Pith. sign in

REVIEW 2 major objections 5 minor 45 references

General AIGC audio-video detection should not rely on audio-visual correspondence; independently modeled modality evidence fused at the decision level is more robust.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:06 UTC pith:GMG7YJB7

load-bearing objection A credible, well-engineered challenge winner whose central claim about decision-level fusion being more robust isn't actually tested by its own ablations. the 2 major comments →

arxiv 2607.25543 v1 pith:GMG7YJB7 submitted 2026-07-28 cs.CV cs.AI

Less is More: Modality-Decoupling for General AIGC Audio-Video Detection

Jielun Peng , Yabin Wang , Yaqi Li , Jincheng Liu , Xiaopeng Hong , Athanasios V. Vasilakos This is my paper
classification cs.CV cs.AI
keywords AIGC detectionaudio-visual deepfake detectionmodality decouplingdecision-level fusionmulti-granularity visual representationtemporal-spectral audio analysiscorrespondence assumptiongeneral scene forgery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the usual assumption behind audio-visual deepfake detection—that real videos show strong correspondence between audio and visual content—breaks down in general AI-generated content, where audio and video are often only weakly or arbitrarily related. It demonstrates this with a pretrained audio-visual similarity model: on human-centric deepfake data, real samples score higher similarity than fakes, but on general-scene data the ordering inverts, so using correspondence alone performs worse than random guessing. The authors conclude that, for general AIGC, fusing per-modality forensic evidence at the decision level is more robust than learning cross-modal features. They build DAV-Det as two independent detectors—visual and audio—and fuse their probabilities with max and product rules. On the challenge leaderboard the system ranks first with a final score of 0.8460.

Core claim

The central discovery is that cross-modal inconsistency is not a reliable forgery cue in general scenes. Using a pretrained model to measure audio-video cosine similarity, the paper reports an AUC of 82.48% on a human-centric deepfake dataset (real samples score higher than fake) but only 40.28% on a general-scene benchmark (real samples score lower than fake), meaning the correspondence signal is actively anti-correlated with ground truth. Consequently, detectors that exploit audio-visual correspondence can be harmed rather than helped. The paper proposes instead to treat each modality as independent forensic evidence, with separate visual and audio detectors, and to fuse only their final d

What carries the argument

The load-bearing design is decision-level fusion. For binary detection, the system takes the maximum of the audio and visual fake probabilities—a sample is fake if either modality seems fake. For four-class labels (real-real, fake-fake, video-fake/audio-real, video-real/audio-fake) it multiplies the independently estimated per-modality probabilities, effectively assuming conditional independence. The visual detector feeds a vision transformer's global, patch, and segment tokens through separate scoring heads to catch multi-granularity spatial artifacts; the audio detector combines temporal and spectral branches with a learned gate. The paper uses an audio-visual correspondence model's simila

Load-bearing premise

The load-bearing premise is that one pretrained audio-visual similarity score on a single general-scene benchmark faithfully captures the correspondence cue that feature-level fusion models actually exploit; if that proxy is wrong, the negative AUC does not generalize to all feature-level fusion methods.

What would settle it

Train a strong feature-level audio-visual fusion detector on the same general-scene training data and show it matches or beats the decoupled system at comparable compute; or, on the same general-scene benchmark, compute audio-visual similarity with a different correspondence model and show real samples score higher than fakes, overturning the reported inversion.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correspondence is unreliable, existing audio-visual detectors that align speech with lip motion will transfer poorly to general scenes, motivating benchmark content with weakly correlated audio and video.
  • Modality-decoupled systems can be deployed as two independent detectors, so a failure or domain shift in one modality does not drag down the other.
  • Decision-level fusion gives an interpretable division of labor: a sample is flagged when either forensic stream is confident, and the four-class scores can point to which modality was likely manipulated.
  • The independent per-modality probabilities yield a simple probabilistic model for the four AIGC types, opening a path to calibrated detection without training a joint fusion network.

