Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Reconstruction residuals from a probe trained only on real speech, fused as a gated add-on to a frozen auditory encoder instead of a competing stream, lower the mean in-the-wild deepfake detection error on the ASVspoof 5-to-ITW transfer to

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 15:06 UTC pith:BYJLDEYD

load-bearing objection Careful, honest empirical paper whose headline complementarity claim is not actually tested by its own experiments; the all-seed auxiliary-supervision interaction is the strongest finding. the 3 major comments →

arxiv 2607.26472 v1 pith:BYJLDEYD submitted 2026-07-29 cs.SD cs.AI

Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection

Haotian Mo , Jie Liu , Siqi Shen , Songzhu Mei , Xinhai Chen , Xiangyang Wang , Yigui Feng , Shuai Li
show 3 more authors
Gencheng Liu Keqi Yang Qinglin Wang
This is my paper
classification cs.SD cs.AI
keywords audio deepfake detectionreconstruction residualdiffusion transformer probeaudio-anchored fusioncross-domain generalizationASVspoof 5WavLMprototype scoring
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.

This paper tries to establish that explicit reconstruction residuals from a generative model trained only on bona fide speech can improve cross-domain audio deepfake detection, but only when they are fused as a gated correction to a strong auditory representation rather than as a rival evidence stream. The authors train a diffusion transformer to reconstruct masked spectrograms of real speech, freeze it, and retain absolute residual maps at three masking ratios. Their audio-anchored fusion passes the frozen WavLM-Large vector into the fusion sum without attenuation and lets a scalar gate scale only the residual correction. On the in-the-wild ITW corpus the complete model achieves a three-seed mean EER of 15.33 ± 2.07%, below a separately optimized WavLM-ResNet18 reference at 18.27 ± 4.98%, while the same auxiliary supervision that helps this design worsens the competitive-fusion alternative from 18.40% to 25.30% in every seed. If the paper is right, the structure of fusion, correction versus competition, is what determines whether fragile reconstruction evidence transfers across domains.

Core claim

The central claim is that domain-sensitive reconstruction evidence should enter a deepfake detector as an additive correction anchored to a stable auditory representation, never as a competing stream whose weight trades off against the auditory branch. The authors build a frozen Diffusion Transformer probe trained exclusively on 2.55 million bona fide utterances, reconstruct 1,024-frame Mel spectrograms at masking ratios 0.5, 0.75, and 0.9, and retain the absolute multi-ratio residual maps. A modified ResNet-18 encodes the residuals; frozen WavLM-Large encodes the waveform; fusion is z_f = z_a_tilde + g z_v_tilde, where the gate g in (0,1) scales only the residual term and cannot down-weight

What carries the argument

The load-bearing object is the 'audio-anchored additive fusion' rule: after L2-normalized projections, the fused vector is z_f = z_a_tilde + g z_v_tilde, with the scalar gate g = sigma(w_g^T [z_v_tilde; z_a_tilde] + b_g) applied only to the residual branch. The WavLM-Large auditory anchor enters the pre-projection sum with coefficient one, so the gate can add correction but cannot explicitly suppress the auditory stream. The companion evidence is the cached multi-ratio residual stack V = [X, R_0.5, R_0.75, R_0.9] produced by the frozen bona-fide-trained flow-matching DiT probe, where each R_r = |X - X_hat_r| is the absolute masked-region reconstruction mismatch. The asymmetry of the fusion i

Load-bearing premise

The whole framing rests on the premise that masked-region DiT reconstruction residuals at three fixed ratios carry a forensic signal that transfers from the ASVspoof 5 domain to ITW when combined with frozen WavLM features, a premise tested only through a system-level contrast whose reference differs in learning rate, sampling, checkpoint protocol, and temporal support, and whose ITW mean difference comes from three seeds with overlapping standard deviations.

