Overview of artificial intelligence model watermarking

Hanzhou Wu, Jie Zhang, Yue Li, Zhaoxia Yin, Xinpeng Zhang, Hui Tian, Bin Li, Weiming Zhang, Nenghai YuJournal of Image and Graphics 2023 · DOI

Abstract

As artificial intelligence models, particularly deep neural networks, become central to commercial products, they represent significant intellectual property (IP) built through massive data and compute resources. However, these models are susceptible to unauthorized theft and tampering. This paper provides a comprehensive review of AI model watermarking—a specialized technique to embed ownership signatures into neural networks. We categorize these methods into white-box, black-box, and box-free approaches, evaluating them based on imperceptibility, robustness, and security.

The Challenge of Protecting AI Assets

Training state-of-the-art models requires specialized expertise and high-end hardware. Unlike traditional media (images or video), AI models are functional entities. Simply modifying weights to embed a watermark can degrade the model's primary task performance. Effective watermarking must survive "attacks" like fine-tuning or pruning while remaining hidden from attackers.

Technical Frameworks for Model Watermarking

The landscape of AI watermarking is generally divided by the level of access a verifier has to the suspect model.

White-box Watermarking

In white-box scenarios, the verifier has full access to the model's internal parameters and architecture. Watermarks are often embedded by applying regularization during training to force specific weight distributions or by utilizing the model's structural properties.

Embedding watermarks in model parameters or structure

Black-box Watermarking

Black-box methods are highly practical for commercial APIs where the verifier only sees inputs and outputs. This is typically achieved through trigger sets. The model is trained to behave normally on standard data but output a specific, pre-defined "incorrect" label when it encounters a secret trigger (e.g., a specific pattern on a stop sign).

Black-box watermarking process using trigger sets and backdoors

Box-free and Fragile Watermarking

  • Box-free: Watermarks are extracted from the model's generated content (e.g., pixels in a GAN-generated image), requiring no direct model interaction.
  • Fragile: Designed to break upon any modification, these act as "tamper-evident seals" to ensure model integrity.

Key Insights and Future Directions

A significant challenge is the "Verification-is-Exposure" Paradox: revealing a watermark to prove ownership might help an attacker remove it. Research is moving toward zero-knowledge proofs to solve this. Additionally, the Lottery Ticket Hypothesis suggests that watermarks can be embedded into specific sparse sub-networks, making them resilient to aggressive pruning.

While current techniques are mature for classification tasks, several frontiers remain:

  1. Model Extraction Attacks: Watermarks often fail when an attacker trains a new "student" model to mimic the "teacher" model's behavior.
  2. Large Language Models (LLMs): Protecting the vast knowledge within Transformers is the next major hurdle.
  3. Standardization: A unified benchmark is needed to objectively compare the robustness of different watermarking schemes.

Summary updated

Abstract

以神经网络为代表的人工智能技术在计算机视觉、模式识别和自然语言处理等诸多应用领域取得了巨大的成功,包括谷歌、微软在内的许多科技公司都将人工智能模型部署在商业产品中,以提升服务质量和经济效益。然而,构建性能优异的人工智能模型需要消耗大量的数据、计算资源和专家知识,并且人工智能模型易于被未经授权的用户窃取、篡改和贩卖。在人工智能技术迅速发展的同时,如何保护人工智能模型的知识产权具有显著学术意义和产业需求。在此背景下,本文主要介绍基于数字水印的人工智能模型产权保护技术。通过与传统多媒体水印技术进行对比,首先概述了人工模型水印的研究意义、基础概念和评价指标;然后,依据水印提取者是否需要掌握目标模型的内容细节以及是否需要和目标模型进行交互,从“白盒”模型水印、“黑盒”模型水印、“无盒”模型水印3个不同的角度分别梳理了国内外研究现状并总结了不同方法的差异,与此同时,对脆弱模型水印也进行了分析和讨论;最后,通过对比不同方法的特点、优势和不足,总结了不同场景下模型水印的共性技术问题,并对发展趋势进行了展望。
BibTeX
@article{wu2023overview,
  author = {Hanzhou Wu and Jie Zhang and Yue Li and Zhaoxia Yin and Xinpeng Zhang and Hui Tian and Bin Li and Weiming Zhang and Nenghai Yu},
  title = {{Overview of artificial intelligence model watermarking}},
  journal = {Journal of Image and Graphics},
  year = {2023},
  doi = {10.11834/jig.230010},
}
Figure