Where Pith is reading between the lines

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

  • An untested consequence is that decision-level fusion should also improve robustness to attacks aimed at a single modality, since an attack on the audio stream cannot corrupt the visual branch's evidence; this could be tested by attacking one modality at a time.
  • If conditional independence fails, the product rule for four-class probabilities will be miscalibrated; a testable extension is to estimate the covariance between audio and visual fake scores and correct the joint probabilities.
  • The paper's negative result on correspondence is based on one similarity model on one general-scene dataset; a stronger claim would require showing the same inversion across multiple correspondence measures and several general-scene datasets.
  • The 'less is more' principle suggests that expensive cross-modal pretraining for general AIGC may be unnecessary, so future challenge entries could cut compute and still remain competitive.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes DAV-Det, a decoupled audio-visual AIGC detection system that independently trains a visual detector and an audio detector and fuses their predictions at the decision level. The motivation is an empirical observation that audio-visual correspondence, the basis of many existing multimodal detectors, is not a reliable cue in general AIGC scenarios: using PEAV cosine similarity as a training-free score yields AUC 82.48% on FakeAVCeleb but only 40.28% on MVAD (Fig. 1). The visual detector uses global, patch, and segment-level representations with multiple-instance learning and degraded-original consistency learning; the audio detector uses a gated temporal-spectral dual-branch architecture. At inference, binary classification uses a max-based rule and four-class classification uses an independence-based product rule (Eq. 24). The system ranks first in the DDL 2.0 challenge (final score 0.8460) and achieves high accuracy on FakeAVCeleb. Ablations (Tabs. 4-6) validate each within-modality component.

Significance. If the central claim were fully demonstrated, the paper would provide a useful design principle for general audio-visual AIGC detection: decouple modalities and fuse at the decision level rather than at the feature level. The challenge result is credible, the component ablations are consistent, and the code release supports reproducibility. The PEAV-based observation is an interesting motivation and a falsifiable finding. However, the central claim comparing fusion strategies is not directly tested; the evidence is indirect (a frozen correspondence score and a complete-system leaderboard comparison). The paper would be strengthened by a matched fusion ablation. The four-class independence assumption is also acknowledged as heuristic and needs empirical justification.

major comments (2)
  1. [§1, §4.4] The central claim that decision-level fusion is more robust than feature-level fusion is not directly tested. The PEAV experiment in Fig. 1 shows that a frozen correspondence model has inverted AUC on MVAD, but trained feature-level fusion models (e.g., AVFF-style) may exploit cues not captured by PEAV cosine similarity. The ablations in Tabs. 4-6 vary within-modality components but never compare decision-level fusion against feature-level fusion using the same audio and visual representations. The challenge result (Table 2) compares complete systems, not fusion strategies. Thus the main claim remains a plausible heuristic, as the Limitations section itself acknowledges. Please add a matched experiment: e.g., concatenate the final visual and audio embeddings and train a feature-level fusion classifier while holding all other components fixed, and report the comparison.
  2. [§3.4, Eq. (24)] The four-class prediction multiplies independently estimated per-modality real/fake probabilities, implicitly assuming conditional independence. The authors flag this in the Limitations, but the four-class score contributes 60% of the official final score, so the validity of this assumption is material. If audio and visual fake probabilities are positively correlated within real or fake samples, the product rule will miscalibrate the four-class rankings. Please provide an empirical check, such as measuring the correlation between p_fake_v and p_fake_a on a validation set, or comparing the product rule with a simple learned calibration that relaxes independence. This would also strengthen the paper's claim about the advantages of decision-level fusion.
minor comments (5)
  1. [Fig. 1 caption] Typo: 'PEA V' should be 'PEAV'.
  2. [§2.2 and Table 3] Inconsistent notation: 'A V oid-DF' in text vs 'A V oiD-DF' in Table 3. Please standardize.
  3. [§3.1] The image sampling scheme (8 images from real videos, 16 from fake videos) is described but its effect on the effective class distribution during visual training is not discussed. Please clarify whether this creates a 1:2 class imbalance and how the loss handles it.
  4. [§3.3, Eq. (16)] The gating network G(·) is defined over the sequence of frame tokens but the exact input aggregation (e.g., concatenation of CLS and frames) is not specified. A short sentence would improve reproducibility.
  5. [Table 2] The team name 'MV ADetection Team' contains an unnecessary space. Also, 'AIGVDete' appears truncated; check the original name.

