Pith. sign in

REVIEW 5 major objections 5 minor 41 references

Cross-Modal Watermarking for Authentic Audio Recovery and Tamper Localization in Synthesized Audiovisual Forgeries

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hidden audio watermark in video frames can restore authentic speech after voice-cloning attacks.

desk verdict Genuinely new task and a credible pipeline, but the semantic recovery headline is underevidenced: SNR/PESQ alone do not show the words came back. read the letter →

arxiv 2507.12723 v1 pith:GW25QDOT submitted 2025-07-17 cs.SD cs.MMeess.AS

Minyoung Kim , Sehwan Park , Sungmin Cha , Paul Hongsuck Seo This is my paper
classification cs.SDcs.MMeess.AS
keywords authenticaudiorecoverytamperlocalizationcross-modalwatermarkinginvertibleneuralnetworkvoicecloninglipsynchronizationsynthesizedaudiovisualforgeriesdeepfakedefense
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

This paper proposes that the original audio of a speech video can be restored after an attacker has cloned the voice and re-synchronized the lips, as long as that audio was invisibly embedded into the visual frames before the attack. It names this task Authentic Audio Recovery and adds Tamper Localization in Audio, the ability to mark which time intervals of the audio were altered. The proposed scheme folds the spectrogram of the authentic audio into the image with a stack of invertible neural network blocks; at test time a learned estimator reconstructs the missing intermediate representation from the tampered frame, and the inverse blocks recover the speech. Tampered segments are found by comparing the recovered and attacked audio in a learned semantic feature space instead of sample by sample. If the claim holds, media forensics can move from flagging fake content to restoring the speaker's actual message and showing exactly what was changed.

What carries the argument

The load-bearing object is the cross-modal watermarking (CMW) transform built from invertible neural network blocks. Each block maps a paired image feature and audio spectrogram through reversible affine coupling: $I^{l+1}=I^l+\phi(A^l)$ and $A^{l+1}=A^l\odot\exp(\sigma(\rho(I^{l+1})))+\eta(I^{l+1})$, with the inversion obtained algebraically without inverting the networks $\phi$, $\rho$, $\eta$. The image side is processed in the discrete-wavelet domain and the audio side as a reshaped short-time Fourier spectrogram, so the two modalities share a common tensor shape. Because the intermediate audio component $A^{L+1}_{\mathrm{org}}$ is discarded in the watermarked video, a learned noise estimator predicts it from the watermarked frame, and the inverse blocks then recover $A_{\mathrm{org}}$ from the tampered frame. Tamper localization is carried by a separate semantic feature extractor that projects recovered and tampered audio into a shared space where cosine similarity marks altered time steps.

What would settle it

Take watermarked videos, replace the entire facial region with a different generated face rather than only re-synchronizing lips, and measure recovery SNR and intelligibility; if the metrics collapse toward the no-mask baseline (SNR about 4.63) or transcripts become wrong, the central claim fails for the stronger visual attack. Alternatively, run automatic speech recognition on recovered audio from 5-second clips and compare word error rate to the original clips, since the paper reports similarity metrics but no transcript-level measure.

Watch

Extended reading notes

Core claim

The central discovery is that cross-modal watermarking converts a detection problem into a recovery problem. By embedding the authentic audio into the visual frame rather than into the audio stream, the method preserves audio integrity (the watermarked video keeps the original audio unchanged) and uses the visual redundancy to survive audio replacement. At evaluation time, with both a voice-cloning attack and a lip-synchronization re-rendering applied, the model reports recovered-audio quality of SNR 17.82 and PESQ 3.18, and tamper-localization IoU above 95 for voice-cloned intervals, while all audio-only baselines can detect or localize but cannot recover the original speech. The paper also reports that the watermark survives several different lip-synchronization pipelines and that training on non-human video and music data yields comparable performance, which would lower the privacy cost of building such a defense.

Load-bearing premise

The premise that everything rests on is that after lip-sync re-rendering the image still contains enough of the embedded watermark for the decoder to reconstruct the speech, and that the video was watermarked before the manipulation began.

