自然杂志 ›› 2022, Vol. 44 ›› Issue (4): 267-273.doi: 10.3969/j.issn.0253-9608.2022.04.002

• 纪念上海大学建校百年专刊 • 上一篇    下一篇

深度模型水印

张新鹏 ，吴汉舟   

  1. 上海大学 通信与信息工程学院，上海 200444
  • 收稿日期:2022-04-27 出版日期:2022-08-25 发布日期:2022-08-27
  • 通讯作者: 张新鹏，研究方向：多媒体信息安全、信息隐藏、数字取证、加密域信号处理、图像处理。

Deep model watermarking

ZHANG Xinpeng, WU Hanzhou   

  1. School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China
  • Received:2022-04-27 Online:2022-08-25 Published:2022-08-27

PDF

484

摘要： 深度神经网络模型凝结了设计者的智慧，需要消耗大量数据和计算资源，是人工智能技术的重要产出物，已被广泛应用于生产和生活当中。然而，作为一种数字产品，如何保护深度神经网络模型免于被非法复制、分发或滥用(即知识产权保护)是人工智能产业化进程中必须面临和解决的难题。文章主要介绍基于数字水印的深度模型产权保护技术，通过总结深度模型水印的发展现状，对深度模型水印的研究趋势进行展望。

关键词: 深度模型, 数字水印, 产权保护, 人工智能安全

Abstract: Deep neural networks (DNNs) condense the wisdom of the designer and consume a lot of data and computing resources. It is an important artificial intelligence technology, and is widely applied in our daily life. However, as a digital asset, how to protect DNN models from being illegally copied, distributed or abused (that is, intellectual property protection) is a difficult problem that must
be faced and solved in the process of artificial intelligence industrialization. This article reviews digital watermarking techniques for
intellectual property protection of DNN models. By summarizing the development status of deep model watermarking, the research
trend of deep model watermarking is further prospected.

法律公告　|　 联系我们　|　 网站地图