Circularity Check

0 steps flagged

No significant circularity: the empirical motivation uses an external fixed feature extractor, and the proposed system is trained and evaluated against external benchmarks.

full rationale

The derivation chain is not circular. The motivating observation (Sec. 1, Fig. 1) uses a fixed external model, PEAV, to compute audio-visual cosine similarities on FakeAVCeleb and MVAD; the reported AUCs (82.48% vs 40.28%) are measurements of an external pretrained representation under a defined ranking protocol, not quantities fitted in this paper. The proposed DAV-Det then trains visual and audio detectors using MVAD-derived data and public labels, and its performance is measured on the official DDL-GAV challenge leaderboard and FakeAVCeleb under a standard 70/30 split; these are external evaluation targets. Internal ablations (Tabs. 4-6) are component-level comparisons, not predictions that reduce to their training inputs. Eq. (24) multiplies per-modality probabilities to form four-class joint probabilities; the paper itself flags this as an independence heuristic in the Limitations section, and a heuristic modeling choice is a correctness/calibration risk, not circularity. The self-citations ([Peng et al. 2026], [Wang et al. 2024/2025/2026]) appear in related work and motivation context only; none is invoked as a uniqueness theorem, a forced ansatz, or a substitute for the paper's own training/evaluation. The paper does not include a matched ablation that replaces decision-level fusion with feature-level fusion, so the comparative claim 'decision-level fusion is more robust' is under-supported as a general thesis; that gap concerns experimental completeness, not circularity. Therefore score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim relies mainly on domain assumptions about where forgery artifacts live and on two heuristic fusion rules. The paper introduces no new physical entities. The PEAV similarity proxy and the independence assumption are the most consequential load-bearing premises; the first motivates the whole decoupling argument, the second affects four-class calibration.

free parameters (6)
  • softmax temperature tau = 0.07
    Chosen for patch and segment softmax weighting in Eq. 1 and Eq. 5; not derived.
  • margin m for margin loss = 0.6
    Hand-set margin in Eq. 11 to separate top-k instances from remaining patches/segments.
  • top-k ratios = 0.05 patches, 0.1 segments
    Selected ratios defining the suspicious subset in Eq. 7; no justification beyond ablations.
  • agglomerative clustering cosine threshold = 0.9
    Threshold for merging segments in §3.2; chosen by hand and affects segment-level representation.
  • DOCL consistency loss weight = 0.05
    Weight on L_con in Eq. 15; chosen by hand.
  • LoRA ranks/alphas = visual rank 32 alpha 16; audio rank 32 alpha 64
    Hyperparameters for parameter-efficient fine-tuning; not derived from theory.
axioms (5)
  • domain assumption PEAV cosine similarity faithfully measures the audio-visual content correspondence that prior detectors rely on.
    Used in §1/Fig. 1 to conclude the correspondence assumption fails on MVAD; if the proxy is wrong, the motivation weakens.
  • domain assumption AIGC visual artifacts appear at global, patch, and segment levels; audio artifacts appear in temporal and spectral irregularities.
    Design premise of §3.2 and §3.3; reasonable but not proven beyond the paper's ablations.
  • ad hoc to paper Max-based binary fusion: a sample is fake if either modality has high fake probability.
    Eq. 22; a heuristic decision rule, acknowledged in limitations as heuristic rather than learned.
  • ad hoc to paper Audio and visual fake probabilities are independent for four-class joint inference.
    Eq. 24 multiplies p_v and p_a probabilities; the authors state in Limitations that this independence assumption may not fully hold.
  • standard math Standard deep learning background: self-attention, softmax, focal loss, LoRA, and pretrained DINOv3/PEAV features transfer forensic cues.
    Background machinery used throughout §3; not proved in this paper but standard in the field.

pith-pipeline@v1.3.0-alltime-deepseek · 13408 in / 11122 out tokens · 110935 ms · 2026-08-01T02:06:31.653513+00:00 · methodology

