enhancement Algorithms

    Kevin Kilgour, Mauricio Zuluaga, Dominik Roblek, and Matthew Sharifi. Fr\’echet audio dis- tance: A metric for evaluating music enhancement algorithms.arXiv preprint arXiv:1812.08466, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  21. [21]

    Uni- formerv2: Unlocking the potential of image vits for video understanding

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uni- formerv2: Unlocking the potential of image vits for video understanding. InICCV, 2023

    2023

  22. [22]

    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. InCVPR, 2020

    2020

  23. [23]

    Celeb-df++: A large-scale challenging video deepfake benchmark for generalizable forensics.arXiv preprint arXiv:2507.18015, 2025

    Yuezun Li, Delong Zhu, Xinjie Cui, and Siwei Lyu. Celeb-df++: A large-scale challenging video deepfake benchmark for generalizable forensics.arXiv preprint arXiv:2507.18015, 2025

    work page arXiv 2025

  24. [24]

    Your one-stop solution for ai-generated video detection.arXiv preprint arXiv:2601.11035, 2026

    Long Ma, Zihao Xue, Yan Wang, Zhiyuan Yan, Jin Xu, Xiaorui Jiang, Haiyang Yu, Yong Liao, and Zhen Bi. Your one-stop solution for ai-generated video detection.arXiv preprint arXiv:2601.11035, 2026

    work page arXiv 2026

  25. [25]

    Detecting ai-generated video via frame consistency

    Long Ma, Zhiyuan Yan, Qinglang Guo, Yong Liao, Haiyang Yu, and Pengyuan Zhou. Detecting ai-generated video via frame consistency. InICME, 2025

    2025

  26. [26]

    Mlaad: The multi-language audio anti-spoofing dataset

    Nicolas M Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger. Mlaad: The multi-language audio anti-spoofing dataset. InIJCNN, 2024

    2024

  27. [27]

    Andreas Nautsch, Xin Wang, Nicholas Evans, Tomi H Kinnunen, Ville Vestman, Massimiliano Todisco, Héctor Delgado, Md Sahidullah, Junichi Yamagishi, and Kong Aik Lee. Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech.IEEE Transactions on Biometrics, Behavior , and Identity Science, 3(2):252–265, 2021

    2019

  28. [28]

    Exploring self-supervised vision trans- formers for deepfake detection: A comparative analysis

    Huy H Nguyen, Junichi Yamagishi, and Isao Echizen. Exploring self-supervised vision trans- formers for deepfake detection: A comparative analysis. InIJCB, 2024

    2024

  29. [29]

    Genvidbench: A 6-million benchmark for ai-generated video detection

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. Genvidbench: A 6-million benchmark for ai-generated video detection. InAAAI, 2026

    2026

  30. [30]

    Attorney general jeff jackson warns north carolinians of investment scams on meta platforms

    North Carolina Department of Justice. Attorney general jeff jackson warns north carolinians of investment scams on meta platforms. Press Release, April 2026

    2026

  31. [31]

    INVESTOR ALERT: Attorney general james warns new yorkers of investment scams on meta platforms

    Office of the New York State Attorney General. INVESTOR ALERT: Attorney general james warns new yorkers of investment scams on meta platforms. Press Release, April 2026

    2026

  32. [32]

    Towards universal fake image detectors that generalize across generative models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. InCVPR, 2023

    2023

  33. [33]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InICML, 2021

    2021

  34. [34]

    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. InICCV, 2019

    2019

  35. [35]

    Oriane Siméoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    An information theoretic approach for attention-driven face forgery detection

    Ke Sun, Hong Liu, Taiping Yao, Xiaoshuai Sun, Shen Chen, Shouhong Ding, and Rongrong Ji. An information theoretic approach for attention-driven face forgery detection. InECCV, 2022

    2022

  37. [37]

    Synthetic audio forensics evaluation (safe) challenge.arXiv preprint arXiv:2510.03387, 2025

    Kirill Trapeznikov, Paul Cummer, Pranay Pherwani, Jai Aslam, Michael S Davinroy, Peter Bautista, Laura Cassani, Matthew Stamm, and Jill Crisman. Synthetic audio forensics evaluation (safe) challenge.arXiv preprint arXiv:2510.03387, 2025. 11

    work page arXiv 2025

  38. [38]

    Towards Accurate Generative Models of Video: A New Metric & Challenges

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  39. [39]

    Cnn- generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn- generated images are surprisingly easy to spot... for now. InCVPR, 2020

    2020

  40. [40]

    Yunet: A tiny millisecond-level face detector.Machine Intelligence Research, 2023

    Wei Wu, Hanyang Peng, and Shiqi Yu. Yunet: A tiny millisecond-level face detector.Machine Intelligence Research, 2023

    2023

  41. [41]

    Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection

    Zhiyuan Yan, Jiangming Wang, Zhendong Wang, Peng Jin, Ke-Yue Zhang, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Effort: Efficient orthogonal modeling for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    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. InICCV, 2023

    2023

  43. [43]

    Deepfakebench: A comprehensive benchmark of deepfake detection.arXiv preprint arXiv:2307.01426, 2023

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

    work page arXiv 2023

  44. [44]

    Deepfake detection that generalizes across benchmarks

    Andrii Yermakov, Jan Cech, Jiri Matas, and Mario Fritz. Deepfake detection that generalizes across benchmarks. InWACV, 2026

    2026

  45. [45]

    Bank of italy warns over deepfake video scams using governor panetta

    Valentina Za. Bank of italy warns over deepfake video scams using governor panetta. Reuters, February 2026

    2026

  46. [46]

    D3: Training-free ai-generated video detection using second-order features

    Chende Zheng, Ruiqi Suo, Chenhao Lin, Zhengyu Zhao, Le Yang, Shuai Liu, Minghui Yang, Cong Wang, and Chao Shen. D3: Training-free ai-generated video detection using second-order features. InICCV, 2025

    2025

  47. [47]

    Open-Sora: Democratizing Efficient Video Production for All

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 12 A Appendix Open-Sora Pyramid-Flow SEINE IPOC SVD LTX Cogvideox1.5 VideoCrafter RepVideo EasyAnimate HunyuanVideo AccVideo Opensora caus...

    work page internal anchor Pith review Pith/arXiv arXiv 2024