Academic homepage
Undergraduate student in Electronic and Computer Engineering
Shenzhen MSU-BIT University · Shenzhen, China
I study trustworthy AI and information security, with a focus on privacy in document multimodal models, agentic search security, LLM uncertainty, and the safety of mental-health dialogue systems.
I am also president of the university's Computer Association. My security work includes binary vulnerability exploitation, IoT security, reverse engineering, and vulnerability reporting.
Recent research and milestones
“Backdoor in the Loop,” our study of malicious retrievers in agentic search, is under review at ICLR 2027. Read the paper
Our work “Beyond Visual Evidence” was accepted as a poster at ACM Multimedia 2026. Paper, code, and poster
“Auditing Implicit Sycophancy in Mental-Health Dialogue” appeared in the EI-MiGA workshop proceedings at IJCAI 2026.
TECP: Token-Entropy Conformal Prediction for LLMs was accepted by Mathematics.
TECP was released on arXiv.
A remote code execution vulnerability was collected by the CNVD Vulnerability Database.
I completed my internship as an IoT Security Engineer at Hillstone Networks Co., Ltd.
A remote code execution vulnerability was collected by CNNVD.
Research papers and poster
ICLR 2027 Under Review · Agentic Search Security
This work exposes a systematic vulnerability in agentic RAG: a compromised retriever can exploit the search feedback loop to suppress useful evidence, repeatedly surface a chosen document, or prolong search, even when the corpus and agent remain unchanged. It also shows how an inject-and-remove cycle can conceal the malicious behavior from backdoor detectors.
ACM MM 2026 Poster · Accepted July 10, 2026
Document MLLMs can reveal correlated private fields even when visual evidence is weak. We study relation-level privacy leakage and introduce DRUF, a dynamic unlearning method that suppresses exposed field pairs while preserving key information extraction.
EI-MiGA @ IJCAI 2026 Workshop · Springer CCIS
This paper audits a hidden dialogue risk: a response may sound empathetic while reinforcing harmful cognition. It introduces a clean matched benchmark and Dynamic Emotional Signature Graphs (DESG) to evaluate clinical-state changes rather than surface wording alone.
Vulnerability analysis and reporting
Alongside AI research, I investigate deployed systems through binary analysis, reverse engineering, and IoT device security.
CVE-2024-52675 · SQL injection
A SQL injection vulnerability discovered and reported in the movie rating prediction system.
Education, experience, and recognition
B.Eng. in Electronic and Computer Engineering.
IoT Security Engineer Intern. Worked on device vulnerability research, CNVD and CNNVD submissions, vulnerability technology tracking articles, and product vulnerability reports.
Internship certificate ↗President, Shenzhen MSU-BIT University.
2024 CCPC Guangzhou Invitational Contest.
Security research notes and technical articles.