0 comments
read the original abstract

Generative AI has rapidly expanded audio-visual forgery beyond human-centric deepfakes into general scenes. Existing AIGC detection methods assume audio-visual content correspondence, identifying forgeries by spotting cross-modal inconsistencies. However, we empirically find that this assumption does not consistently hold in general scenarios. We argue that, for general audio-visual AIGC detection, decision-level fusion is a more robust alternative to feature-level fusion. Therefore, we propose DAV-Det, a decoupled audio-visual AIGC detection system that independently models forensic evidence from each modality. The visual detector leverages multi-granularity representations at global, patch, and segment levels to capture spatial forgery cues, while the audio detector exploits both temporal and spectral irregularities via a gated temporal-spectral dual-branch architecture to model acoustic artifacts. Our method ranks 1st in the General AIGC Audio-Video Detection Challenge of the IJCAI-ECAI 2026 DDL 2.0 Workshop, with a final score of 0.8460. Code is available at https://github.com/tuffy-studio/DAV-Det.

Figures

Figures reproduced from arXiv: 2607.25543 by Athanasios V. Vasilakos, Jielun Peng, Jincheng Liu, Xiaopeng Hong, Yabin Wang, Yaqi Li.

Figure 1
Figure 1. Figure 1: Distributions of cosine similarities between audio and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed DAV-Det. (a) The visual AIGC detector models global, patch-, and segment-level representations to capture forgery cues at multiple spatial granularities. (b) The audio AIGC detector characterizes temporal dynamics and spectral anomalies in the audio stream for comprehensive AIGC detection. detect forgeries by measuring the matching degree between audio and video, without requiring … view at source ↗

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

45 extracted references · 11 linked inside Pith

  1. [1]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations.Advances in neural information processing sys- tems, 33:12449–12460,

    [Baevskiet al., 2020 ] Alexei Baevski, Yuhao Zhou, Abdel- rahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations.Advances in neural information processing sys- tems, 33:12449–12460,

    2020

  2. [7]

    Insid3: Training-free in-context segmentation with dinov3

    [Cuttanoet al., 2026 ] Claudia Cuttano, Gabriele Trivigno, Christoph Reich, Daniel Cremers, Carlo Masone, and Ste- fan Roth. Insid3: Training-free in-context segmentation with dinov3. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 21638–21648,

    2026

  3. [8]

    Pia: Deepfake detection using phoneme-temporal and identity-dynamic analysis

    [Dattaet al., 2025 ] Soumyya Kanti Datta, Tanvi Ranga, Chengzhe Sun, and Siwei Lyu. Pia: Deepfake detection using phoneme-temporal and identity-dynamic analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1596–1606,

    2025

  4. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    [Dosovitskiyet al., 2020 ] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Min- derer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Pith/arXiv arXiv 2020

  5. [11]

    Self-supervised video forensics by audio-visual anomaly detection

    [Fenget al., 2023 ] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10491–10503,

    2023

  6. [12]

    Allm4add: Unlocking the capabilities of audio large language models for audio deepfake detec- tion

    [Guet al., 2025 ] Hao Gu, Jiangyan Yi, Chenglong Wang, Jianhua Tao, Zheng Lian, Jiayi He, Yong Ren, Yujie Chen, and Zhengqi Wen. Allm4add: Unlocking the capabilities of audio large language models for audio deepfake detec- tion. InProceedings of the 33rd ACM International Con- ference on Multimedia, pages 11736–11745,

    2025

  7. [13]

    Lips don’t lie: A generalisable and robust approach to face forgery detection

    [Haliassoset al., 2021 ] Alexandros Haliassos, Konstantinos V ougioukas, Stavros Petridis, and Maja Pantic. 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, pages 5039– 5049,

    2021

  8. [16]

    Mvad: A comprehensive multimodal video-audio dataset for aigc detection.arXiv preprint arXiv:2512.00336,

    [Huet al., 2025 ] Mengxue Hu, Yunfeng Diao, Changtao Miao, Jianshu Li, Zhe Li, and Joey Tianyi Zhou. Mvad: A comprehensive multimodal video-audio dataset for aigc detection.arXiv preprint arXiv:2512.00336,

    Pith/arXiv arXiv 2025

  9. [17]