Editorial extensions

If this is right

  • Because the authentic audio is carried in the visual channel, the original speech can be recovered even if the tampered audio stream is completely replaced; audio-only watermarking schemes cannot do this.
  • The reported tamper-localization accuracy (average precision above 98 in the voice-cloning setting) suggests that comparing recovered and tampered audio in semantic feature space is sufficient to mark altered intervals precisely enough for downstream use.
  • The method degrades gracefully across different lip-synchronization methods while staying above 90 in intersection-over-union, so it is not tuned to a single forgery pipeline.
  • Because non-human data gives comparable results, a deployment does not need to collect faces or voices from real people, which addresses the privacy and portrait-rights obstacle to forensic training data.
  • Watermarking is imperceptible (PSNR 41.53, SSIM 0.98), which matters for practical adoption because the defensive layer does not visibly degrade the content being protected.

Reading between the lines

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

  • The paper defines Authentic Audio Recovery only for videos that were watermarked by this system before manipulation; a reader should not infer that unwatermarked legacy footage can be restored. An open direction is recovering authentic audio from arbitrary existing forgeries, which would require a different source of ground truth.
  • If the visual watermark survives lip-sync re-rendering because facial masking mimics realistic removal during training, then attacks that fully replace the face rather than moving the lips should degrade recovery markedly; that boundary can be tested directly.
  • The same visual-channel capacity could carry additional provenance metadata, such as capture time or device identity, through the same tampering pipeline, turning the watermark from an audio-recovery tool into a general integrity certificate.
  • Feeding the recovered audio to automatic speech recognition would provide an end-to-end check of semantic preservation; the paper reports SNR and PESQ but not word error rate, so an ASR evaluation is a natural extension.
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

5 major / 5 minor

Summary. The paper introduces Authentic Audio Recovery (AAR) and Tamper Localization in Audio (TLA) for Synthesized Audiovisual Forgeries (SAVFs), where both audio and video are manipulated by voice cloning and lip synchronization. The proposed method, Cross-Modal Watermarking (CMW), embeds the authentic audio into the visual frame using invertible neural network blocks and a noise estimator that predicts the discarded transformed audio from the watermarked image. At test time, the recovered audio is compared with the tampered audio in a semantic feature space to localize tampered intervals. The authors report strong localization performance (IoU above 95 for voice cloning) and audio recovery quality (SNR 17.82, PESQ 3.18) on the HDTF dataset, and they additionally show that the method can be trained on non-human datasets without much degradation.

Significance. If the central claim holds, the paper opens a useful direction: moving beyond tamper detection and localization to actual recovery of the original speech semantics in manipulated audiovisual content. The method is clearly specified, the experimental protocol is mostly transparent, and the code is promised to be public. The paper also provides evidence of domain generalization by training on non-human data, which is a practical strength for privacy-sensitive deployment. The main value is the proposed task formulation and the cross-modal embedding design; these are likely to be of interest to the multimedia forensics and watermarking community.

