Ph.D. Student Data Science William & Mary

Zhi Yang 阳芷

Trustworthy AI, multimodal learning and
interactive AI systems.

Portrait of Zhi Yang

About

My research centers on trustworthy AI, multimodal learning and interactive AI systems, with a focus on privacy, fairness and explainability. I study how privacy protection and its costs are distributed across groups — for example, whether differentially private training exposes some subgroups to higher privacy risk than others — and how to measure and mitigate such disparities. I am also interested in trustworthy large language models and vision-language models, including LLM personalization, multi-agent LLM systems, and how redundant visual information affects multimodal reasoning.

Before W&M, I received my M.Eng. in Computer Science from the Southern University of Science and Technology (SUSTech), where I was advised by Dr. Changwu Huang, Prof. Ke Tang and Prof. Xin Yao, and my B.Eng. in Software Engineering from Yunnan University.

Research interests

Core

  • Trustworthy AI
  • Multimodal learning
  • Interactive AI systems

Also

  • Differential privacy & privacy auditing
  • Algorithmic fairness
  • Explainable AI
  • Trustworthy LLMs
  • Vision-language models
  • LLM personalization
  • Multi-agent LLM systems

Publications

  1. Preprint2025

    On the Fairness of Privacy Protection: Measuring and Mitigating the Disparity of Group Privacy Risk for Differentially Private Machine Learning

    Zhi Yang, Changwu Huang, Ke Tang, Xin Yao

    arXiv preprint arXiv:2510.09114, 2025

    Introduces a membership inference game that efficiently audits worst-case privacy risk to measure privacy-risk disparities across groups, and an adaptive group-specific gradient clipping strategy for DP-SGD that reduces them.

    arXivPDF

  2. ICONIP 2024

    Towards Private and Fair Machine Learning: Group-Specific Differentially Private Stochastic Gradient Descent with Threshold Optimization

    Zhi Yang, Changwu Huang, Xin Yao

    International Conference on Neural Information Processing (ICONIP), 2024, pp. 66–80. Springer Nature Singapore.

    A group-specific DP-SGD training framework combined with classification threshold optimization that jointly reduces accuracy disparities and achieves outcome fairness under differential privacy.

    PDFSlidesCode

  3. ICANN 2023

    An Explainable Feature Selection Approach for Fair Machine Learning

    Zhi Yang, Ziming Wang, Changwu Huang, Xin Yao

    International Conference on Artificial Neural Networks (ICANN), 2023, pp. 75–86. Springer Nature Switzerland.

    ExFS improves outcome fairness by recursively removing features that contribute to unfairness, guided by feature attribution explanations of model predictions.

    PDFSlides

Research projects

Education

Honors & awards

Teaching