Pith. sign in

REVIEW 4 major objections 5 minor 48 references

KLASSify to Verify: Audio-Visual Deepfake Detection Using SSL-based Audio and Handcrafted Visual Features

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A lightweight audio-visual deepfake detector reaches 92.78% AUC on AV-Deepfake1M++ testA by calibrating and maxout-fusing an SSL audio score with a handcrafted-feature video score.

desk verdict A competent challenge-paper write-up whose headline fusion gain is inflated by test-set selection; worth refereeing mainly to force the authors to show the complementarity claim is real. read the letter →

arxiv 2508.07337 v1 pith:BZPOARIR submitted 2025-08-10 eess.AS cs.CV

Ivan Kukanov , Jun Wah Ng This is my paper
classification eess.AScs.CV
keywords deepfakedetectionaudio-visualfusiontemporallocalizationself-supervisedlearninggraphattentionnetworkshandcraftedfeaturesconvolutionalnetworkscorecalibration
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

The paper sets out to show that a computationally modest audio-visual system can perform competitively on a large-scale temporal deepfake benchmark by keeping the modalities separate and combining them only at the decision level. It reports that on AV-Deepfake1M++ testA the audio model alone, a Wav2Vec 2.0 front end with an AASIST graph attention head trained with codec-style augmentations, reaches 82.91% AUC, and the video model alone, a 124K-parameter temporal convolutional network over eight handcrafted mouth-region features, reaches 73.11% AUC. After Platt-calibrating the two score streams and taking their maximum, the fused system reaches 92.78% AUC, beating simple score averaging (91.97%) as well as either modality on its own. For the localization task, the authors adapt a boundary-aware attention model using audio only and report IoU 0.3536 on testA. The broader point is that real-world deepfake detection may not require heavyweight multimodal encoders, and that handcrafted cues plus calibrated score fusion can be both interpretable and competitive.

What carries the argument

The load-bearing mechanism is score-level max-out fusion of calibrated unimodal probabilities. Each modality outputs a softmax score; Platt sigmoid scaling with parameters fit on 40% of the validation split turns these into calibrated probabilities, and the final video-level decision is the larger of the two calibrated probabilities. The audio side is powered by Wav2Vec 2.0 XLSR-53 self-supervised embeddings classified by AASIST, a spectro-temporal graph attention network; the video side is a lightweight temporal convolutional network over eight handcrafted features targeting mouth-region artifacts left by lip-sync generators. For localization, the key mechanism is the boundary-aware attenti

What would settle it

Compute per-clip binary errors of the audio-only and video-only models on testA and measure their overlap; then re-fit the Platt calibration on a different 40% of validation and recompute maxout AUC. If the two error sets overlap heavily, or the gain over averaging disappears on the new split, the 92.78% fusion result is not a stable multimodal complementarity.

Watch

Extended reading notes

Core claim

The paper's central claim is that a deliberately simple multimodal system can detect and localize temporal deepfakes in AV-Deepfake1M++ without heavy fusion backbones. On the classification task, it decouples the two modalities: a Wav2Vec 2.0 XLSR-53 front end feeding an AASIST graph attention head is trained on audio with codec-style augmentations (82.91% AUC on testA), while a 124K-parameter temporal convolutional network is trained on eight handcrafted mouth-region features such as mouth blurriness, non-mouth MSE, color shift, and landmark kinematics (73.11% AUC on testA). The two scores are Platt-calibrated on 40% of the validation split and combined by a max-out operation, reaching 92.7

Load-bearing premise

The fusion's edge over score averaging rests on the audio and video models making errors on different clips, and the paper does not report per-sample error agreement; if their errors overlap, the maxout gain could be an artifact of the 40% validation subset used to fit calibration.

