Backward Compatibility in Zero-Bit Neural Audio Watermarking
The authors report preserving detection of older audio watermarks while restoring near-perfect rejection of embedders outside a detector’s intended set.
Their framework trains each new detector against a sequence of frozen legacy embedders, weighting sampling by performance. That preserves detection across earlier versions but causes an “embedder-collapse” effect that severely degrades rejection of out-of-chain embedders. The proposed Angular Diversity Loss forces new embedders into distinct regions of the output space. The authors report no measurable cost to detection accuracy or perceptual audio quality.
ArXiv · AI/CL/LG's note
Their framework trains each new detector against a sequence of frozen legacy embedders, weighting sampling by performance. That preserves detection across earlier versions but causes an “embedder-collapse” effect that severely degrades rejection of out-of-chain embedders. The proposed Angular Diversity Loss forces new embedders into distinct regions of the output space. The authors report no measurable cost to detection accuracy or perceptual audio quality.
ArXiv · AI/CL/LG's note
score 4