About me

  • Hi! I am a junior year student studying computer engineering at Zhejiang University - University of Illinois Urbana Champaign Institute. I’m currently an exchange student at UIUC.

Education

  • Zhejiang University
    • Bachelor of Science in Computer Engineering (GPA: 3.85 / 4.00)
    • Rank: 5 / 65
    • Location: Zhejiang, China
    • Period: Aug 2021 - Present
  • University of Illinois Urbana-Champaign (double degree)
    • Bachelor of Science in Computer Engineering (GPA: 4.00 / 4.00)
    • Location: Illinois, US
    • Period: Aug 2021 - Present
    • Relevant Courses: Data Structure, Computer Systems Engineering, Applied Parallel Programming, Distributed Systems, Database Systems, Deep Learning, Computer Networks

Research Interests

  • I have a wide-ranging interest in generative models, with a particular emphasis on LLM security and interpretability

Publications

  1. Qiusi Zhan, Zhixiang Liang, Zifan Ying, Daniel Kang. “InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents.” Submitted to ACL 2024. arXiv:2403.02691
  2. Yidan Wang, Zhaochun Ren, Weiwei Sun, Jiyuan Yang, Zhixiang Liang, et al. “Content-Based Collaborative Generation for Recommender Systems.” arXiv:2403.18480

Research Experiences

  • InjecAgent: Exposing Vulnerabilities in Large Language Model Agents
    • Supervisor: Prof. Daniel Kang
    • Location: Illinois, US
    • Period: Jan 2024 – Present
    • Designed a benchmark to assess the vulnerability of tool-integrated Large Language Models (LLMs) to Indirect Prompt Injection (IPI) attacks. Our benchmark included 1,054 test cases across 17 user tools and 62 attacker tools, providing a comprehensive framework for evaluating agent resilience against IPI threats and categorized attack intentions into two primary types: direct harm to users and stealing of private data.
    • Analyzed the resilience of different LLM agents to IPI attacks and explores the impact of enhanced settings where attacker instructions are reinforced with hacking prompts. We found that fine-tuned agents, such as fine-tuned GPT-4, show significantly lower attack success rates compared to prompted agents, suggesting that fine-tuning may offer a more secure approach to deploying LLM agents.
  • Content-Based Collaborative Generation for Recommender Systems
    • Supervisor: Prof. Xin Xin
    • Location: Shandong, China
    • Period: June 2023 – Aug 2023
    • Proposed a generative recommendation model which unifies both item content information and user-item collaborative interaction signals in a sequence-to-sequence generation framework.
    • Contributed to manuscript refinement by complementing the ‘Related Work’ section to provide a comprehensive overview of current developments in generative models for recommendation.

Course Projects

  • CUDA-Optimized Forward Propagation of CNN Layers
    • Technologies: C++, CUDA, Nsight-Compute
    • Period: Nov 2023 - Dec 2023
    • Implemented forward pass of convolutional layers in LeNet-5 using CUDA-C++.
  • DZ OS: A Linux-like Operating System
    • Technologies: C, x86 assembly
    • Period: Oct 2023 - Dec 2023
    • Developed various kernel components and won 2nd place in the final competition.

Honors and Rewards

  • Zhejiang Provincial Government Scholarship (2023)
  • Third Class Scholarship of Zhejiang University (2022)
  • Dean’s List, UIUC (2022, 2023)

Technical Skills

  • Programming Languages: C, C++, Python, SQL, CUDA, x86 assembly
  • Frameworks/Tools: Git, SVN, Markdown, LaTeX, PyTorch, MySQL
  • Languages: Mandarin and English (TOFEL 97)