Black Box Watermarking Technology of the AI Model

Abstract

Today's rapidly developing AI models, black box watermarking technology is becoming an important means of intellectual property protection as it does not require access to the internal structure of the model. The theme of this paper is the black box watermarking technology of artificial intelligence models. First, the paper explains its theory, utilizing the ‘memory’ of the model to achieve identification through embedding and verification processes. Second, based on the different watermark embedding methods, two mainstream embedding methods are analyzed. In the training embedding stage, the model ‘remembers’ the mapping relationship, while in the fine-tuning stage, the watermark is embedded with a small number of parameters. Besides, this paper explores key performance indicators including robustness fidelity. Finally, for the current technological development situation, suggestions such as dynamically triggering sample generation are proposed, and future development trends are pointed out. This paper may help the researchers who are working in this field.