Large Language Model Assisted Kernel Data Race Detection

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "Large Language Model Assisted Kernel Data Race Detection"

By

Mr. Qiao ZHANG


Abstract:

Data races in the Linux kernel remain a persistent security threat yet 
detecting them is notoriously difficult due to complex thread interleavings 
and deep execution paths. Traditional fuzzing often struggles to reach 
specific conflict points, while static analysis suffers from high false 
positive rates. We introduce Kendace, a novel framework that leverages 
Large Language Models (LLMs) to guide directed data race detection. Kendace 
operates in three phases: first, it employs module-aware static analysis 
with MemorySSA to efficiently identify potential conflict domains. Second, 
it utilizes LLMs to analyze path constraints and generate targeted C 
programs, successfully triggering deep kernel paths that evade traditional 
methods. Finally, a modified Kernel Concurrency Sanitizer (KCSAN) with 
precise thread synchronization verifies the races. Evaluation on the 
Mainline Linux Kernel and Distribution Kernels (e.g., EulerOS) demonstrates 
that Kendace effectively triggers difficult execution paths and successfully 
detect real-world vulnerabilities.


Date:                   Tuesday, 27 January 2026

Time:                   3:00pm - 5:00pm

Venue:                  Room 5501
                        Lifts 25/26

Chairman:               Dr. Dimitris PAPADOPOULOS

Committee Members:      Dr. Dongdong SHE (Supervisor)
                        Dr. Xiaomin OUYANG