major comments (5)
  1. [Sec. 4.1, Table 1] The central claim that recovered audio 'directly conveys the semantic content of the message' (Sec. 1) is not supported by the evaluation. The AAR metrics are SNR and PESQ only, which measure waveform fidelity and perceptual quality, not word-level intelligibility or semantic correctness. For a voice-cloning SAVF, the tampered audio says different words, so the task is to recover the original words; a signal with correct pitch/energy but wrong phonemes could score well on PESQ while failing the task. The paper should report word error rate (WER) or character error rate (CER) against the original transcript, or a human transcription study, to establish semantic recovery.
  2. [Sec. 3.3, Table 3] The statement that the invertible property allows 'perfectly recover the original inputs' is overstated. Exact inversion requires A^{L+1}_org, which is discarded during embedding; in the actual pipeline A^{L+1}_org is estimated from I_wm by a noise estimator, and I_wm itself may be altered by lip-sync forgery. Table 3 shows SNR dropping from 28.20 without lip-sync to 16.06 with Wav2Lip, confirming that recovery is lossy under the intended operating conditions. The text should be revised to distinguish conditional invertibility (exact only when A^{L+1}_org is available) from the approximate recovery actually achieved.
  3. [Table 1, Sec. 4.2] All reported numbers are single runs without standard deviations or significance tests. Given the modest margin over AudioSeal in the voice-cloning TLA results (IoU 95.40 vs 91.78), the claim of consistent superiority needs error bars or multiple-seed evaluations. Additionally, ThinImg [23], cited in Related Work as the closest cross-modal steganography prior (hiding audio in images), is not included as a baseline; either it should be evaluated or the exclusion should be explicitly justified.
  4. [Sec. 1, Sec. 5] The framing as a general 'robust defense against misinformation' overclaims the scope. AAR is only defined for videos that were watermarked at creation by this specific system; for arbitrary, unwatermarked SAVFs the method cannot recover any authentic audio. This is an inherent limitation of the proposed task and should be stated prominently in the abstract and conclusion, so that the claim is not read as a universal solution to audiovisual misinformation.
  5. [Table 2, Sec. 4.2] The masking strategy is essential to the method's robustness, but its hyperparameters (number of masks, side-length range 20-150, use of LightFace for facial masks) are described only briefly and no sensitivity analysis is provided. Since Table 2 shows that training without masking fails catastrophically (SNR 4.63), the choice of masking parameters is load-bearing. The authors should report the effect of mask size and coverage, or at least justify the chosen values.
minor comments (5)
  1. [Sec. 4.2] The text says tampering uses 'AS-inserting... and VS-modifying...', but Table 1 uses 'AS' and 'VC'; 'VS' appears to be a typo for 'VC' (voice cloning) and should be corrected.
  2. [Abstract and throughout] The term 'SA VFs' is inconsistently spaced; it should be 'SAVFs' for readability.
  3. [Fig. 1] The notation 'M' in the masking operation is introduced in Sec. 3.4 but the figure shows 'M' and '1-M' without a definition; a brief caption note would help.
  4. [Sec. 3.4] The phrase 'unsupervised manner' is imprecise since the training uses reconstruction and contrastive losses; 'self-supervised' or 'without tamper-localization annotations' would be more accurate.
  5. [References] Reference [24] renders the venue as 'cvpr' lowercase, and other references have inconsistent capitalization; a final formatting pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the watermark recovery is an end-to-end learned decode task evaluated on held-out data, with no load-bearing self-citation or construction-level reduction.

full rationale

The paper's central derivation is a trained cross-modal watermark encoder/decoder, not a parameter-free first-principles claim. At training time the original audio A_org is available to supervise the noise estimator (Sec. 3.4, L_ARL = ||A_org - A_rec||^2_2), and at test time the same audio is used only as ground truth for evaluation (Sec. 4.1). The INN inversion equations (3)-(4) indeed give exact recovery of A_org only when A^{L+1}_org is known; the actual AAR pipeline estimates A^{L+1}_org from the watermarked visual frame, as the paper explicitly states in Sec. 3.3. Thus AAR is a learned decoding task on held-out HDTF videos, not an identity that makes the result true by construction. Tamper localization compares the recovered audio with the tampered audio; this is narrower than forensic recovery from arbitrary unwatermarked forgeries, but narrowing scope is not circularity. The references to prior INN steganography [22], feature extractors [25], and contrastive loss [26] are external; there are no load-bearing self-citations. The one flagged weakness, the Table 3 statement that 'speech remains intelligible (refer to Supplementary Material)' without reporting WER or a similar intelligibility metric, is a missing-support or validity concern about the semantic-content claim, not a circularity. Overall, no circular step is identifiable from the paper's equations or argumentation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The paper's dependence is on channel assumptions (capacity, survivability) plus hand-set hyperparameters, not on fitted physical constants or invented forces.

