Reproduce “WAKE: Watermarking Audio with Key Enrichment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yaoxun Xu, Jianwei Yu, Hangting Chen, Zhiyong Wu, Xixin Wu, Dong Yu, Rongzhi Gu, Yi Luo
The WAKE (Watermarking Audio with Key Enrichment) framework introduces the first key-controllable audio watermarking system. It enables simultaneous embedding and independent extraction of multiple watermarks while supporting variable lengths without retraining, outperforming baselines like WavMark and AudioSeal in audio quality and decoding accuracy.
Key insights
- Using an Invertible Neural Network (INN) where each key bit controls a distinct invertible block enables a cryptographic-like mechanism for secure, key-dependent audio transformations.
- A specialized Predict Module within the INN enhances decoding accuracy by generating audio-dependent features that guide the extraction process, reducing reliance on random sampling.
- A dual training strategy alternating between single and sequential double watermark scenarios allows the model to robustly handle multiple independent embeddings simultaneously.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.