WatermarkLab: A Comprehensive Framework for Robust Image Watermarks Benchmarking and Development

Jiale Chen, Xiaoyan Yuan, Wei Wang, Yuanman Li, Li Dong, Feng Liang, Chongyang Shi, Xiping Hu

Keywords: robust image watermarking, benchmark;

Abstract: With the growing demand for multimedia content, image protection has become increasingly important. Robust image watermarking, as a core technology for copyright protection, has attracted extensive attention. To advance research in this field, we propose WatermarkLab, a comprehensive framework for systematic benchmarking of robust image watermarks and the development of new methods. WatermarkLab supports benchmarking of all types of blind robust image watermarks, including in-generation watermarks and post-generation watermarks. Beyond benchmarking, WatermarkLab integrates 10 representative watermarking methods for systematic comparison. It also includes 34 attackers for benchmarking and 28 differentiable attackers for development. Furthermore, we evaluate the robustness of 9 watermarking methods under 34 attackers and give their weaknesses, assisting researchers in enhancing more robust watermarking methods and designing new watermark removal attackers. In addition, the framework provides auxiliary tools such as arithmetic coding and reversible data hiding commonly used in robust reversible watermarking. For result visualization, WatermarkLab offers comprehensive visualization tools and an interactive website, enabling researchers to intuitively analyze and compare benchmarking results. In summary, WatermarkLab is a powerful framework, aiming to establish a comprehensive, fair, open, and extensive platform for blind robust image watermark benchmarking and development. Interactive visualization website code is available at: https://anonymous.4open.science/r/watermarklab-website.

Primary Area: datasets and benchmarks

Submission Number: 3641