Codable watermarking method oriented to data set intellectual property protection

According to the data set intellectual property protection-oriented encodable watermarking method provided by the invention, third parties providing machine learning as a service can be verified, and when data leakage occurs, suspicious third parties leaking data can be traced to the source. Specifically, in the field of deep learning, the method uses an automatic codec to embed a backdoor trigger as a watermark into data, and uses a clean tag backdoor attack method to deduce whether a third-party model uses user data for training. And the backdoor trigger is bound with the code, so that the data leakage party can be traced according to the code. In addition, in the multimedia field, data with watermarks can be directly decoded to obtain corresponding codes for verification and responsibility investigation. According to the watermarking method, no extra label information needs to be provided, and verification can be carried out only by providing a small amount of data. 本发明提出一种面向数据集知识产权保护的可编码水印方法，可以对提供机器学习即服务的