Editorial extensions

If this is right

  • A 124K-parameter TCN over handcrafted features can out-generalize a much larger visual baseline (Xception) on unseen attack settings, suggesting lightweight visual detectors are viable for constrained deployments.
  • Codec-style audio augmentation yields the best single audio model on testA (82.91% AUC), better than an ensemble of audio models and a model trained with many augmentations, indicating compression artifacts are a powerful training signal.
  • Platt calibration plus maxout adds about 0.8 AUC over score averaging at fusion time, so multimodal gains can come from decision-level fusion rather than joint feature learning.
  • Audio-only boundary-aware localization reaches IoU 0.3536 with AP@0.5 of 0.5117, while video-only localization reaches only IoU 0.1139, indicating the audio stream carries most of the localization signal for these short edits.
  • Utterance-level models trained on fully fake or fully real audio degrade sharply on partial deepfakes; segment-aware training and evaluation are necessary.

Reading between the lines

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

  • Because the paper never reports per-sample agreement between the audio and video models, the stability of the maxout gain is untested; a learned or confidence-weighted fusion might outperform maxout if the two streams are unequally reliable, since audio alone is about 10 AUC points stronger than video alone.
  • The weak video-localization result (IoU 0.1139) compared with audio localization (0.3536) suggests the handcrafted visual features are better for whole-clip classification than for finding segment boundaries; fusing audio boundaries with visual cues could improve temporal precision.
  • The dataset's fake segments average 0.33 seconds, about one word, so a practical system must operate at near-phoneme granularity; this work's sliding-window evaluation could be pushed toward streaming or online detection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper addresses the AV-Deepfake1M++ 2025 challenge. For Task 1 (video-level deepfake classification), it proposes a video stream that extracts eight handcrafted mouth/face features and feeds them to a lightweight 124K-parameter TCN, an audio stream based on Wav2Vec 2.0 XLSR-53 with an AASIST graph-attention head and codec augmentations, and score-level fusion using Platt calibration followed by max-out. On the testA split, the paper reports an AUC of 92.78% for the fused system, compared with 82.91% for audio-only, 73.11% for video-only, and 91.97% for simple score averaging. For Task 2 (temporal localization), the paper adapts the boundary-aware attention mechanism BAM for audio-only localization, reporting IoU 0.3536 on testA, and a TCN with a tagging head for video-only localization, reporting IoU 0.1139. The abstract and conclusion present the system as efficient, robust, and competitive, with an emphasis on interpretability through handcrafted features.

Significance. If the reported fusion result is taken at face value as the performance of a fixed pipeline, the paper offers a useful lightweight audio-visual detector and a clean ablation of handcrafted visual features. The strengths are the explicit ablations over feature groups and audio augmentations, the calibration analysis, the very small video model (124K parameters), and the adaptation of partial-spoof localization methods to a much larger benchmark. However, the central claim of a 9.87-point testA AUC gain from fusion over audio-only is currently not backed by protocol evidence: the configuration appears to be selected on testA, no error bars or multiple runs are reported, and no per-sample complementarity analysis is provided. The significance of the paper therefore depends on additional validation.

