报告人
Zongjie Li
Hong Kong University of Science and Technology
时间
2025年9月16日 星期二
下午 14:00-15:00
地点
602会议室
Abstract
The rapid development of Large Language Models (LLMs) has introduced critical challenges in ensuring their security and reliability. We present a systematic approach to enhancing LLM safety, centered on the philosophy of "defining consistency," which involves establishing clear, measurable criteria for reliable behavior at each phase of the LLM lifecycle: training, testing, and deployment. We address key issues such as the scarcity of high-quality training data, the inherent biases in LLM-based evaluation, and the significant intellectual property risks during deployment.
Our research provides novel solutions across these stages, including a method for controllable training data synthesis using compiler techniques, a "split-and-merge" framework to mitigate position bias in LLM evaluators, and the first watermarking scheme for code models to protect intellectual property by ensuring functional consistency. Additionally, we introduce a pioneering technique to extract proprietary training data by identifying behavioral differences between base and fine-tuned models. These contributions systematically advance the development of more robust and secure LLM systems.
Biography
Zongjie Li is a fourth-year Ph.D. candidate in the School of Computer Science and Engineering at the Hong Kong University of Science and Technology, supervised by Professor Shuai Wang. He graduated from the Harbin Institute of Technology, Shenzhen, where he earned his bachelor's degree, ranking first in his class in 2021.
During his doctoral studies, Zongjie was a visiting researcher at ETH Zürich, working with Professor Zhendong Su. His primary research focuses on the security oflLarge language models systems.
He has published 23 papers in top-tier international security and software engineering conferences, including CCS, USENIX Security, ICSE, FSE, ISSTA, and ASE. Among these, eight are first-author papers in CCF-A ranked conferences.
Zongjie's research has received several honors, including a Best Paper Award Nomination at ICSE 2022 and the Best Paper Award for the Industrial Challenge at ASE 2023. Furthermore, he was the runner-up in the IEEE (Hong Kong) Computational Intelligence Chapter Competition.
