B.Sc

Education

  • 2024.08 - 2024.12

    Montreal, Quebec, CAN

    Visiting Student
    McGill University
    Computer Science
  • 2021.09 - 2025.06

    Kowloon, Hong Kong

    Bachelor of Science (Hons)
    Hong Kong Baptist University
    Computer Science & Applied Mathematics

Research

  • 2025.02 - 2025.05
    Research Assistant
    Dept. of Computer Science, HKBU
    Analyzed gradient-manifold alignment in model inversion attacks and supported NeurIPS 2025 experiments.
    • Advised by Bo Han.
  • 2024.08 - 2024.10
    Research Assistant
    Dept. of Computer Science, HKBU
    Studied optimization instability and name-image matching issues in multimodal model evaluation.
    • Advised by Bo Han.
  • 2024.06 - 2025.05
    Undergraduate Research Project
    Dept. of Computer Science, HKBU
    Developed a final-year thesis framework for robust training with OOD detection and confidence calibration.
    • Advised by Bo Han.
  • 2024.05 - 2024.08
    Undergraduate Research Assistant
    Research Office, HKBU
    Studied shortcut learning in robust training and ran summer research experiments in the TMLR group.
    • Advised by Bo Han.
  • 2024.02 - 2024.06
    Student Research Assistant
    Dept. of Computer Science, HKBU
    Implemented experiments for pseudo-private data guided model inversion, contributing to a NeurIPS 2024 paper.
    • Advised by Bo Han.
  • 2023.05 - 2023.08
    Research Assistant
    Dept. of Computer Science, HKBU
    Surveyed model inversion attacks and defenses, contributing to an undergraduate research survey.
    • Advised by Bo Han.

Work

  • 2025.01 - 2025.04
    Student Helper
    Dept. of Computer Science, HKBU
    Tutored COMP1005 students in Python programming and computational thinking.
  • 2024.01 - 2024.04
    Student Helper
    Dept. of Computer Science, HKBU
    Supported COMP1005 and COMP2016 tutorials on Python, Java, SQL, and databases.
  • 2023.01 - 2023.04
    Student Helper
    Dept. of Computer Science, HKBU
    Tutored COMP1005 students in Python programming and computational thinking.

Mentor

  • 2023.08 - 2023.11
    Peer Mentor 2023
    Hong Kong Baptist University
    Mentored first-year students on academic adjustment and university life.
  • 2023.08 - 2024.04
    Mental Health First Aid Faculty Captain
    Faculty of Science, HKBU
    Supported Science students with mental health first-aid guidance and referrals.
  • 2023.08 - 2024.04
    Peer Tutor
    Dept. of Computer Science, HKBU
    Tutored junior CS students and advised them on coursework and academic planning.
  • 2022.08 - 2022.11
    Peer Mentor 2022
    Hong Kong Baptist University
    Mentored first-year students on academic adjustment and university life.

Awards

Certificates

Mental Health First Aid (MHFA) Standard Course
The Mental Health Association of Hong Kong 2023-08
Study Tour to Singapore
Dept. of Computer Science, HKBU 2023-05

Publications

  • 2025.09
    Generative Model Inversion Through the Lens of the Manifold Hypothesis
    NeurIPS 2025
    This paper explains how generative priors improve model inversion by projecting noisy inversion gradients onto the tangent space of the generator manifold. We analyzed gradient-manifold alignment and developed an alignment-promoting training objective and a training-free correction for inversion.
  • 2024.12
    Pseudo-Private Data Guided Model Inversion Attacks
    NeurIPS 2024
    This paper introduces a model inversion attack that uses pseudo-private data to approximate the target model's private training distribution and reconstruct its training samples more accurately. We implemented the training pipeline and conducted the main experiments and ablation studies.
  • 2024.11
    Model Inversion Attacks: A Survey of Approaches and Countermeasures
    arXiv preprint arXiv:2411.10023
    This survey reviews model inversion attacks and defenses across image, text, and graph domains. We categorized attacks by threat model and methodology, synthesized defenses and evaluation protocols, and curated the accompanying repository.
  • 2024.09
    Single-Step Diffusion Model-Based Generative Model Inversion Attacks
    Submitted to ICLR 2025
    This work uses a distilled single-step diffusion generator to reconstruct private training samples in model inversion attacks. We implemented the diffusion-based attack pipeline and evaluated privacy leakage in multimodal encoders, including CLIP.