LLM Paternity Test: Generated Text Detection with LLM Genetic Inheritance
Large language models (LLMs) can generate texts that carry the risk of various misuses, including plagiarism, planting fake reviews on e-commerce platforms, or creating inflammatory false tweets. Detecting whether a text is machine-generated has thus become increasingly important. While existing detection methods exhibit superior performance, they often lack generalizability due to their heavy dependence on training data. To alleviate this problem, we propose a model-related generated text detection method, the LLM Paternity Test (LLM-Pat). Specifically, given any candidate text (\textit{child}), LLM-Pat employs an intermediary LLM (\textit{parent}) to reconstruct a \textit{sibling} text corresponding to the given text and then measures the similarity between candidate texts and their sibling texts. High similarity indicates that the candidate text is machine-generated, akin to genetic traits. We have constructed datasets encompassing four scenarios: student responses in educational settings, news creation, academic paper writing, and social media bots to assess the performance of LLM-Pat. The experiments show that LLM-Pat outperforms the existing detection methods and is more robust against paraphrasing attacks and re-translating attacks. Besides, LLM-Pat can also be used to trace which large language model the text was generated by. The constructed dataset and code will be released to benefit the community.
PDF Abstract