What would settle it

Run the exactly matched pair of experiments the paper identifies as missing: audio-anchored fusion versus dynamic competitive fusion with identical residual routing, visual encoder, and all other settings fixed, on the same three seeds and ITW protocol. If the anchored rule does not beat competitive fusion at least as often as reported, or if gate statistics show g barely moves on ITW samples, the anchoring claim collapses to a protocol or routing artifact. A second decisive check is testing on any additional target corpus beyond ITW: if the residual correction raises EER there, the transfer c

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim is right, reconstruction-based forensic evidence should be fused with a strong auditory encoder as a gated additive correction, not as a competing stream whose weight inversely scales the auditory one.
  • Multi-ratio residual inputs (0.5/0.75/0.9) beat every single-ratio input on the unseen-attack Eval set even though single ratios can win on Dev, so ratio selection and related settings should be judged on held-out attack families rather than Dev EER.
  • Source auxiliary supervision (attack-ID) interacts with fusion structure: it helps the anchored design on ITW (17.13% to 15.33% mean EER) and hurts the competitive design in every seed (18.40% to 25.30%).
  • Keeping absolute residual magnitude at the input (no InstanceNorm) preserves forensic energy relations; ablations show input normalization consistently degrades Eval performance.
  • With a fixed, uncalibrated bona fide prototype from the training set, the system transfers to ITW without any target-domain center update or threshold fitting, the protocol used for all reported scores.

Where Pith is reading between the lines

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

  • The paper does not report the values of the gate g; inspecting gate statistics by domain would show whether the anchoring advantage is a genuine fusion effect or a routing no-op, since a gate that barely moves on ITW samples would mean the correction term is nearly inert.
  • The non-competitive fusion principle is stated for DiT residuals but likely generalizes to any domain-sensitive auxiliary evidence, such as other generative priors or anomaly scores; a swap test with a different frozen probe would reveal whether the fusion rule rather than the probe's power is the active ingredient.
  • The detector uses only 4 seconds of waveform and about 10.24 seconds of spectrogram, with WavLM support being the shorter; whole-utterance inference is a plausible immediate improvement and the paper itself lists it as a priority.
  • System rankings on Eval do not reproduce on ITW (the competitive system has the best primary-run Eval EER but the worst ITW behavior), suggesting Eval improvement is a weak proxy for in-the-wild robustness and that additional target corpora are needed before treating the ITW gain as a general law.

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

3 major / 5 minor

Summary. The paper proposes a cross-domain audio deepfake detector that augments a frozen WavLM-Large auditory stream with multi-ratio reconstruction residuals from a Diffusion Transformer (DiT) trained only on bona fide speech. The residuals are computed at masking ratios 0.5, 0.75, and 0.9, concatenated with the original Mel spectrogram, and encoded by a spectro-temporal ResNet. An 'audio-anchored' fusion rule adds a scalar-gated residual vector to the auditory vector, preventing explicit down-weighting of the auditory branch. Training uses batch-hard triplet loss with an optional attack-ID auxiliary head; inference is prototype-based against the ASVspoof 5 Train bona fide center. The paper reports three-seed means on ASVspoof 5 Eval and ITW Full, comparing against a separately optimized WavLM-ResNet18 reference and a dynamic competitive fusion variant. The central claim is that reconstruction residuals serve as complementary evidence for ASVspoof 5-to-ITW transfer, with the strongest evidence said to be the all-seed interaction between auxiliary supervision and fusion structure.

Significance. If the result holds, the paper offers a plausible and relatively careful case that a bona-fide-only DiT residual stream can be added to a frozen SSL branch through an asymmetric gate, improving cross-corpus transfer at the system level. The paper's strengths are real: it fixes seed 42 before repeated-seed validation, reports three-seed means and standard deviations without claiming significance, discloses protocol differences between the residual systems and the single-stream reference, caches residuals so all residual-based systems share the identical input, and labels post-hoc diagnostics as descriptive. The contribution is, however, currently a system-level demonstration, not a validated demonstration of residual complementarity, because no matched auditory-only baseline exists under the shared protocol. The paper is transparent about this limitation in the Discussion, but the Abstract and Conclusion reach beyond what the experiments establish.

