Entropy-aware dynamic bias watermarking for LLM-generated emotional content

The application of large language models(LLMs) in affective computing-ranging from empathetic chatbots to creative writing-has intensified the demand for distinguishing and authenticating AI-generated emotional content. Watermarking, by embedding detectable signals into the outputs of language models, offers a promising solution. However, a critical challenge persists: emotional texts often exhibit low entropy or complex spiky entropy distributions, which severely undermine the performance of existing watermarking methods. Unlike prior works that primarily treated spiky entropy as an external metric, we focus specifically on its role within the text generation process itself. To address the challenges of watermarking under low-entropy and complex entropy distributions, we propose DBW (Dynamic Bias Watermarking)-an entropy-aware watermarking algorithm for LLMs. DBW dynamically adjusts the watermarking bias in real time based on the entropy of each token. This innovation ensures a stronger watermark signal (increased green token count) in high-entropy contexts, while minimizing interference and quality degradation in fragile low-entropy emotional segments. Experimental results demonstrate that the proposed DBW algorithm outperforms the KGW watermarking method in both complex entropy distribution and low-entropy text generation scenarios. DBW achieves higher detection accuracy without sacrificing text quality. Furthermore, comparative experiments show that our proposed DBW algorithm demonstrates superior robustness under different attacks. Our work provides a reliable and adaptive tool for safeguarding emotion-AI generated content, contributing to the secure and trustworthy deployment of large-scale pre-trained models in affective computing.