CRII: CIF: Automated and Robust Image Watermarking: A Deep Learning Approach

数字图像水印是指将信息秘密地嵌入到封面图像中，并从标记图像中提取信息的过程，它被应用于从秘密通信到身份验证到安全的各种应用领域。虽然有许多手工制作的水印方案，但这些传统的方法由于人工设计固有的有限范围而遇到了困难。为了实现适应日益多样化的应用场景的图像水印，本项目旨在基于深度学习的思想来开发新的方案。将解决两个主要问题，即(I)最小化对领域知识的需求，(Ii)在没有先验知识的情况下实现健壮性。该项目的成果将有助于开发新一代强大的智能水印工具，这些工具可以支持摄像头扫描和安全物联网设备自注册等尖端应用。将拟议的研究活动纳入大学课程编制和其他教育方案，将有助于各级的STEM教育。该项目旨在通过开发图像水印方案来推动基于数字图书馆的图像水印的发展，该方案实现了对水印规则的稳健概括，而不需要关于标签、原始图像或失真的信息。研究议程围绕两个相辅相成的研究活动进行：(I)基于距离函数、鉴别器分类器或度量学习的相似性度量的基于动态链接库的自动图像水印；(Ii)基于动态链接库的稳健图像水印，探索不变的图像潜在空间和自动纠正。将制定的计划将在不同的应用程序上进行测试，以确定其实用性。这些研究活动有望在多个方面加深我们对水印的理解，即(I)如何设计深度学习组件(如体系结构和层)和新算法(通过相似性度量)来充分概括图像水印过程中的图像特征和功能；(Ii)如何设计DL组件以实现对图像水印中不同类型的失真的稳健性，而不需要先验知识或相反的例子；以及(Iii)这些设计如何能够实现各种新颖的水印应用场景和使用案例。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Digital image watermarking refers to the process of covertly embedding information into a cover-image and extracting it from it the marked-image; it is used in various application areas ranging from covert communication to authentication to security. Although many handcrafted watermarking schemes are available, these traditional methods run into difficulties due to the limited scope inherent to manual design. To implement image watermarking which adapts to the demands of increasingly diverse application scenarios, this project aims to develop novel schemes based on ideas from deep learning (DL). Two major problems will be addressed, namely (i) minimizing the requirement of domain knowledge, and (ii) achieving robustness without prior knowledge. Outcomes of this project will contribute to a new generation of robust and intelligent watermarking tools that can support cutting-edge applications such as camera scans and secured Internet-of-Things device on-boarding. The integration of the proposed research activities into university curriculum development and other educational programs will contribute to STEM education at various levels. This project seeks to advance the state-of-the-art in DL—based image watermarking through the development of image watermarking schemes that achieve a robust generalization of watermarking rules without requiring information about labeling, the original images, or distortions. The research agenda is structured around two complementary research activities: (i) DL—based automated image watermarking with similarity measures of distance functions, discriminator classifiers, or metric learning; and (ii) DL—based robust image watermarking that explores invariant image latent spaces and automatic rectification. The schemes to be developed will be tested on different applications to confirm their practicality. These research activities are expected to advance our understanding of watermarking on a number of fronts, namely (i) how to design deep learning components (such as architectures and layers) and novel algorithms (through similarity measures) to fully generalize image features and functions for image watermarking processes; (ii) how to design DL components to achieve robustness to different types of distortions in image watermarking, without requiring prior knowledge or adversarial examples; and (iii) how these designs can enable various novel watermarking application scenarios and use cases.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.