major comments (4)
  1. [Section 5.2, Table 2] The reported 92.78% testA AUC is likely a configuration-selection artifact rather than a fixed-pipeline result. The text states that "score averaging across all windows yields the best performance on the TestA set," and Table 2 reports testA AUC for multiple audio augmentations and two fusion rules. If the final choice of augmentations, fusion rule, and windowing strategy was made after inspecting testA numbers, the reported number is a selected maximum over a family of testA scores. The paper must either evaluate on the held-out testB split, perform nested validation, or otherwise justify that the choices were made without testA feedback.
  2. [Section 4.1, Section 5.2, Table 2] The claimed fusion gain is not supported by error-correlation evidence. The video model has testA AUC 73.11, substantially lower than the audio model's 82.91, so a max-out fusion reaching 92.78 implies that the video model is correct on many samples where the audio model is wrong. No per-sample agreement, confusion-analysis, or failure-analysis between modalities is reported. Moreover, on validation the fused system (98.04) is below the audio-only system (99.71), so the complementarity only appears on testA. The authors should report per-sample error overlap or at least per-modality coverage statistics to demonstrate that the fusion gain is not an artifact of calibration or configuration selection.
  3. [Section 6, Table 2] There is no uncertainty quantification or reproducibility evidence for the main result. All numbers are single-run measurements without error bars, multiple seeds, or significance tests; the 92.78% vs 91.97% difference between max-out and averaging could be within run-to-run noise. For a benchmark claim, the authors should provide variance estimates or confirm the result on testB. Additionally, the testA entry for Baseline Wav2Vec-AASIST-2 is missing in Table 2, making the comparison incomplete.
  4. [Section 6, Table 2] The large validation-to-testA drop for the video model is not explained: validation AUC is 88.41 while testA AUC is 73.11. This 15-point drop is important because the fusion result depends on the video model's testA behavior. The paper should discuss the distribution shift or overfitting that causes this drop, and report whether the same drop occurs on testB. Also, the text claims baselines achieve "63.53%" on testA, but Table 2 reports 60.53%; this discrepancy should be corrected.
minor comments (5)
  1. [References] References [3] and [4] appear to be the same paper (same title, same journal, same year, same page range). Please merge or distinguish them.
  2. [Table 1] In the audio_modified row of the validation column, the percentage "24,49%" uses a comma as decimal separator, while other entries use a period. Please make formatting consistent.
  3. [Section 5.2] The phrase "using torchvision1 library" has a formatting error (the superscript reference marker is not rendered properly), and the audio preprocessing is underspecified: sampling rate, number of channels, and conversion details should be stated.
  4. [Figure 5] The calibration curves show that the video model's Brier score slightly worsens after Platt scaling (0.1354 to 0.1368), but the text says the video scores are close to perfect calibration. Please clarify whether the calibrated video scores are actually used in the max-out fusion and how the worse Brier score is reconciled with the claimed benefit of calibration.
  5. [Section 6] The sentence "Achieving the IoU of 0.1139 on testA, using only handcrafted video features. Our approach was 3% lower than the baseline method, BA-TFD+ results the IoU of 0.1471" is grammatically awkward. Also, '3% lower' should be stated as a 3.32-point IoU difference (0.1139 vs 0.1471), not a relative percentage.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the central claims are external benchmark measurements; the only self-citation is minor and non-load-bearing.

full rationale

The paper's central claims are benchmark measurements on the external AV-Deepfake1M++ testA and validation splits, not derivations from assumptions that already contain the outcome. The audio pipeline uses the published Wav2Vec-AASIST (Tak et al.) with standard augmentations; the video pipeline is a lightweight TCN on handcrafted features; fusion is score-level Platt calibration fit on the validation split followed by a fixed maxout rule. None of these steps defines the target AUC in terms of itself: the calibration parameters are fit to validation labels, and the testA numbers are evaluated rather than generated by the fitting equations. The one self-citation ([19], Kukanov et al.) is used only as background for a baseline/training line and is not load-bearing; the architecture's primary references are [31] for audio and [45] for localization. No uniqueness theorem or ansatz is imported from same-author work to force the choice. The paper also explicitly acknowledges limitations, such as cross-dataset generalization of localization, which indicates the results are empirical rather than definitional. The sentence 'Empirically, it was found that score averaging across all windows yields the best performance on the TestA set' hints at possible test-set-based configuration selection, but that is a data-snooping concern, not a circular reduction, and does not change the circularity verdict. Overall, no circular step can be exhibited; score 2 reflects only the presence of a minor, non-load-bearing self-citation.