major comments (3)
  1. [Results and Analysis, Table 3 and Eq. (8)] The central claim that 'reconstruction residuals are complementary evidence' is not directly tested. The only no-residual comparator is the separately optimized WavLM-ResNet18 reference, which differs in learning rate (1e-4 vs 1e-3), batch sampling (class-balanced vs attack-balanced), Dev checkpoint-selection protocol (random crops with Dev center vs leading crops with Train prototype), and temporal support (4s waveform only vs 10.24s spectrogram-residual input). The residual-based rows all include the residual stream and the original Mel, so Table 3 cannot attribute the 15.33% vs 18.27% ITW gap to residuals. The internal audio-anchored vs dynamic-competitive comparison (Eq. (9)-(10)) also changes routing and visual organization, so it does not isolate residual contribution. To support the abstract's claim, the manuscript needs a matched auditory-only branch trained under the identical p
  2. [Abstract and Discussion, 'Evaluation scope'] The Abstract states 'The results support reconstruction residuals as complementary evidence,' but the Discussion's 'Evaluation scope' explicitly says 'Missing componentwise fusion controls, matched reconstructors, whole-utterance tests, and additional targets define the claim as system level.' This is a direct tension. The manuscript should either soften the Abstract to state that a system-level comparison supports the motivation for a non-competitive auditory path, or provide the missing matched no-residual control. As written, the headline claim is stronger than the evidence and inconsistent with the manuscript's own limitation statement.
  3. [Table 5 and 'Repeated-seed robustness'] The three-seed ITW mean comparison (15.3328±2.0719% vs 18.2738±4.9847%) has overlapping standard deviations, and the paired comparison is not in the proposed method's favor for seed 2026. The paper explicitly disclaims significance, which is appropriate, but the Abstract and Conclusion still present the lower mean as the outcome supporting the approach. With n=3, a paired per-seed presentation of the difference and, ideally, a matched-baseline difference would be needed before the 'complementary evidence' framing is justified. The all-seed interaction with auxiliary supervision is a legitimate observation, but it concerns the fusion rule given that residuals are present; it does not by itself show that residuals help at all.
minor comments (5)
  1. [Method, Eq. (4)] The text says 'Equation (4) measures masked-region mismatch,' but Eq. (4) defines the absolute difference over the full tensor. The explanation that visible cells are copied from X and therefore contribute zero is given in the text and Appendix A, but the main-method phrasing could be clarified to avoid implying the equation itself masks the computation.
  2. [Table 4 caption] The table compares single-ratio and multi-ratio inputs, but the caption does not specify that all rows use the shared protocol and the same auxiliary supervision setting. Please state this explicitly so readers do not infer a cross-configuration comparison.
  3. [Experimental Setup, 'Strong single-stream reference'] The difference between 'class-balanced' (reference) and 'attack-balanced' (residual systems) batch sampling is central to the protocol comparison but is defined only informally. A one-sentence definition of each would help.
  4. [Discussion, 'Auxiliary-label scope'] The 33-output head with positive targets only for bona fide and A01-A08, while A09-A32 indices exist in the softmax denominator, is unusual. The main text explains it, but a reader may wonder whether the unused indices affect gradient behavior; a brief note that they act purely as constant negative classes would be useful.
  5. [Figure 2] The figure is dense. The gate estimator input is shown as concatenated [z_v; z_a], but the fusion equation (Eq. (7)) uses normalized projections. Consider labeling the normalization and projection steps explicitly in the figure.

Circularity Check

0 steps flagged

No significant circularity: the residual probe is bona-fide-trained and frozen, the detector is trained on ASVspoof5 Train with conventional prototype scoring, and the acknowledged unmatched reference is an external-validity limitation, not a definitional reduction.

full rationale

The paper's load-bearing derivation is the proposal that explicit multi-ratio DiT reconstruction residuals, computed by Eq. (4) from a frozen bona-fide-only probe, can be fused with WavLM features as a gated additive correction (Eqs. 7-8). This is not circular: the probe training objective (Eq. 3) uses only bona fide spectrograms and no spoof labels or attack identities; the residuals are then cached and consumed by a separately trained supervised detector on ASVspoof5 Train; scoring uses the Train bona fide prototype (Eq. 13) and Eval/ITW are read-only. No parameter is fitted to the target-domain labels, and no result asserted as a prediction is equal by construction to an input. The one comparison that could have been circular—claiming residual complementarity from the WavLM–ResNet18 reference—is explicitly disclaimed: 'It is therefore a strong reference, not a one-branch ablation,' and the paper states 'The comparison with the single-stream reference therefore does not isolate temporal coverage' and 'Missing componentwise fusion controls... define the claim as system level.' Those are limitations in experimental control (the residual contribution is not isolated under a matched no-residual baseline), not circularity: the gap could be caused by protocol/backend differences, but that does not mean the claimed result is entailed by its inputs. The internal audio-anchored vs dynamic-competitive contrast shares data, cache, auditory branch, objectives, and optimizer, and the routing differences are disclosed; it supports an interaction, not a componentwise causal claim. The post-hoc subband diagnostics are explicitly after model choices and not used for selection. Self-citations to prior reconstruction work (e.g., Wang et al. 2024b) are contextual and not load-bearing. Therefore no circular step can be quoted, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No invented entities. The central result is an empirical architecture; the free parameters are hand-chosen hyperparameters that could affect conclusions (masking ratios, loss weights, LR). The key domain assumptions are the transferability of the DiT residual and the sufficiency of the prototype scoring.