free parameters (4)
  • Training loss weights = lambda_WL=10, lambda_ARL=10, lambda_VRL=0.1, lambda_SFCL=1
    Hand-set coefficients controlling watermark invisibility, audio reconstruction, visual reconstruction, and feature contrast; reported once in Sec. 4.1 with no sensitivity analysis.
  • Masking strategy hyperparameters = 1 to 3 masks, side lengths 20 to 150 pixels
    Chosen by hand in Sec. 3.4 to simulate watermark removal; Table 2 shows the choice strongly affects both AAR and TLA, so robustness claims are tied to these values.
  • STFT analysis parameters = window 510, hop 128
    Selected to map the audio segment into the H/4 x W/4 x 1 spectrogram shape required by the INN blocks in Sec. 3.2; changing them changes embedding capacity and fidelity.
  • Number of INN blocks = L=6
    Architecture depth chosen by hand with no ablation study; the invertibility property holds for any L, but capacity and reconstruction fidelity depend on this value.
assumptions (5)
  • standard math Affine coupling blocks with any phi, rho, eta are invertible via Eq. 3-4.
    Used in Sec. 3.2; the inversion only requires shape matching, which holds by construction.
  • standard math DWT and STFT with the chosen parameters are invertible transforms.
    Sec. 3.2 constructs the intermediate representations from these transforms and relies on IDWT to produce the visible frame.
  • domain assumption Visual frames have enough capacity to carry a full audio spectrogram imperceptibly.
    Necessary for the whole watermarking scheme; Table 1's PSNR 41.53 shows the frame is close to original, but capacity limits are not analyzed.
  • domain assumption Lip-sync and voice-cloning attacks leave enough watermark intact for recovery.
    Sec. 3.4's masking strategy simulates this; Table 3 shows degradation as attacks get stronger, so the assumption is partially validated but attack-dependent.
  • domain assumption Semantic feature inner-product similarity reliably separates tampered from untampered audio regions.
    Sec. 3.3 uses a semantic feature extractor from [25] and a contrastive loss; no analysis of failure cases or feature alignment errors is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Modal Watermarking for Authentic Audio Recovery and Tamper Localization in Synthesized Audiovisual Forgeries." pith.science (2026). https://pith.science/paper/GW25QDOT

@misc{pith2026250712723,
  author       = {Pith},
  title        = {Pith review of: Cross-Modal Watermarking for Authentic Audio Recovery and Tamper Localization in Synthesized Audiovisual Forgeries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GW25QDOT}},
  note         = {Machine review of arXiv:2507.12723}
}
read the original abstract

Recent advances in voice cloning and lip synchronization models have enabled Synthesized Audiovisual Forgeries (SAVFs), where both audio and visuals are manipulated to mimic a target speaker. This significantly increases the risk of misinformation by making fake content seem real. To address this issue, existing methods detect or localize manipulations but cannot recover the authentic audio that conveys the semantic content of the message. This limitation reduces their effectiveness in combating audiovisual misinformation. In this work, we introduce the task of Authentic Audio Recovery (AAR) and Tamper Localization in Audio (TLA) from SAVFs and propose a cross-modal watermarking framework to embed authentic audio into visuals before manipulation. This enables AAR, TLA, and a robust defense against misinformation. Extensive experiments demonstrate the strong performance of our method in AAR and TLA against various manipulations, including voice cloning and lip synchronization.

Figures

Figures reproduced from arXiv: 2507.12723 by the authors.

Figure 1
Figure 1. Overall architecture of our model. The framework comprises three main processes: cross-modal watermarking (CMW), authentic audio recovery, and tamper localization. In the CMW process, CMW embed the authentic audio within a visual frame. For authentic audio recovery, noise estimators predict the transformed audio output from watermarked visual frame, enabling the inverse CMW to recover the authentic audio embedded in… view at source ↗
Figure 2
Figure 2. Qualitative Examples The watermarked frames and recovered audio closely resemble the original AVS, ensuring im￾perceptible embedding and authentic audio recovery. performs the best SOTA method, Audioseal, in both settings and surpasses other baselines by a large margin. Unlike Wavmark and Audioseal, which degrade audio quality by embedding wa￾termarks directly in the audio stream, our approach preserves audio integr… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [23]

    Wavmark: Watermarking for audio generation,

    G. Chen, Y . Wu, S. Liu, T. Liu, X. Du, and F. Wei, “Wavmark: Watermarking for audio generation,” 2023

    work page 2023

  2. [1]

    Introduction Recent advancements in generative speech models [1, 2, 3] have enabled the synthesis of high-fidelity audio content that closely resembles real-world speech. Among these, voice cloning tech- niques [4, 5, 6] can replicate a speaker’s unique vocal charac- teristics from just a few audio samples, facilitating personalized content generation. Wh...

