In recent years, there have been notable advancements in the field of natural language generation, particularly with the development of large-scale PLMs like ChatGPT [1] and GPT-4 [2]. The texts produced by these models are of such exceptional quality that it can be challenging for humans to discern them from those written by people. In fact, according to a technical report by OpenAI, the majority of texts generated by GPT-2 were already indistinguishable from those written by humans [3]. These PLMs have a broad range of applications, including story [4] and dialogue generation [5], as well as code writing [6]. Nonetheless, they can also be easily exploited by malicious actors to fabricate fake news [7, 8, 9] and comments [10] for personal profit or political interference, thereby posing a significant threat to society. Therefore, it is imperative to explore automatic methods for detecting machine-generated text to identify disinformation and mitigate the likelihood of abuse [11].