WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck

TL;DR

WaterVIB introduces a theoretically grounded watermarking framework that learns minimal sufficient representations to enhance robustness against generative attacks, outperforming existing methods in zero-shot scenarios.

Contribution

The paper proposes WaterVIB, a novel variational information bottleneck approach that filters out cover details, improving watermark robustness against regeneration-based attacks.

Findings

Outperforms state-of-the-art watermarking methods

Achieves superior zero-shot resilience against diffusion-based editing

Theoretically guarantees robustness through information bottleneck optimization

Abstract

Robust watermarking is critical for intellectual property protection, whereas existing methods face a severe vulnerability against regeneration-based AIGC attacks. We identify that existing methods fail because they entangle the watermark with high-frequency cover texture, which is susceptible to being rewritten during generative purification. To address this, we propose WaterVIB, a theoretically grounded framework that reformulates the encoder as an information sieve via the Variational Information Bottleneck. Instead of overfitting to fragile cover details, our approach forces the model to learn a Minimal Sufficient Statistic of the message. This effectively filters out redundant cover nuances prone to generative shifts, retaining only the essential signal invariant to regeneration. We theoretically prove that optimizing this bottleneck is a necessary condition for robustness against distribution-shifting attacks. Extensive experiments demonstrate that WaterVIB significantly outperforms state-of-the-art methods, achieving superior zero-shot resilience against unknown diffusion-based editing.