![[2602.21508] WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21508] WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck
Summary
The paper introduces WaterVIB, a framework for robust watermarking that utilizes the Variational Information Bottleneck to enhance resilience against generative attacks, outperforming existing methods significantly.
Why It Matters
As digital content becomes increasingly vulnerable to unauthorized use and manipulation, robust watermarking is essential for protecting intellectual property. WaterVIB addresses the shortcomings of current methods by focusing on learning minimal sufficient statistics, making it a crucial advancement in the field of watermarking and machine learning.
Key Takeaways
- WaterVIB reformulates watermarking through Variational Information Bottleneck for improved robustness.
- The method filters out redundant details, focusing on essential signals that withstand generative shifts.
- Extensive experiments show WaterVIB's superior performance against state-of-the-art watermarking techniques.
Computer Science > Machine Learning arXiv:2602.21508 (cs) [Submitted on 25 Feb 2026] Title:WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck Authors:Haoyuan He, Yu Zheng, Jie Zhou, Jiwen Lu View a PDF of the paper titled WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck, by Haoyuan He and 3 other authors View PDF HTML (experimental) 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 superio...
