Large Language Model-Assisted Security Testing for Embedded Systems

Fuzzing of embedded systems in a system simulation environment is currently a mainstream security testing method. To address the limitations of traditional embedded system fuzzing, such as random byte perturbation existing in traditional embedded system fuzzing technologies, this paper proposes a large language model-assisted fuzzing method. Firstly, static analysis is used to extract the semantic information of data register-related functions and decompiled code, which serve as the input basis for constructing prompts; Secondly, various types of peripheral protocols are analyzed to generate initial seeds that are more in line with the firmware input format; Finally, dynamic detection of coverage is performed, and the large language model is used to assist in generating test cases in a timely manner, helping to increase the number of paths explored by firmware execution. Experiments show that this method improves the code coverage of the tested firmware by 14.2%~52.9% and effectively enhances the generation quality of test cases.