MFD: Multi-Feature Detection of LLM-Generated Text

Abstract With the rapid development of large language models, their powerful capabilities have led to their rapid popularity in society. However, it not only brings great convenience to people’s life and work but also provides a favorable tool for criminals to carry out malicious behaviors. Therefore, to prevent the malicious use of large language models, there is a growing demand for a detector that can efffciently discriminate texts generated by large language models. In this paper, Multi-Feature Detection (MFD), a new zero-shot method, is introduced. MFD comprehensively considers log-likelihood, log-rank, entropy, and LLM-Deviation. LLM-Deviation is a new statistical feature proposed in this paper and has a clear distribution difference between texts generated by LLMs and those written by humans. Experiments show MFD is more effective than the existing zero-shot method. MFD improves the detection performance by 1.02 F1 score on average on the HC3-English dataset. In generalization ability, MFD is also very competitive compared with the existing zero-shot method.