    Aasist: Audio anti-spoofing us- ing integrated spectro-temporal graph attention networks

    [Junget al., 2022 ] Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu, and Nicholas Evans. Aasist: Audio anti-spoofing us- ing integrated spectro-temporal graph attention networks. InICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 6367–6371. IEEE,

    2022

  10. [18]

    Fakeavceleb: A novel audio- video multimodal deepfake dataset.arXiv preprint arXiv:2108.05080,

    [Khalidet al., 2021 ] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S Woo. Fakeavceleb: A novel audio- video multimodal deepfake dataset.arXiv preprint arXiv:2108.05080,

    Pith/arXiv arXiv 2021

  11. [19]

    iwax: interpretable wav2vec-aasist-xgboost framework for voice spoofing detection.Scientific reports, 15(1):40491,

    [Leeet al., 2025 ] Seungeun Lee, Sunmook Choi, Taein Kang, Sanghyeok Chung, Soyul Han, Jaejin Seo, Seoy- oung Park, Eujin Kim, Seungsang Oh, and Il-Youp Kwak. iwax: interpretable wav2vec-aasist-xgboost framework for voice spoofing detection.Scientific reports, 15(1):40491,

    2025

  12. [20]

    Speechforensics: Audio- visual speech representation learning for face forgery de- tection.Advances in Neural Information Processing Sys- tems, 37:86124–86144,

    [Lianget al., 2024 ] Yachao Liang, Min Yu, Gang Li, Jian- guo Jiang, Boquan Li, Feng Yu, Ning Zhang, Xiang Meng, and Weiqing Huang. Speechforensics: Audio- visual speech representation learning for face forgery de- tection.Advances in Neural Information Processing Sys- tems, 37:86124–86144,

    2024

  13. [21]

    Focal loss for dense ob- ject detection

    [Linet al., 2017 ] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll ´ar. Focal loss for dense ob- ject detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988,

    2017

  14. [23]

    Multi-modal deepfake detection via multi-task audio-visual prompt learning

    [Miaoet al., 2025 ] Hui Miao, Yuanfang Guo, Zeming Liu, and Yunhong Wang. Multi-modal deepfake detection via multi-task audio-visual prompt learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 612–621,

    2025

  15. [24]

    Modern hierarchical, agglomerative clustering algorithms.arXiv preprint arXiv:1109.2378,

    [M¨ullner, 2011] Daniel M ¨ullner. Modern hierarchical, agglomerative clustering algorithms.arXiv preprint arXiv:1109.2378,

    Pith/arXiv arXiv 2011

  16. [26]

    Avff: Audio- visual feature fusion for video deepfake detection

    [Oorloffet al., 2024 ] Trevine Oorloff, Surya Koppisetti, Nicol`o Bonettini, Divyaraj Solanki, Ben Colman, Yaser Yacoob, Ali Shahriyari, and Gaurav Bharaj. Avff: Audio- visual feature fusion for video deepfake detection. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27102–27112,

    2024

  17. [27]

    Leave no stone unturned: Un- covering holistic audio-visual intrinsic coherence for deep- fake detection

    [Penget al., 2026 ] Jielun Peng, Yabin Wang, Yaqi Li, Long Kong, and Xiaopeng Hong. Leave no stone unturned: Un- covering holistic audio-visual intrinsic coherence for deep- fake detection. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 6655–6666,

    2026

  18. [28]

    Learning transferable visual models from nat- ural language supervision

    [Radfordet al., 2021 ] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR,

    2021

  19. [29]

    Robust speech recognition via large-scale weak supervision

    [Radfordet al., 2023 ] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. InInternational conference on machine learning, pages 28492–28518. PMLR,

    2023

  20. [30]

    Learning audio-visual speech representation by masked multimodal cluster pre- diction.arXiv preprint arXiv:2201.02184,

    [Shiet al., 2022 ] Bowen Shi, Wei-Ning Hsu, Kushal Lakho- tia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster pre- diction.arXiv preprint arXiv:2201.02184,

    Pith/arXiv arXiv 2022

  21. [31]