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

The paper's central claims rest on a handful of domain assumptions about feature universality, context-window coverage, and calibration transfer, plus several hand-tuned hyperparameters. It introduces no new entities. The largest implicit burden is that handcrafted visual artifacts persist across generators and that maxout of calibrated scores yields complementarity; neither is demonstrated with cross-generator or error-correlation evidence.

free parameters (5)
  • Platt calibration parameters a,b = not reported
    Eq. (2) in Section 4.1; fitted on 40% of validation split, used for maxout fusion. Without these, audio and video scores cannot be compared.
  • Temporal tagging loss weights = [0.05, 0.30, 0.30, 0.35]
    Section 5.2; chosen by hand (not derived) for O/B/I/L tags in localization head; affect frame boundary predictions.
  • Audio training duration cutoff = 6 s
    Section 5.2; selected so the 4 s context window covers a fake segment, but no evidence this guarantees coverage for all fake durations or short videos.
  • Video training duration cutoff = 10.24 s (256 frames)
    Section 5.2; training subset chosen for TCN, not justified with ablation.
  • Handcrafted feature set = 8 features grouped as 4 families
    Section 4.1; features 1-4 chosen empirically, no individual ablation beyond grouped Table 2 results.
assumptions (5)
  • domain assumption Wav2Vec 2.0 XLSR-53 features plus AASIST graph attention can detect short (mean 0.33s) partially spoofed segments from video-level labels.
    Section 4.1 audio modality. The paper acknowledges no prior work on Wav2Vec-AASIST in weakly supervised scenarios; the 4s context coverage is assumed sufficient.
  • domain assumption The eight handcrafted visual features (blurriness, non-mouth MSE, color shift, kinematics) are universal deepfake artifacts that persist across generators.
    Section 4.1: 'we empirically observed'; no quantitative cross-generator evidence is provided. Video model testA AUC 73.11 suggests limited transfer.
  • domain assumption Platt calibration trained on 40% of validation and maxout selection transfer to testA.
    Section 4.1 Eq(2) and Section 5.2. No evidence that calibration is stable across splits or that testA and validation have the same score distribution.
  • domain assumption The AV-Deepfake1M++ labels and train/val/test splits are correct as provided.
    Whole paper relies on challenge dataset [5] for training and evaluation.
  • domain assumption The published BAM checkpoint trained on PartialSpoof is a useful initialization for AV-Deepfake1M++ after finetuning.
    Section 5.2 audio localization: 'pre-initialized with checkpoint weights from [45]'. The authors report the pretrained model fails without finetuning (EER 77.52%), so the value of this init is empirical.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KLASSify to Verify: Audio-Visual Deepfake Detection Using SSL-based Audio and Handcrafted Visual Features." pith.science (2026). https://pith.science/paper/BZPOARIR

@misc{pith2026250807337,
  author       = {Pith},
  title        = {Pith review of: KLASSify to Verify: Audio-Visual Deepfake Detection Using SSL-based Audio and Handcrafted Visual Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZPOARIR}},
  note         = {Machine review of arXiv:2508.07337}
}
read the original abstract

The rapid development of audio-driven talking head generators and advanced Text-To-Speech (TTS) models has led to more sophisticated temporal deepfakes. These advances highlight the need for robust methods capable of detecting and localizing deepfakes, even under novel, unseen attack scenarios. Current state-of-the-art deepfake detectors, while accurate, are often computationally expensive and struggle to generalize to novel manipulation techniques. To address these challenges, we propose multimodal approaches for the AV-Deepfake1M 2025 challenge. For the visual modality, we leverage handcrafted features to improve interpretability and adaptability. For the audio modality, we adapt a self-supervised learning (SSL) backbone coupled with graph attention networks to capture rich audio representations, improving detection robustness. Our approach strikes a balance between performance and real-world deployment, focusing on resilience and potential interpretability. On the AV-Deepfake1M++ dataset, our multimodal system achieves AUC of 92.78% for deepfake classification task and IoU of 0.3536 for temporal localization using only the audio modality.

Figures

Figures reproduced from arXiv: 2508.07337 by the authors.

