Abstract: Present methods for protocol format extraction analyze the execution traces of programs at syntax level, which leads to redundancy and conflict in the results of field identification. In order to improve the accuracy of field identification, a semantic level method was proposed for protocol format extraction. The method firstly translated the binary instructions into equivalent intermediate language, and then tracked the parsing process of field semantics through fine-grained dynamic taint analysis. Further, it extracted protocol format using semantic level policies of field identification, based on the semantic indivisibility of fields. Experimental results show that the proposed method can achieve high identification accuracy with low complexity.