ECG: Augmenting Embedded Operating System Fuzzing via LLM-Based Corpus Generation
Qiang Zhang,Yuheng Shen,Jianzhong Liu,Yiru Xu,Heyuan Shi,Yu Jiang,Wanli Chang
DOI: https://doi.org/10.1109/tcad.2024.3447220
IF: 2.9
2024-11-09
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Abstract:Embedded operating systems (Embedded OSs) power much of our critical infrastructure but are, in general, much less tested for bugs than general-purpose operating systems. Fuzzing Embedded OSs encounter significant roadblocks due to much less documented specifications, an inherent ineffectiveness in generating high-quality payloads. In this article, we propose ECG, an Embedded OS fuzzer empowered by large language models (LLMs) to sufficiently mitigate the aforementioned issues. ECG approaches fuzzing Embedded OS by automatically generating input specifications based on readily available source code and documentation, instrumenting and intercepting execution behavior for directional guidance information, and generating inputs with payloads according to the pregenerated input specifications and directional hints provided from previous runs. These methods are empowered by using an interactive refinement method to extract the most from LLMs while using established parsing checkers to validate the outputs. Our evaluation results demonstrate that ECG uncovered 32 new vulnerabilities across three popular open-source Embedded OS (RT-Linux, RaspiOS, and OpenWrt) and detected ten bugs in a commercial Embedded OS running on an actual device. Moreover, compared to Syzkaller, Moonshine, KernelGPT, Rtkaller, and DRLF, ECG has achieved additional kernel code coverage improvements of 23.20%, 19.46%, 10.96%, 15.47%, and 11.05%, respectively, with an overall average improvement of 16.02%. These results underscore ECG's enhanced capability in uncovering vulnerabilities, thus contributing to the overall robustness and security of the Embedded OS.
engineering, electrical & electronic,computer science, interdisciplinary applications, hardware & architecture