以神经网络为代表的人工智能技术在诸多应用领域取得了巨大的成功,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,作为一种数字资产,人工智能模型易于被窃取、篡改和贩卖,在人工智能技术迅速发展的同时,如何保护其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,通过概述模型水印的研究意义、基础概念、评价指标和技术分类,梳理国内外研究现状并总结差异,同时对比优势和劣势, 对模型水印的发展趋势进行展望。

Artificial intelligence technologies represented by neural networks have achieved great success in many areas, and many technology companies have deployed artificial intelligence models in commercial products to improve service quality and economic benefits. However, as a digital asset, artificial intelligence models are easy to be stolen, tampered with and sold. With the rapid development of artificial intelligence technology, how to protect the intellectual property of advanced artificial intelligence models has important academic significance and industrial needs. In this context, this paper mainly introduces the intellectual property protection of artificial intelligence models based on digital watermarking. We first introduce the research significance, basic concepts, evaluation metrics and technical classification of artificial intelligence model watermarking. The international and domestic development status of artificial intelligence model watermarking is then summarized and discussed. By analyzing and comparing the advantages and disadvantages of international and domestic research, the prospects for the development trend of model watermarking are finally provided.