Figure 1
Figure 1. KLASS Solution Pipeline for Deepfake Classification. Self-supervised Learning (SSL)-based features followed by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the lightweight TCN [21] used for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the lightweight TCN [21] with a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Histogram of the most frequently edited words in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Calibration curves for audio and video models be [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 30 canonical work pages

  1. [1]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. Wav2Vec 2.0: A Framework for Self-Supervised Learning of Speech Represen- tations. In Advances in Neural Information Processing Systems . H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, (Eds.) Vol. 33. Curran Associates, Inc., 12449–12460

    work page 2020

  2. [2]

    Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. 2024. AV-Deepfake1M: A Large-scale LLM-driven Audio-Visual Deepfake Dataset. In Proceedings of the 32nd ACM International Conference on Multimedia , 7414–7423

    work page 2024

  3. [3]

    Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. 2023. Glitch in the matrix: A large scale benchmark for content driven audio–visual forgery detection and localization. Computer Vision and Image Understanding, 236, 103818. doi:https://doi.org/10.1016/j.cviu.2023.10381 8

    work page arXiv 2023

  4. [4]

    Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. 2023. Glitch in the matrix: A large scale benchmark for content driven audio–visual forgery detection and localization. Computer Vision and Image Understanding, 236, 103818

    work page 2023

  5. [5]

    Zhixi Cai, Kartik Kuckreja, Shreya Ghosh, Akanksha Chuchra, Muhammad Haris Khan, Usman Tariq, Tom Gedeon, and Abhinav Dhall. 2025. AV-Deep- fake1M++: A Large-Scale Audio-Visual Deepfake Benchmark with Real-World Perturbations. (2025). arXiv: 2507.20579

    work page Pith review arXiv 2025

  6. [6]

    Zhixi Cai, Kalin Stefanov, Abhinav Dhall, and Munawar Hayat. 2022. Do You Really Mean That? Content Driven Audio-Visual Deepfake Dataset and Multi- modal Method for Temporal Forgery Localization. In 2022 International Con- ference on Digital Image Computing: Techniques and Applications (DICTA) , 1– 10

    work page 2022

  7. [7]

    Edresson Casanova, Julian Weber, Christopher Shulby, Arnaldo Candido Junior, Eren Gölge, and Moacir Antonelli Ponti. 2023. YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for everyone. (2023). arXiv: 2112.02418 [cs.SD]

    arXiv 2023

  8. [8]

    Edresson Casanova et al. 2024. XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model. (2024). arXiv: 2406.04904 [eess.AS]

    arXiv 2024

Show all 48 references
  1. [10]

    Sanyuan Chen et al. 2021. WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing. CoRR, abs/2110.13900

    2021

  2. [11]

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

    2017

  3. [12]

    Soumyya Kanti Datta, Shan Jia, and Siwei Lyu. 2024. Exposing Lip-syncing Deepfakes from Mouth Inconsistencies. (2024). arXiv: 2401.10113 [cs.CV]

    2024 arXiv

  4. [13]

    Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. 2021. Spatiotemporal Inconsistency Learning for DeepFake Video Detection. (2021). arXiv: 2109.01860 [cs.CV]

    2021 arXiv

  5. [14]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE / ACM Trans. Audio, Speech and Lang. Proc. , 29, (Oct. 20...

    2021

  6. [15]

    Jee Weon Jung, Hee Soo Heo, Hemlata Tak, Hye Jin Shim, Joon Son Chung, Bong Jin Lee, Ha Jin Yu, and Nicholas Evans. 2022. AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks. In 2022 IEEE International Conference on Acoustics, Speech, and Sig...