    Detecting deepfakes with self-blended im- ages

    [Shiohara and Yamasaki, 2022] Kaede Shiohara and Toshi- hiko Yamasaki. Detecting deepfakes with self-blended im- ages. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 18720–18729,

    2022

  22. [32]

    Dinov3.arXiv preprint arXiv:2508.10104,

    [Sim´eoniet al., 2025 ] Oriane Sim ´eoni, Huy V V o, Maxim- ilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104,

    Pith/arXiv arXiv 2025

  23. [33]

    Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning

    [Smeuet al., 2025 ] Stefan Smeu, Dragos-Alexandru Bold- isor, Dan Oneata, and Elisabeta Oneata. Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 18815– 18825,

    2025

  24. [34]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation

    [Taket al., 2022 ] Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee-weon Jung, Junichi Yamagishi, and Nicholas Evans. Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation. arXiv preprint arXiv:2202.12233,

    Pith/arXiv arXiv 2022

  25. [35]

    Attention is all you need.Advances in neural information processing systems, 30,

    [Vaswaniet al., 2017 ] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30,

    2017

  26. [36]

    Pushing the frontier of audiovisual per- ception with large-scale multimodal correspondence learn- ing

    [Vyaset al., 2026 ] Apoorv Vyas, Heng-Jui Chang, Cheng- Fu Yang, Po-Yao Huang, Luya Gao, Julius Richter, Sanyuan Chen, Matthew Le, Piotr Doll ´ar, Christoph Fe- ichtenhofer, et al. Pushing the frontier of audiovisual per- ception with large-scale multimodal correspondence learn- ing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2026

  27. [37]

    Facex-zoo: A pytorch toolbox for face recognition

    [Wanget al., 2021 ] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. InProceedings of the 29th ACM international conference on Multimedia, pages 3779–3782,

    2021

  28. [38]

    Linguistic profiling of deep- fakes: An open database for next-generation deepfake de- tection.arXiv preprint arXiv:2401.02335,

    [Wanget al., 2024 ] Yabin Wang, Zhiwu Huang, Zhiheng Ma, and Xiaopeng Hong. Linguistic profiling of deep- fakes: An open database for next-generation deepfake de- tection.arXiv preprint arXiv:2401.02335,

    Pith/arXiv arXiv 2024

  29. [39]

    Asvspoof: The automatic speaker verification spoofing and countermeasures challenge

    [Wuet al., 2017 ] Zhizheng Wu, Junichi Yamagishi, Tomi Kinnunen, Cemal Hanilc ¸i, Mohammed Sahidullah, Alek- sandr Sizov, Nicholas Evans, Massimiliano Todisco, and Hector Delgado. Asvspoof: The automatic speaker verification spoofing and countermeasures challenge. IEEE Journal of Selected Topics in Signal Processing, 11(4):588–604,

    2017

  30. [40]

    Detect all-type deepfake audio: Wavelet prompt tuning for enhanced auditory perception

    [Xieet al., 2026 ] Yuankun Xie, Ruibo Fu, Xiaopeng Wang, Zhiyong Wang, Songjun Cao, Long Ma, Haonan Cheng, and Long Ye. Detect all-type deepfake audio: Wavelet prompt tuning for enhanced auditory perception. InPro- ceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 35922–35930,

    2026

  31. [41]

    Orthogo- nal subspace decomposition for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633,

    [Yanet al., 2024b ] Zhiyuan Yan, Jiangming Wang, Peng Jin, Ke-Yue Zhang, Chengchun Liu, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Orthogo- nal subspace decomposition for generalizable ai-generated image detection.arXiv preprint arXiv:2411.15633,

    Pith/arXiv arXiv

  32. [42]

    Avoid-df: Audio-visual joint learning for detecting deepfake.IEEE Transactions on Information Forensics and Security, 18:2015–2029,

    [Yanget al., 2023 ] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. Avoid-df: Audio-visual joint learning for detecting deepfake.IEEE Transactions on Information Forensics and Security, 18:2015–2029,

    2023

  33. [43]

    Fine- grained multimodal deepfake classification via heteroge- neous graphs.International Journal of Computer Vision, 132(11):5255–5269,

