Organized by the IEEE Computational Intelligence Society Shenzhen Chapter.

Activity Aim & Talk

In this activity, Guangsheng Bao will talk about Fast-DetectGPT. Large language models like ChatGPT and GPT-4 have brought convenience to people’s work and lives across various sectors. However, concerns about their misuse, including issues like fake news, malicious product reviews, and plagiarism, have also arisen. This paper introduces a novel text detection method Fast-DetectGPT that operates without training and utilizes open-source small language models to detect text generated by various large language models. Fast-DetectGPT increases the detection speed by 340 times and improves the detection accuracy by 75% relative to previous methods, establishing it as the new state-of-the-art (SOTA). In the detection of text generated by widely used models such as ChatGPT and GPT-4, it surpasses the accuracy of the commercial system GPTZero. Fast-DetectGPT achieves high accuracy, high speed, low cost, and versatility, clearing the hurdles for practical application.

Meet the Speaker

Guangsheng Bao is a second-year Ph.D. student in Professor Yue Zhang’s lab at Westlake University (https://frcchang.github.io/). His personal webpage can be found at https://baoguangsheng.github.io/. His primary interests lie in trustworthy natural language generation technologies, including controllability, interpretability, security, and reasoning of models. He has published several papers in top conferences and journals in the field, including ICLR 24, ACL 23, ACL 21, EMNLP 23, AAAI 21, and TASLP 23. Before his doctoral studies, Bao Guangsheng worked for several years at Microsoft (China) and Alibaba, where he was involved in the research and development of Microsoft’s personal assistant, Cortana.