Codec representations that preserve watermark-aligned latent directions

Determine which neural codec representations preserve watermark-aligned directions in latent space, with the aim of improving the generalization of latent audio watermarking across codecs.

Background

The paper introduces DeltaMark, a content-independent watermarking method that adds a payload-dependent perturbation to the continuous latent representation of a frozen EnCodec model. Experiments show that learned perturbations survive EnCodec resynthesis more effectively than equal-norm random perturbations, and that preservation depends strongly on the allocation of perturbation energy across latent channels.

However, robustness does not transfer uniformly to other neural codecs: DeltaMark remains detectable under DAC but fails under TiCodec-1g4r. The authors therefore identify the codec representations responsible for preserving watermark-aligned directions as an unresolved issue relevant to developing higher-fidelity and more broadly transferable latent watermarking systems.