free parameters (4)
  • Masking ratios R={0.5,0.75,0.9} = 0.5, 0.75, 0.9
    Hand-selected reconstruction difficulties; no systematic sweep is reported. The multi-ratio choice is motivated by single-ratio Dev/Eval reversal but was not tuned on a held-out set.
  • Aux loss weight lambda_aux = 0.2
    Fixed value with label smoothing 0.1; paper notes no weight sweep, so findings are specific to this setting.
  • Triplet margin m = 0.3
    Standard margin, chosen without reported tuning.
  • Learning rate / optimizer settings = 1e-3, AdamW, 5 epochs, cosine decay
    Chosen for residual systems; the reference uses 1e-4, so not matched.
axioms (4)
  • domain assumption A DiT trained only on bona fide speech produces residual maps that are informative about synthetic speech across domains
    The whole method depends on this; the evidence is an Eval single-ratio/multi-ratio comparison and system-level ITW results, not a direct cross-corpus stability analysis.
  • domain assumption Frozen WavLM-Large features are a robust and complementary auditory anchor
    The reference baseline also uses WavLM, but the complementary benefit of residuals plus WavLM is not measured against a matched single-stream control.
  • domain assumption ASVspoof5 Train bona fide prototype scoring is a fair uncalibrated ranking for ITW
    The paper states scoring is uncalibrated and no target-domain adaptation is performed, so the ITW numbers depend on cross-corpus embedding alignment.
  • standard math Conditional flow matching as implemented (F5-TTS backbone, 16-step Euler) approximates the true flow
    The DiT training uses standard flow-matching losses; the paper does not derive a new result, just applies the known framework.

pith-pipeline@v1.3.0-daily-deepseek · 22469 in / 11989 out tokens · 123357 ms · 2026-08-01T15:06:00.114092+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection." pith.science (2026). https://pith.science/paper/BYJLDEYD

@misc{pith2026260726472,
  author       = {Pith},
  title        = {Pith review of: Audio-Anchored Fusion of Multi-Ratio DiT Reconstruction Residuals for Cross-Domain Audio Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BYJLDEYD}},
  note         = {Machine review of arXiv:2607.26472}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Audio deepfake detectors often degrade when generators, corpora, or recording conditions change. We use a Diffusion Transformer (DiT), trained only on bona fide speech, as a frozen reconstruction probe. Reconstructions at masking ratios 0.5, 0.75, and 0.9 yield explicit multi-ratio residual maps. Because these residuals are domain sensitive, our audio-anchored detector passes the projected frozen-WavLM auditory representation into the fusion sum without gate-based attenuation and uses residuals only as a scalar-gated additive correction. The pre-specified seed-42 run obtains 6.5442% EER / 0.18456 min-DCF on ASVspoof 5 Eval and 13.8372% / 0.36921 on ITW Full; three-seed means are 6.8885 (0.3308)% and 15.3328 (2.0719)%. The latter is below a separately optimized WavLM-ResNet18 reference under both supervision settings. Auxiliary supervision raises dynamic competitive fusion from 18.4007% to 25.2968% mean ITW EER, worsening all three seeds. The results support reconstruction residuals as complementary evidence and motivate a non-competitive auditory path for ASVspoof 5-to-ITW transfer, without claiming a componentwise causal ablation of anchoring alone.

