An Effective Framework for Intellectual Property Protection of NLG Models

Natural language generation (NLG) models combined with increasingly mature and powerful deep learning techniques have been widely used in recent years. Deployed NLG practical applications may be stolen or illegally, watermarking has become an important tool to protect the Intellectual Property (IP) of these models. Watermarking technique designs algorithms embed watermark information extracts for IP identification can seen as a symmetric signal processing problem. In terms protection models, however, existing approaches cannot provide reliable timely model prevent illegal users from utilizing original performance addition, quality watermarked text sequences generated by some is not high. view these, this paper proposes two embedding schemes hidden memory state RNN different tasks. Besides, we add loss decoder improve grammatical correctness output sequences. During experiments, it proved that our approach does compromise on corresponding datasets outputs high-quality sequences, while forged secret keys will generate unusable thus defeating purpose infringement. also conduct sufficient experiments prove proposed strong robustness under attacks.