Haonan Shi

I am a fourth-year Ph.D. candidate at Case Western Reserve University, advised by Prof. An Wang and co-advised by Dr. Tu Ouyang. Prior to this, I received my bachelor's degree from South China University of Technology, where I conducted research on privacy under the supervision of Prof. Hongyun Xu.
My research focuses on Large Language Model Safety, Agent Safety and Machine Learning Privacy. I am broadly interested in developing methods that make LLM/AI systems safer and more privacy-preserving as they are adapted and deployed in real-world settings.
I am passionate about applying my research to real-world problems and open to internship opportunities and research collaborations.

Email: haonan.shi3[AT]case.edu (please replace "[AT]" by "@")

Email  /  Google Scholar  /  LinkedIn

We propose PACT, a token-level constrained fine-tuning framework that preserves safety alignment of large language models by regularizing safety-critical tokens while allowing other general tokens to adapt to downstream tasks.

We propose EASE, a practical safety alignment post training framework for small language models that selectively activates safety reasoning for adversarial jailbreak queries while preserving efficiency for benign interactions.

We propose GuardedTuning, a set of privacy-preserving fine-tuning designs that navigate trade-offs among model utility, privacy guarantees, and tuning costs for large language models.

We reveal privacy risks in public dataset-assisted federated distillation and introduce attacks that infer clients' label distributions and membership information from their predictions on public datasets.

We propose LDC-MIA, a learning-based difficulty calibration method that improves membership inference attacks at low false positive rates while maintaining practical attack costs.

We propose KPDP, a trajectory k-anonymity model based on point density and partition, to protect trajectory datasets against re-identification attacks while reducing data utility loss.

• Journal Reviewer: Pattern Recognition(2026)
• Conference Reviewer: KDD(2026), ICLR FM-Wild(2026), ICLR FM4Sci(2026)

Outside of research, I love to spend time with Yazai, Meituo, and Tuotuo, who keep my life warm and lively.

Yazai & Meituo

Tuotuo