WaterVIB: Learning minimal sufficient watermark representations via a variational information bottleneck

WaterVIB learns a minimal sufficient watermark representation of the message through a variational information bottleneck to resist regeneration-based AIGC attacks. The method refactors the encoder into an information sieve, filtering out high-frequency carrier textures that are easily rewritten by generative purification and retaining only signals invariant to regeneration; the paper states that optimizing this bottleneck is a necessary condition for defending against distribution-shift attacks. Experiments show that its zero-shot robustness under unknown diffusion model editing is significantly better than existing methods.