LAW: A Training-Free Latent Watermark for Diffusion-Generated Audio Attribution
IEEE Signal Processing Letters · 2026 · 33 · 2580-2584
DOI: 10.1109/lsp.2026.3702794 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Ruizhe Zhu und 5 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Zhejiang Province Key R&D Project, Hangzhou Municipal Key R&D Project.
Abstract (Originalsprache)
The increasing realism of generative audio models has made it difficult to distinguish synthetic audio from authentic recordings, creating a need for reliable attribution. This letter presents a training-free latent audio watermarking method for diffusion-generated audio that embeds source information during generation rather than after waveform synthesis. The method perturbs the initial latent variables through a tailored rearrangement strategy, thereby preserving perceptual quality while maintaining the generative process of the frozen model. A lightweight inversion-based detector is then used to recover and verify the embedded message from the generated audio. Because the watermark is introduced in the latent space, the method remains robust to distortions caused by vocoding, signal processing, and regeneration. Experimental results show that the proposed method preserves audio quality and achieves an average extraction accuracy above 95% under diverse post-processing attacks in high-capacity settings. These results indicate that the method provides an effective solution for synthetic audio attribution and provenance verification without requiring model retraining.
Förderung
- National Natural Science Foundation of China
- Zhejiang Province Key R&D Project
- Hangzhou Municipal Key R&D Project
- Sino-French Joint Laboratory for Digital Media Forensics of Zhejiang Province
- Zhejiang Provincial Key Laboratory for Sensitive Data Security Protection and Confidentiality Management
Zitieren
APA
Zhu, R., Zou, L., Lai, S., Shen, Z., Li, Z., & Qiao, T. (2026). LAW: A Training-Free Latent Watermark for Diffusion-Generated Audio Attribution. IEEE Signal Processing Letters, 33, 2580-2584. https://doi.org/10.1109/lsp.2026.3702794
Vancouver
Zhu R, Zou L, Lai S, Shen Z, Li Z, Qiao T. LAW: A Training-Free Latent Watermark for Diffusion-Generated Audio Attribution. IEEE Signal Processing Letters. 2026;33:2580-2584. doi:10.1109/lsp.2026.3702794
BibTeX
@article{Zhu2026,
author = {Zhu, Ruizhe and Zou, Lingyong and Lai, Shuhong and Shen, Zhangyi and Li, Zhuang and Qiao, Tong},
title = {LAW: A Training-Free Latent Watermark for Diffusion-Generated Audio Attribution},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2580--2584},
doi = {10.1109/lsp.2026.3702794}
}
EUVIMED