.png){kind=icon}
.png){kind=icon}
Return
MisConfDetect: Deterministic Misconfiguration Detection via Automated Validation Code Synthesis
.png){kind=icon}
PRE
AI
DOI：10.1109/tce.2026.3689049.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid expansion of consumer electronics (CE) has forged highly interconnected digital ecosystems where networking devices act as the foundational security backbone. However, the increasing complexity of network configurations has turned these devices into a primary source of security anomalies—misconfigurations—that not only lead to network outages but also create security concerns and privacy risks in these environments. While Large Language Models (LLMs) have recently been adopted for anomaly detection, existing solutions lack the trustworthiness and resilience required for safety-critical CE environments due to hallucination risks, lack of interpretability, and high runtime overhead. To address these challenges, we propose MisConfDetect, a proactive threat hunting framework designed to secure CE infrastructure through deterministic misconfiguration detection powered by LLMs. Our framework introduces a two-stage strategy: an Offline Synthesis Stage, where configuration knowledge is extracted from vendor manuals and transformed into structured domain-specific language specifications to generate deterministic validation code; and an Online Detection Stage, where device configurations are automatically validated by executing the pre-generated code. By decoupling LLM reasoning from online detection, MisConfDetect mitigates hallucination risks, improves efficiency, and ensures interpretable and reproducible results. Experiments demonstrate that MisConfDetect achieves 12% and 26% improvements in Detection Rate at the command and function levels, respectively, over the best-performing baseline, while maintaining reliable code execution with 100% precision.
Keywords:
Consumer electronics
anomaly detection
large language model
Journal
IF:
10.9
Papers:
5.1K
Citations:
6.8K