- category
- watermark-remover
- maturity
- emerging
- approach
- Academic method that, from a handful of clean/watermarked audio pairs, computes cheap structural probes to diagnose where a watermark is embedded (its domain), then applies a single domain-matched attack instead of a blind distortion sweep; also produces a threshold-free 'fragility score' (area under the accuracy-versus-quality trade-off curve) and can identify which of ten watermarking schemes is present with 84% accuracy from the probe signatures alone.
- access
- free
- owns
- A research method (arXiv, submitted 2026-08-17) demonstrating that magnitude/carrier-domain neural audio watermarks, including Meta's AudioSeal, can be erased or have their detection flag removed by a single targeted, training-free attack at high perceptual quality (PESQ >= 3.6), while latent-domain schemes resist the same attacks.
- distinctFrom
- Unlike marknull (an image-domain watermark-removal paper using on-manifold latent manipulation against SynthID-Image), this targets audio watermarks specifically and is diagnostic rather than blind: it first probes to identify the watermark's embedding domain (magnitude/carrier vs. latent) before choosing an attack, and is the first entry in this catalogue to directly attack the already-tracked meta-audioseal scheme.
- notes
- Submitted 2026-08-17, two days before this lens's 2026-08-19 window start; included as a gap-fill following the same precedent as the already-tracked marknull entry (paper-only research method, access marked 'free' rather than 'open-source' since no code/tool release is mentioned anywhere on the arXiv page). Single-author paper (Likhith Kumara); no independent replication found.
