Correlation-Aware and Gaussianity-Preserving Robust Latent Angular Watermarking for Diffusion Models

TL;DR - LAW embeds angular watermarks in diffusion-model latents while preserving Gaussianity and limiting correlation degradation. Its geometry-based encoding is designed to improve robustness without sacrificing generation fidelity.

• Encodes bits as antipodal angles between disjoint pairs of latent elements.
• Maximizes bit separation using π-spaced encoding and derives angular-error variance proportional to (1/\rho^2).
• LAW-M selects high-magnitude, geometrically stable latent dimensions for added robustness.
• Derives a closed-form autocorrelation structure, with induced correlations confined to sparse off-diagonal entries of fixed (\pm\pi/4).