    2022

  7. [16]

    Jaehyeon Kim, Jungil Kong, and Juhee Son. 2021. Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech. (2021). arXiv: 2106.06103 [cs.SD]

    2021 arXiv

  8. [17]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Optimization. (2017). arXiv: 1412.6980

    2017 arXiv

  9. [18]

    Seltzer, and Sanjeev Khudanpur

    Tom Ko, Vijayaditya Peddinti, Daniel Povey, Michael L. Seltzer, and Sanjeev Khudanpur. 2017. A Study on Data Augmentation of Reverberant Speech for Robust Speech Recognition. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 5220–5224

    2017

  10. [19]

    Ivan Kukanov, Janne Laakkonen, Tomi Kinnunen, and Ville Hautamaki. 2024. Meta-Learning Approaches For Improving Detection of Unseen Speech Deep- fakes. In 2024 IEEE Spoken Language Technology Workshop (SLT) , 1173–1178. doi:10.1109/SLT61566.2024.10832350

    2024

  11. [20]

    Galina Lavrentyeva, Sergey Novoselov, Andzhukaev Tseren, Marina Volkova, Artem Gorlanov, and Alexandr Kozlov. 2019. STC Antispoofing Systems for the ASVspoof2019 Challenge. (2019). arXiv: 1904.05576 [cs.SD]

    2019 arXiv

  12. [21]

    Colin Lea, Rene Vidal, Austin Reiter, and Gregory D. Hager. 2016. Temporal Convolutional Networks: A Unified Approach to Action Segmentation. (2016). arXiv: 1608.08242 [cs.CV]

    2016 arXiv

  13. [22]

    Chunyu Li, Chao Zhang, Weikai Xu, Jingyu Lin, Jinghui Xie, Weiguo Feng, Bingyue Peng, Cunjian Chen, and Weiwei Xing. 2025. LatentSync: Taming Audio-Conditioned Latent Diffusion Models for Lip Sync with SyncNet Super- vision. (2025). arXiv: 2412.09262 [cs.CV]

    2025 arXiv

  14. [23]

    Jiatong Liu, Lina Wang, Run Wang, Jianpeng Ke, Xi Ye, Yadi Wu, and Yu-an Tan

  15. [24]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regulariza- tion. (2019). arXiv: 1711.05101 [cs.LG]

    2019 arXiv

  16. [25]

    Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. (2017). arXiv: 1608.03983 [cs.LG]

    2017 arXiv

  17. [26]

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McCormick, Alexander Roff, Jonathon Shlens, Michael Mazzocchi, Ming Guang Thornton, and Yannis Papakonstantinou. 2019. MediaPipe: A Framework for Building Perception Pipelines. (2019). arXiv: 1906.08172 [cs.CV]

    2019 arXiv

  18. [27]

    Soumik Mukhopadhyay, Saksham Suri, Ravi Teja Gadde, and Abhinav Shrivas- tava. 2023. Diff2Lip: Audio Conditioned Diffusion Models for Lip-Synchroniza- tion. (2023). arXiv: 2308.09716 [cs.CV]

    2023 arXiv

  19. [28]

    Müller, Pavel Czempin, Franziska Dieckmann, Adam Froghyar, and Konstantin Böttinger

    Nicolas M. Müller, Pavel Czempin, Franziska Dieckmann, Adam Froghyar, and Konstantin Böttinger. 2024. Does Audio Deepfake Detection Generalize? (2024). arXiv: 2203.16263

    2024

  20. [29]

    John C. Platt. [n. d.] Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods. In Advances in Large Margin Classifiers. MIT Press, 61–74

  21. [30]

    Lev Ratinov and Dan Roth. 2009. Design Challenges and Misconceptions in Named Entity Recognition. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009) , 147–155

    2009

  22. [31]

    Hemlata Tak, Massimiliano Todisco, et al. 2022. Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation. In Odyssey

    2022

  23. [32]

    Lingfeng Tan, Yunhong Wang, Junfu Wang, Liang Yang, Xunxun Chen, and Yuanfang Guo. 2023. Deepfake video detection via facial action dependencies estimation. In (AAAI’23/IAAI’23/EAAI’23) Article 589. AAAI Press, 9 pages. isbn: 978-1-57735-880-0. doi:10.1609/aaai.v37i4.25658

    2023 doi

  24. [33]

    Massimiliano Todisco et al. 2019. ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection. In Interspeech

