基于所有者身份识别的自然语言处理模型水印算法OA

In order to ensure that model developers and data providers can effectively protect their intellectual assets,this paper proposes a watermarking framework for Natural Language Pro-cessing NLP)models based on owner identity OIRW)to address the security and robustness chal-lenges of model copyright protection.Specifically,during the model training stage,a digital signa-ture is generated using the user's copyright information and a secret key to link the trigger,which is then inserted into the dataset to train the watermarked model.In the model verification stage,the owner's identity is first verified via the digital signature,copyright information,etc.;subsequently the triggered test data is fed into the remote model to obtain the watermark verification result.To e-valuate the watermarking performance,watermark information was embedded into three common lan-guage models on the AgNews and SST-2 datasets.The experimental results show that the watermark verification accuracy is close to 100％,demonstrating strong robustness under attack scenarios such as model fine-tuning,pruning,and overwriting.