摘要
1 Introduction Recent advancements in encoder-decoder based text generation technology,like ChatGPT by OpenAI,and PaLM[1]by Google,have garnered attention in the AI community.Pay-per-use APIs offer access to these models,but research shows they are prone to imitation attacks,where malicious users train their models through skillfully crafted queries to get responses from lawful APIs.Such attacks violate the intellectual property(IP)and deter further research[2].Recent work introduced lexical watermarking(LW)methods to protect legal APIs’IP.LW modifies the original outputs and uses null-hypothesis test for ownership verification on imitation models[2,3].High-frequency words are selected,and WordNet synonyms replace them,but this one-size-fits-all approach neglects rational substitutes.
基金
This research was partially supported by the National Natural Science Foundation of China(Grant Nos.62076217 and U22B2037)
the Blue Project of Yangzhou University.