    2019

  25. [34]

    Tan, and Haizhou Li

    Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T. Tan, and Haizhou Li

  26. [35]

    Xin Wang et al. 2024. ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale. ArXiv, abs/2408.08739

    2024 arXiv

  27. [36]

    Sophia Koepke, and Andrew Zisserman

    Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. 2018. X2Face: A network for controlling face generation by using images, audio, and pose codes. (2018). arXiv: 1807.10550 [cs.CV]

    2018 arXiv

  28. [37]

    Xinqi Xiong, Prakrut Patel, Qingyuan Fan, Amisha Wadhwa, Sarathy Selvam, Xiao Guo, Luchao Qi, Xiaoming Liu, and Roni Sengupta. 2025. TalkingHead- Bench: A Multi-Modal Benchmark & Analysis of Talking-Head DeepFake Detection. (2025). arXiv: 2505.24866 [cs.CV]

    2025

  29. [38]

    Junichi Yamagishi, Xin Wang, et al. 2021. ASVspoof 2021: Accelerating Progress in Spoofed and Deepfake Speech Detection. In Proc. 2021 Edition of the Auto- matic Speaker Verification and Spoofing Countermeasures Challenge

    2021

  30. [39]

    Lin Zhang, Xin Wang, Erica Cooper, Nicholas Evans, and Junichi Yamagishi

  31. [40]

    Lin Zhang, Xin Wang, Erica Cooper, and Junichi Yamagishi. 2021. Multi-task Learning in Utterance-level and Segmental-level Spoof Detection. In Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 9–15. doi:10.21437/ASVSPOOF.2021-2

    2021 doi

  32. [41]

    Lin Zhang, Xin Wang, Erica Cooper, Junichi Yamagishi, Jose Patino, and Nicholas Evans. 2021. An Initial Investigation for Detecting Partially Spoofed Audio. In Proc. Interspeech 2021 , 4264–4268. doi:10.21437/Interspeech.2021-738

    2021 doi

  33. [42]

    doi:10.1109/TASLP.20 22.3233236

    The PartialSpoof Database and Countermeasures for the Detection of Short Fake Speech Segments Embedded in an Utterance.IEEE/ACM Transactions on Audio, Speech, and Language Processing , 31, 813–825. doi:10.1109/TASLP.20 22.3233236

    doi

  34. [43]

    You Zhang, Baotong Tian, Lin Zhang, and Zhiyao Duan. 2025. PartialEdit: iden- tifying partial deepfakes in the era of neural speech editing. InProc. Interspeech

    2025

  35. [44]

    Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. 2021. Exploring Temporal Coherence for More General Video Face Forgery Detection. (2021). arXiv: 2108.06693 [cs.CV]

    2021 arXiv

  36. [45]

    Rui Zhang, Hongxia Wang, Mingshan Du, Hanqing Liu, Yang Zhou, and Qiang Zeng. 2023. UMMAFormer: A Universal Multimodal-adaptive Transformer Framework for Temporal Forgery Localization. In Proceedings of the 31st ACM International Conference on Multimedia (MM ’23). ACM, (Oct. 2...

    2023

  37. [46]

    Xiaojin Zhu. 2005. Semi-Supervised Learning Literature Survey. Tech. rep. 1530. Computer Sciences, University of Wisconsin-Madison. Received 1 August 2025

    2005

  38. [48]

    Jiafeng Zhong, Bin Li, and Jiangyan Yi. 2024. Enhancing Partially Spoofed Audio Localization with Boundary-aware Attention Mechanism. arXiv: 2407.21611

    2024 arXiv

  39. [2023]

    Seeing What You Said: Talking Face Generation Guided by a Lip Reading Expert. (2023). arXiv: 2303.17480 [cs.CV]

    2023 arXiv

  40. [2025]

    Exposing the Forgery Clues of DeepFakes via Exploring the Inconsistent Expression Cues. 2025. doi:10.1155/int/7945646

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.