Backward Compatibility in Zero-Bit Neural Audio Watermarking
Legacy-chain training plus angular diversity loss makes neural audio watermarks backward compatible.
Neural audio watermarking usually trains coupled embedder-detector pairs that fail when versions are updated independently. The authors formalize backward compatibility for zero-bit watermarking and train a new detector on a performance-weighted sequence of frozen legacy embedders. That coupling induces embedder collapse, which sharply lowers the true-negative rate against out-of-chain embedders. An Angular Diversity Loss forces each new embedder into a distinct output region, restoring near-perfect TNR with no measured loss of detection accuracy or perceptual quality.
- Formalizes backward compatibility for zero-bit neural audio watermarking.
- New detectors train against frozen legacy embedders with weighted sampling.
- Legacy coupling causes embedder collapse and poor out-of-chain true negatives.
- Angular Diversity Loss restores near-perfect TNR without hurting quality.
From the original article
1 min read at arxiv.orgNeural watermarking systems are typically trained as tightly coupled embedder-detector pairs and implicitly assume static deployment.
By Admir Krnić · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.