GAI-AntiCopy: Infrequent Transformation Aided Accuracy-Consistent Copyright Protection for Generative AI Instructions in NGN
Yixin Fan, Jun Wu · IEEE Transactions on Cognitive Communications and Networking · 2025
Generative artificial intelligence (GAI) brings an unprecedented revolution to the next-generation networks (NGN) from resource allocation to network traffic monitoring. With its powerful creative content generation capabilities, GAI significantly enhances the interaction and quality of customized services in NGN. Currently, benefiting from the thriving GAI services, it is possible to build personalized GAIs through designing GAI instructions without the need for training models from scratch. Meanwhile, infringements like pirating are emerging, necessitating effective copyright protection schemes. However, current schemes suffer from an unacceptable decrease in task processing accuracy when applied to GAIs, and the success rate of watermarking is extremely low on GAI instructions. Therefore, we propose an infrequent transformation aided accuracy-consistent copyright protection scheme for GAI instructions. We first build a comprehensive GAI instruction copyright protection system for NGN, designing a complete watermarking and verification mechanism. Additionally, we integrate copyright watermark messages with the syntactic features of GAI instructions to select the embedding positions. Watermarks are embedded through emphasis and passivization, which are infrequent transformations that minimize semantic distortion. Finally, we conduct experiments on real GAI instructions datasets and compare our scheme with existing works to demonstrate that ours effectively realizes accuracy-consistent copyright protection for GAI instructions in NGN.