    work page

  3. [2]

    Cross-Modal Watermarking for Authentic Audio Recovery and Tamper Localization in Synthesized Audiovisual Forgeries

    Related Works V oice Cloned Audio LocalizationRecent advances in voice cloning have intensified the challenge of localizing manipulated audio segments. Approaches like BAM [14] and CFPRF [15] utilize boundary-aware attention and coarse-to-fine refinement to detect tampering; however, their reliance on specific training manipulations limits robustness agai...

    work page Pith review arXiv 2025

  4. [3]

    Method 3.1. Overview In this work, we particularly focus on SA VFs utilized to ma- nipulate the original message of the speaker by modifying real speech videos through voice cloning and lip synchroniza- tion. Our task prioritizes Authentic Audio Recovery (AAR) and Tamper Localization in Audio (TLA) because speech con- veys core semantic content, serving a...

    work page

  5. [4]

    Semantic Feature Contrastive LossTo ensure robust tam- per localization, we compare the tampered and recovered audio streams in a semantic feature space

    These terms are critical to train the noise estimator forA L+1 org introduced in Section 3.2. Semantic Feature Contrastive LossTo ensure robust tam- per localization, we compare the tampered and recovered audio streams in a semantic feature space. Specifically, we enforce proximity between temporally aligned featuresf t org andf t rec us- ing a contrastiv...

    work page

  6. [5]

    No Mask” refers to training without a mask. Our masking strategies outperform “No Mask

    Experiments 4.1. Experimental Setup DatasetWe use the HDTF dataset [28], which consists of 410 talking face videos with synchronized speech, totaling 16 hours of audiovisual data. As one of the primary benchmarks for lip- synchronization [28, 29], HDTF provides high-quality, diverse speaker recordings, making it well-suited for evaluating the ef- fectiven...

    work page

  7. [6]

    To achieve this, we propose cross-modal watermarking method not only localizing tampered regions but also recover- ing authentic audio

    Conclusion In this paper, we introduced a novel task of recovering authentic audio from SA VFs, moving beyond mere detection and local- ization. To achieve this, we propose cross-modal watermarking method not only localizing tampered regions but also recover- ing authentic audio. Our model demonstrated state-of-the-art localization performance while effec...

    work page

  8. [7]

    Acknowledgements This research was supported by IITP grants (IITP-2025- RS-2020-II201819, IITP-2025-RS-2024-00436857, IITP- 2025-RS-2024-00398115, IITP-2025-RS-2025-02263754, IITP-2025-RS-2025-02304828), and the KOCCA grant (RS- 2024-00345025) funded by the Korea government (MSIT, MOE and MSCT)

    work page 2025

Show all 41 references
  1. [8]

    V oice- box: Text-guided multilingual universal speech generation at scale,

    M. Le, A. Vyas, B. Shi, B. Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokaret al., “V oice- box: Text-guided multilingual universal speech generation at scale,”NeurIPS, vol. 36, 2024

    2024

  2. [9]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” inICML. PMLR, 2021, pp. 5530–5540

    2021

  3. [10]

    Prompttts++: Controlling speaker identity in prompt-based text-to-speech using natural lan- guage descriptions,

    R. Shimizu, R. Yamamoto, M. Kawamura, Y . Shirahata, H. Doi, T. Komatsu, and K. Tachibana, “Prompttts++: Controlling speaker identity in prompt-based text-to-speech using natural lan- guage descriptions,” inICASSP. IEEE, 2024, pp. 12 672–12 676

    2024

  4. [11]

    Openvoice: Versatile instant voice cloning,

    Z. Qin, W. Zhao, X. Yu, and X. Sun, “Openvoice: Versatile instant voice cloning,”arXiv preprint arXiv:2312.01479, 2023

    2023 arXiv

  5. [12]

    Paddlespeech: An easy-to-use all-in-one speech toolkit,

    H. Zhang, T. Yuan, J. Chen, X. Li, R. Zheng, Y . Huang, X. Chen, E. Gong, Z. Chen, X. Huet al., “Paddlespeech: An easy-to-use all-in-one speech toolkit,” inProceedings of the 2022 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Huma...