Figures

Figures reproduced from arXiv: 2607.26472 by Gencheng Liu, Haotian Mo, Jie Liu, Keqi Yang, Qinglin Wang, Shuai Li, Siqi Shen, Songzhu Mei, Xiangyang Wang, Xinhai Chen, Yigui Feng.

Figure 1
Figure 1. Figure 1: Mean DiT reconstruction residual spectra on the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Audio-anchored dual-stream detector. The scalar-gated spectrogram–residual representation is added to the frozen [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Masked self-supervised training of the DiT recon [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustrative t-SNE of 1,000 ASVspoof 5 Dev embed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Post-hoc per-attack subband EER diagnostic for [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] 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

37 extracted references · 7 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2408.08739 , year =

    Wang, Xin and Delgado, Hector and Tak, Hemlata and others , title =. arXiv preprint arXiv:2408.08739 , year =

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2601.03944 , year =

    Wang, Xin and Delgado, Hector and Evans, Nicholas and others , title =. arXiv preprint arXiv:2601.03944 , year =

    Pith/arXiv arXiv

  3. [3]

    Proceedings of ICASSP , pages =

    Tak, Hemlata and Patino, Jose and Todisco, Massimiliano and others , title =. Proceedings of ICASSP , pages =

  4. [4]

    Proceedings of ICASSP , pages =

    Jung, Jee-Weon and Heo, Hee-Soo and Tak, Hemlata and others , title =. Proceedings of ICASSP , pages =

  5. [5]

    IEEE Journal of Selected Topics in Signal Processing , volume =

    Chen, Sanyuan and Wang, Chengyi and Chen, Zhengyang and others , title =. IEEE Journal of Selected Topics in Signal Processing , volume =

  6. [6]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =

    Hsu, Wei-Ning and Bolte, Benjamin and Tsai, Yao-Hung Hubert and others , title =. IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =

  7. [7]

    Does Audio Deepfake Detection Generalize? , booktitle =

    M. Does Audio Deepfake Detection Generalize? , booktitle =

  8. [8]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =

    Liu, Xuechen and Wang, Xin and Sahidullah, Md and others , title =. IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =

  9. [9]

    Proceedings of Interspeech , pages =

    Wang, Xin and Fu, Ruibo and Wen, Zhengqi and others , title =. Proceedings of Interspeech , pages =

  10. [10]

    Proceedings of ACL , pages =

    Chen, Yushen and Niu, Zhikang and Ma, Ziyang and others , title =. Proceedings of ACL , pages =

  11. [11]

    Advances in Neural Information Processing Systems , year =

    Huang, Po-Yao and Xu, Hu and Li, Juncheng and others , title =. Advances in Neural Information Processing Systems , year =

  12. [12]

    Proceedings of CVPR , pages =

    He, Kaiming and Chen, Xinlei and Xie, Saining and others , title =. Proceedings of CVPR , pages =

  13. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Zhang, Kai and Hua, Zhen and Lan, Rushi and others , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =

  14. [14]

    arXiv preprint arXiv:2409.11909 , year =

    Wang, Ziyang and Fu, Ruibo and Wen, Zhengqi and others , title =. arXiv preprint arXiv:2409.11909 , year =

    Pith/arXiv arXiv

  15. [15]

    IEEE Signal Processing Letters , volume =

    Zhang, You and Jiang, Fei and Duan, Zhiyao , title =. IEEE Signal Processing Letters , volume =

  16. [16]

    Proceedings of Interspeech , pages =

    Pascu, Oana and Stan, Adriana and Oneata, Dan and Oneata, Elisabeta and Cucu, Horia , title =. Proceedings of Interspeech , pages =

  17. [17]

    Advances in Neural Information Processing Systems , volume =

    Baevski, Alexei and Zhou, Yuhao and Mohamed, Abdelrahman and Auli, Michael , title =. Advances in Neural Information Processing Systems , volume =

  18. [18]

    arXiv preprint arXiv:1904.05441 , year =

    Todisco, Massimiliano and Wang, Xin and Vestman, Ville and others , title =. arXiv preprint arXiv:1904.05441 , year =

    Pith/arXiv arXiv 1904

  19. [19]

    arXiv preprint arXiv:1912.06670 , year =

    Ardila, Rosana and Branson, Megan and Davis, Kelly and others , title =. arXiv preprint arXiv:1912.06670 , year =

