黄恩浩
PhD Student at Zhejiang University
LLM agent security · Trustworthy AI
I am a PhD student in Computer Science and Technology at Zhejiang University, advised by Prof. Kui Ren (AAAS, ACM, CCF, and IEEE Fellow) at the State Key Laboratory of Blockchain and Data Security. I began my PhD in 2026. I received my B.E. in Information Security from Zhejiang University in 2026 (top 10%), and studied Innovation and Entrepreneurship Management at Chu Kochen Honors College.
My research focuses on LLM agent security and trustworthy AI: how agents can act reliably in real environments while remaining under human oversight. I collaborate with Zhan Qin and Yiming Li.
In 2024, I co-founded Hangzhou Zhulong Zhiyuan Technology Co., Ltd., where I work on bringing intelligent agents into practical use. We raised nearly RMB 5 million across pre-seed and seed rounds in 2025–2026.
I believe AGI should be trustworthy and accessible to everyone. This conviction guides both my research and the systems I build.
10 papers published or accepted at international conferences including CCS, KDD, WWW, IEEE S&P, and ICLR. A selection of my work is listed below.
CCS · First Author
A black-box method for auditing copyright infringement and tracing buyers of LLM agent skills. It embeds ownership evidence and buyer-specific fingerprints into structured skill documents, then verifies ownership and identifies buyers through paired probes and error-correcting decoding.
PaperCode
KDD · Dataset & Benchmark Track · First Author
A holistic benchmark for evaluating LLM capabilities across nine Web3 subdomains, combining objective knowledge assessment with open-ended reasoning tasks. It introduces a contamination-aware evaluation pipeline and analyzes the accuracy, reasoning reliability, and cost efficiency of 31 leading models.
WWW · Demo Track · Co-first Author
An LLM-driven multi-agent simulation framework that jointly models transient affective responses and persistent cognitive beliefs in public-opinion propagation. Evaluations on 15 real-world crises show that it reconstructs opinion trajectories more faithfully than existing simulation baselines.
ICASSP · First Author
A physics-inspired Transformer for complex-valued signals that explicitly incorporates relative phase and wave interference into self-attention. The architecture preserves amplitude–phase consistency and demonstrates strong performance on PolSAR classification and wireless-channel prediction.
Former Chief Scientist
I led post-training of the DMind model series and the development of DMind Benchmark.
DMind Benchmark previously reached No. 1 on the Hugging Face datasets leaderboard.
DMind Benchmark ↗Agent security evaluation
I contributed extensively to the writing of CAMEL’s research paper and was responsible for agent security evaluation.
National Natural Science Foundation of China · Young Student Research Program
Focusing on intrinsic security risks in LLM agents, this project develops a three-layer active defense framework spanning the full perception–reasoning–action pipeline.
SRTP National Innovation Project
This project developed domain-embedding-based model watermarking that reprograms a target model into an owner-controlled latent space. It preserves harmlessness while remaining robust against a range of model-removal attacks, and received the Best Creative Project award at the National College Student Innovation Annual Conference—one of only 12 projects nationwide.
Our paper “Trace: Optimizing Long-Context Agents via Task-Adaptive Information Extraction” received the Best Student Paper Award at AIRS 2026, held in conjunction with ICPP 2026 in Singapore.
I successfully passed the final project defense for the NSFC Young Student Research Program (Undergraduate).
As an undergraduate, I was invited to give a 15-minute talk on LLM agent security at ACM TURC 2026.
In 2024, I co-founded Hangzhou Zhulong Zhiyuan Technology Co., Ltd., where I built an enterprise-grade agent R&D and delivery team from the ground up and continue to explore pathways from research to real-world products.
Zhulong Technology ranked first in its group and joined the innovation showcase.
China International College Students’ Innovation Competition · National Gold Award
Related research outcomes have been widely covered by dozens of leading media outlets in China and abroad, including China Education News Network, Zhejiang Satellite TV, Hunan Satellite TV, Yahoo Finance, the Associated Press, and Caitong News.
PromptDefender, a platform for protecting the IP of system prompts.
Zhulong Technology ranked first in its group and joined the innovation showcase.
Led the development of a robust backdoor defense system.
Led the development of NetDP, a differential-privacy data processing system.
Awarded for “Trace: Optimizing Long-Context Agents via Task-Adaptive Information Extraction” at AIRS, held in conjunction with ICPP 2026, on September 28, 2026 in Singapore.
The model-watermarking project was selected among 12 national winners.
Recognized for academic excellence, research innovation, and student leadership.
Received the college’s highest student honor and nine university-level exemplary distinctions.
Invited reviewer
News photos carry your connection IP and generation time as a watermark. Watermark data is not stored separately.
