About Me

I am a final year Ph.D. student in the Department of Computer Science at Rice University. My advisor is Prof. Xia Hu, who leads the DATA Lab and Rice and Rice D2K Lab. I obtained my bachelor’s degree from the Department of Automation, Tsinghua University, advised by Prof. Jiwen Lu.

My primary research interests lie in the domain of Trustworthy AI, with an emphasis on robustness, privacy, security, explainability, and fairness. I have particularly explored these facets within the context of healthcare applications. 

My recent work pivots towards investigating the intersection of trustworthy machine learning and large language models (LLMs). This includes regulatory considerations for powerful LLMs, as well as strategies for mitigating potential risks related to privacy and social bias in these foundational models. 

I’m graduating in Fall 2024 and am actively seeking a tenure-track faculty position. 

Research Experiences

• 09/2023: Two papers have been accepted by NeurIPS 2023. We proposed a honeypot mechanism to defend against backdoor attacks in language models.

• 09/2023: Our paper "The Science of Detecting LLM Generated Texts" has been accepted by the Communications of the ACM. We provide an overview of existing LLM-generated text detection techniques and enhance the control and regulation of language generation models. 

• 08/2023: One paper has been accepted by ECML-PKDD 2023. We proposed a serial key protection mechanism for safeguarding DNN models.

• 08/2023: Two papers have been accepted by CIKM 2023. We proposed a transferable watermark for defending against model extraction attacks.

• 07/2023: Three Papers have been accepted by AMIA 2023 Annual Symposium. We investigated methods for harnessing the capabilities of LLMs in several healthcare tasks, such as Named Entity Recognition, Relation Extraction, and Patient-Trial Matching. 

• 05/2023: One paper has been accepted by ACL 2023. We reveal that LLMs are "lazy learners" that tend to exploit shortcuts in prompts for downstream tasks. Additionally, we uncover a surprising finding that larger models are more likely to utilize shortcuts in prompts during inference. 

• 04/2023: One paper has been accepted by SIGKDD Exploration 2023. We proposed a clean-label backdoor-based watermarking framework for safeguarding training datasets. 

• 03/2023: One paper has been accepted by ICHI 2023. We proposed a deep ensemble framework for improving phenotype prediction from multi-modal data.