    [Yinet al., 2024 ] Qilin Yin, Wei Lu, Xiaochun Cao, Xi- angyang Luo, Yicong Zhou, and Jiwu Huang. Fine- grained multimodal deepfake classification via heteroge- neous graphs.International Journal of Computer Vision, 132(11):5255–5269,

    2024

  34. [44]

    Multi-attentional deepfake detection

    [Zhaoet al., 2021 ] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194,

    2021

  35. [45]

    Cross- modality and within-modality regularization for audio- visual deepfake detection

    [Zouet al., 2024 ] Heqing Zou, Meng Shen, Yuchen Hu, Chen Chen, Eng Siong Chng, and Deepu Rajan. Cross- modality and within-modality regularization for audio- visual deepfake detection. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4900–4904. IEEE, 2024

    2024

  36. [2011]

    Towards generalizable deepfake detection via forgery-aware audio-visual adapta- tion: A variational bayesian approach.IEEE Transactions on Information Forensics and Security,

    [Nieet al., 2026 ] Fan Nie, Jiangqun Ni, Jian Zhang, Bin Zhang, Weizhe Zhang, and Bin Li. Towards generalizable deepfake detection via forgery-aware audio-visual adapta- tion: A variational bayesian approach.IEEE Transactions on Information Forensics and Security,

    2026

  37. [2017]

    Mcl: multimodal contrastive learning for deep- fake detection.IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2803–2813,

    [Liuet al., 2023 ] Xiaolong Liu, Yang Yu, Xiaolong Li, and Yao Zhao. Mcl: multimodal contrastive learning for deep- fake detection.IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2803–2813,

    2023

  38. [2018]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3,

    [Huet al., 2022 ] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3,

    2022

  39. [2020]

    Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection

    [Chenet al., 2022 ] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 18710–18719,

    2022

  40. [2021]

    Squeeze- and-excitation networks

    [Huet al., 2018 ] Jie Hu, Li Shen, and Gang Sun. Squeeze- and-excitation networks. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 7132–7141,

    2018

  41. [2022]

    Rawbmamba: End-to-end bidirectional state space model for audio deep- fake detection.arXiv preprint arXiv:2406.06086,

    [Chenet al., 2024 ] Yujie Chen, Jiangyan Yi, Jun Xue, Chen- glong Wang, Xiaohui Zhang, Shunbo Dong, Siding Zeng, Jianhua Tao, Lv Zhao, and Cunhang Fan. Rawbmamba: End-to-end bidirectional state space model for audio deep- fake detection.arXiv preprint arXiv:2406.06086,

    Pith/arXiv arXiv 2024

  42. [2023]

    Unsupervised cross-lingual representation learning for speech recognition.arXiv preprint arXiv:2006.13979,

    [Conneauet al., 2020 ] Alexis Conneau, Alexei Baevski, Ro- nan Collobert, Abdelrahman Mohamed, and Michael Auli. Unsupervised cross-lingual representation learning for speech recognition.arXiv preprint arXiv:2006.13979,

    Pith/arXiv arXiv 2020

  43. [2024]

    De- mamba: Ai-generated video detection on million-scale genvideo benchmark.Science China Information Sci- ences, 69(6):162103,

    [Chenet al., 2026 ] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. De- mamba: Ai-generated video detection on million-scale genvideo benchmark.Science China Information Sci- ences, 69(6):162103,

    2026

  44. [2025]

    Simlbr: Learning to detect fake images by learning to detect real images

    [Dhakalet al., 2026 ] Aayush Dhakal, Subash Khanal, Sriku- mar Sastry, Jacob Arndt, Philipe Dias, Dalton Lunga, and Nathan Jacobs. Simlbr: Learning to detect fake images by learning to detect real images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 35472–35482,

    2026

  45. [2026]

    V oice-face homogeneity tells deepfake.ACM Transactions on Mul- timedia Computing, Communications and Applications, 20(3):1–22,

    [Chenget al., 2023 ] Harry Cheng, Yangyang Guo, Tianyi Wang, Qi Li, Xiaojun Chang, and Liqiang Nie. V oice-face homogeneity tells deepfake.ACM Transactions on Mul- timedia Computing, Communications and Applications, 20(3):1–22,

    2023