    2022

  6. [13]

    V oice cloning: a multi-speaker text-to-speech synthesis approach based on trans- fer learning,

    G. Ruggiero, E. Zovato, L. Di Caro, and V . Pollet, “V oice cloning: a multi-speaker text-to-speech synthesis approach based on trans- fer learning,”arXiv preprint arXiv:2102.05630, 2021

    2021 arXiv

  7. [14]

    A lip sync expert is all you need for speech to lip generation in the wild,

    K. R. Prajwal, R. Mukhopadhyay, V . P. Namboodiri, and C. Jawahar, “A lip sync expert is all you need for speech to lip generation in the wild,” inProceedings of the 28th ACM International Conference on Multimedia, ser. MM ’20. New York, NY , USA: Association for Computing Mac...

    2020

  8. [15]

    Diff2lip: Audio conditioned diffusion models for lip- synchronization,

    S. Mukhopadhyay, S. Suri, R. T. Gadde, and A. Shrivas- tava, “Diff2lip: Audio conditioned diffusion models for lip- synchronization,” inWACV, January 2024, pp. 5292–5302

    2024

  9. [16]

    Pose-controllable talking face generation by implicitly modular- ized audio-visual representation,

    H. Zhou, Y . Sun, W. Wu, C. C. Loy, X. Wang, and Z. Liu, “Pose-controllable talking face generation by implicitly modular- ized audio-visual representation,” inCVPR, 2021

    2021

  10. [17]

    Transface: Unit-based audio-visual speech synthesizer for talking head translation,

    X. Cheng, R. Huang, L. Li, T. Jin, Z. Wang, A. Yin, M. Li, X. Duan, Z. Zhaoet al., “Transface: Unit-based audio-visual speech synthesizer for talking head translation,”arXiv preprint arXiv:2312.15197, 2023

    2023 arXiv

  11. [18]

    Synctalklip: Highly synchronized lip-readable speaker generation with multi-task learning,

    X. Yang, X. Cheng, D. Fu, M. Fang, J. Zuo, S. Ji, Z. Zhao, and J. Tao, “Synctalklip: Highly synchronized lip-readable speaker generation with multi-task learning,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 8149– 8158

    2024

  12. [19]

    Editguard: Versatile image watermarking for tamper localization and copy- right protection,

    X. Zhang, R. Li, J. Yu, Y . Xu, W. Li, and J. Zhang, “Editguard: Versatile image watermarking for tamper localization and copy- right protection,” inCVPR, 2024, pp. 11 964–11 974

    2024

  13. [20]

    Proactive detection of voice cloning with localized watermarking,

    R. San Roman, P. Fernandez, H. Elsahar, A. D´efossez, T. Furon, and T. Tran, “Proactive detection of voice cloning with localized watermarking,”ICML, 2024

    2024

  14. [21]

    Enhancing partially spoofed audio localization with boundary-aware attention mechanism,

    J. Zhong, B. Li, and J. Yi, “Enhancing partially spoofed audio localization with boundary-aware attention mechanism,”arXiv preprint arXiv:2407.21611, 2024

    2024 arXiv

  15. [22]

    Coarse- to-fine proposal refinement framework for audio temporal forgery detection and localization,

    J. Wu, W. Lu, X. Luo, R. Yang, Q. Wang, and X. Cao, “Coarse- to-fine proposal refinement framework for audio temporal forgery detection and localization,” inProceedings of the 32nd ACM In- ternational Conference on Multimedia, 2024, pp. 7395–7403

    2024

  16. [24]

    Hiding data in images by simple lsb substitution,

    C.-K. Chan and L. Cheng, “Hiding data in images by simple lsb substitution,”Pattern Recognition, vol. 37, no. 3, pp. 469–474, 2004. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S003132030300284X

    2004

  17. [25]

    Complete video quality-preserving data hiding,

    K. Wong, K. Tanaka, K. Takagi, and Y . Nakajima, “Complete video quality-preserving data hiding,”IEEE Transactions on Cir- cuits and Systems for Video Technology, vol. 19, no. 10, pp. 1499– 1512, 2009

