Yanbo Dai
 |
Yanbo Dai, Ph.D. Student |
About me
I am a first-year Ph.D. student at the CSE department of HKUST supervised by Prof. Shuai Wang. Before that, I received my Mphil degree from the IoT Thrust, Information Hub under the supervision of Prof. Songze Li. I received my bachelor degree of electronic information engineering from the Excellence class, School of Microelectronics and Communication Engineering, Chongqing University, where I worked with Prof. Hailin Cao.
Research
My research interests include
AI Security with a recent focus on LLMs
Model Editing
Backdoors Attacks & Defenses in Federated Learning
My research journey began with developing secure federated learning (FL) systems, which are especially resilient to backdoor attacks. For attack algorithms, I investigated how adversaries could exploit sample relationships to inject more durable backdoors (ICML’23). For defensive mechanisms, I proposed proactive defenses via out-of-distribution data, which has stronger defense performance under strong non-IID settings or long-term continusous injection scenarios. After receiving client updates, the server can either employ indicator-based screening (Usenix Security’24) or direct aggregation (arXiv’25) to effectively eliminate backdoors.
During my PhD, my focus shifted to enhancing the security and reliability of large language models (LLMs). I see model editing as a powerful technique to hack or safeguard LLMs with minimal computational and data overhead. My work identifies root causes behind the failures of existing model editing methods in large-scale and context-rich settings and proposes targeted countermeasures (arXiv’25).
I am committed to advancing secure and trustworthy LLM systems. Feel free to reach out if you’re interested in discussing my research.
News
[05-30-2024] One paper is accepted by Usenix Security’24!
[04-25-2023] One paper is accepted by ICML’23!
Education
Ph.D. in Computer Science Engineering, HKUST, Aug. 2024 -
M.phil. in Internet of Things, HKUST(GZ), Sep. 2022 - Jun. 2024
B.Eng. in Electronic Information Engineering, Chongqing University, Sep. 2018 - Jun. 2022
Experience
Research Assistant, School of Cyber Science and Engineering, Southeast University, hosted by Prof. Songze Li, Jun. 2024 - Aug. 2024
Summer Research Internship, Department of EEE, The University of Hong Kong, hosted by Prof. Y.C. Wu, Jun. 2021 - Aug. 2021
Selected Publications (Full at Google Scholar)
(* indicates equal contribution)
NAMET: Robust Massive Model Editing via Noise-Aware Memory Optimization
Yanbo Dai, Zhenlan Ji, Zongjie Li, Shuai Wang
Preprint
TrojanDam: Detection-Free Backdoor Defense in Federated Learning through Proactive Model Robustification utilizing OOD Data
Yanbo Dai, Songze Li, Zihan Gan, Xueluan Gong
Preprint
BackdoorIndicator: Leveraging OOD Data for Proactive Backdoor Detection in Federated Learning
Songze Li*, Yanbo Dai*
Usenix Security ’24.
Chameleon: Adapting to Peer Images for Planting Durable Backdoors in Federated Learning
Yanbo Dai, Songze Li
ICML’23
Deep Domain Adaptation for Predicting Intra-AbdominalPressure with Multichannel Attention Fusion Radar Chip
Hao Tang*, Yanbo Dai*, Dongchu Zhao, Zhiwei Sun, Fuqiang Chen, Yiliang Zhu,Huaping Liang, Hailin Cao, and Lianyang Zhang
Advanced Intelligent Systems.