    Pith/arXiv arXiv 1912

  20. [20]

    Proceedings of ICASSP , pages =

    Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev , title =. Proceedings of ICASSP , pages =

  21. [21]

    Proceedings of Interspeech , pages =

    Nagrani, Arsha and Chung, Joon Son and Zisserman, Andrew , title =. Proceedings of Interspeech , pages =

  22. [22]

    Proceedings of ICML , pages =

    Kim, Jaehyeon and Kong, Jungil and Son, Juhee , title =. Proceedings of ICML , pages =

  23. [23]

    arXiv preprint arXiv:2009.09761 , year =

    Kong, Zhifeng and Ping, Wei and Huang, Jiaji and Zhao, Kexin and Catanzaro, Bryan , title =. arXiv preprint arXiv:2009.09761 , year =

    Pith/arXiv arXiv 2009

  24. [24]

    Proceedings of ICML , pages =

    Popov, Vadim and Vovk, Ivan and Gogoryan, Vladimir and Sadekova, Tasnima and Kudinov, Mikhail , title =. Proceedings of ICML , pages =

  25. [25]

    Proceedings of ICASSP , pages =

    Deng, Jiang and Ren, Yong and Zhang, Tong and Zhu, Haibin and Sun, Zhenan , title =. Proceedings of ICASSP , pages =

  26. [26]

    Proceedings of the ASVspoof Workshop , pages =

    Chan, Po-Chun and Chen, Wei-Yu and Wang, Jhing-Fa , title =. Proceedings of the ASVspoof Workshop , pages =

  27. [27]

    Proceedings of ACM Multimedia , pages =

    Gu, Haowen and Yi, Jiangyan and Wang, Chenglong and others , title =. Proceedings of ACM Multimedia , pages =

  28. [28]

    IEEE Open Journal of Signal Processing , year =

    Jung, Jee-Weon and Wu, Yuxuan and Wang, Xin and others , title =. IEEE Open Journal of Signal Processing , year =

  29. [29]

    Leveraging Frequency Analysis for Deep Fake Image Recognition , booktitle =

    Frank, Joel and Eisenhofer, Thorsten and Sch. Leveraging Frequency Analysis for Deep Fake Image Recognition , booktitle =

  30. [30]

    Advances in Neural Information Processing Systems , year =

    Liang, Yuchen and Yu, Ming and Li, Guangyuan and others , title =. Advances in Neural Information Processing Systems , year =

  31. [31]

    Proceedings of EMNLP , pages =

    Li, Yiming and Zhang, Ming and Ren, Meng and others , title =. Proceedings of EMNLP , pages =. 2024 , doi =

    2024

  32. [32]

    Proceedings of the ASVspoof Workshop , year =

    Zhu, Yuxin and Goel, Chinmay and Koppisetti, Sree Harsha and others , title =. Proceedings of the ASVspoof Workshop , year =

  33. [33]

    Robust Audio Deepfake Detection: Exploring Front-/Back-End Combinations and Data Augmentation Strategies for the

    Sch. Robust Audio Deepfake Detection: Exploring Front-/Back-End Combinations and Data Augmentation Strategies for the. Proceedings of the ASVspoof Workshop , year =

  34. [34]

    Applied Sciences , volume =

    Tahaoglu, Gokhan , title =. Applied Sciences , volume =. 2025 , doi =

    2025

  35. [35]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =

    Automated Data Augmentation for Audio Classification , author =. IEEE/ACM Transactions on Audio, Speech, and Language Processing , volume =. 2024 , doi =

    2024

  36. [36]

    arXiv preprint arXiv:2604.26465 , year =

    Cheng, Bo and Cao, Songjun and Zhang, Xiaoming and Chen, Jie and Ma, Long and Chen, Fei , title =. arXiv preprint arXiv:2604.26465 , year =

    Pith/arXiv arXiv

  37. [37]

    Proceedings of Interspeech , pages =

    Grinberg, Petr and Kumar, Ankur and Koppisetti, Surya and Bharaj, Gaurav , title =. Proceedings of Interspeech , pages =. 2025 , doi =

    2025