    2009

  18. [26]

    Steganogan: High capacity image steganography with gans,

    K. A. Zhang, A. Cuesta-Infante, and K. Veeramachaneni, “Steganogan: High capacity image steganography with gans,” arXiv preprint arXiv:1901.03892, 2019. [Online]. Available: https://arxiv.org/abs/1901.03892

    1901 arXiv

  19. [27]

    Nice: Non-linear independent components estimation,

    L. Dinh, D. Krueger, and Y . Bengio, “Nice: Non-linear independent components estimation,” 2015. [Online]. Available: https://arxiv.org/abs/1410.8516

    2015 arXiv

  20. [28]

    Hinet: Deep image hiding by invertible network,

    J. Jing, X. Deng, M. Xu, J. Wang, and Z. Guan, “Hinet: Deep image hiding by invertible network,” inICCV, 2021, pp. 4733– 4742

    2021

  21. [29]

    Large-capacity and flexible video steganography via invertible neural network,

    C. Mou, Y . Xu, J. Song, C. Zhao, B. Ghanem, and J. Zhang, “Large-capacity and flexible video steganography via invertible neural network,” inCVPR, 2023, pp. 22 606–22 615

    2023

  22. [30]

    Thinimg: Cross-modal steganography for presenting talking heads in images,

    L. Zhao, H. Li, X. Ning, and X. Jiang, “Thinimg: Cross-modal steganography for presenting talking heads in images,” inWACV, 2024, pp. 5553–5562

    2024

  23. [31]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” incvpr, 2017

    2017

  24. [32]

    Cnn architectures for large-scale audio classification,

    S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seyboldet al., “Cnn architectures for large-scale audio classification,” in2017 ieee international conference on acoustics, speech and signal pro- cessing (icas...

    2017

  25. [33]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” 2019. [Online]. Available: https://arxiv.org/abs/1807.03748

    2019 arXiv

  26. [34]

    Lightface: A hybrid deep face recognition framework,

    S. I. Serengil and A. Ozpinar, “Lightface: A hybrid deep face recognition framework,” in2020 Innovations in Intelligent Systems and Applications Conference (ASYU). IEEE, 2020, pp. 23–27. [Online]. Available: https://ieeexplore.ieee.org/document/ 9259802

    2020

  27. [35]

    Flow-guided one-shot talk- ing face generation with a high-resolution audio-visual dataset,

    Z. Zhang, L. Li, Y . Ding, and C. Fan, “Flow-guided one-shot talk- ing face generation with a high-resolution audio-visual dataset,” inCVPR, 2021, pp. 3661–3670

    2021

  28. [36]

    Lrs3-ted: a large-scale dataset for visual speech recognition,

    T. Afouras, J. S. Chung, and A. Zisserman, “Lrs3-ted: a large-scale dataset for visual speech recognition,”arXiv preprint arXiv:1809.00496, 2018

    2018 arXiv

  29. [37]

    V2a- mark: Versatile deep visual-audio watermarking for manipulation localization and copyright protection,

    X. Zhang, Y . Xu, R. Li, J. Yu, W. Li, Z. Xu, and J. Zhang, “V2a- mark: Versatile deep visual-audio watermarking for manipulation localization and copyright protection,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 9818– 9827

    2024

  30. [38]

    Musetalk: Real-time high quality lip synchorization with latent space inpainting,

    Y . Zhang, M. Liu, Z. Chen, B. Wu, Y . Zeng, C. Zhan, Y . He, J. Huang, and W. Zhou, “Musetalk: Real-time high quality lip synchorization with latent space inpainting,”arxiv, 2024

    2024

  31. [39]

    Video enhancement with task-oriented flow,

    T. Xue, B. Chen, J. Wu, D. Wei, and W. T. Freeman, “Video enhancement with task-oriented flow,”International Journal of Computer Vision, vol. 127, pp. 1106–1125, 2019

    2019

  32. [40]

    FMA: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” in18th International Society for Music Information Retrieval Conference (ISMIR),

  33. [2017]

    Available: https://arxiv.org/abs/1612.01840

    [Online]. Available: https://arxiv.org/abs/1612.01840

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.