Wenyun Li
About Me
I am a Ph.D. candidate in Electronic Information at Harbin Institute of Technology (Shenzhen) and Peng Cheng Laboratory. My research focuses on trustworthy machine learning and AI safety, with particular interests in adversarial learning, jailbreak attacks on large language models, hallucination detection and mitigation, and safety for agentic AI systems.
Education
- Ph.D. Candidate in Electronic Information, Harbin Institute of Technology (Shenzhen) and Peng Cheng Laboratory, Sep. 2023 - Present
- M.Sc. in Computer Science, University of Macau, Sep. 2020 - Jan. 2023
- B.Sc. in Electronic Information, University of Electronic Science and Technology of China, Sep. 2016 - Sep. 2020
Research Interests
- AI Safety: adversarial learning, jailbreak attacks, constitutional refusal, agentic AI safety
- Trustworthy Machine Learning: hallucination detection and mitigation, robust and reliable ML systems
- Multimodal Learning: LLMs, MLLMs, and safety evaluation across vision-language settings
Selected Highlights
- [2026] “Bolster Hallucination Detection via Prompt-Guided Data Augmentation” is accepted to AAAI 2026.
- [2025] Two papers on adversarial attack and healthcare image retrieval are published in AAAI 2025 and IEEE Transactions on Consumer Electronics.
- [2023] Published works in BIBM, ICASSP, and ICIP on multimodal fusion and cross-modal retrieval.
Publications
-
Latent Safety Boundary Probing for Automated Jailbreak Attacks on Large Language ModelsIEEE Transactions on Information Forensics and Security, under review.Under Review
-
Direct Alignment Is Insufficient: Architecting Constitutional Refusal in Agentic AIAdvances in Neural Information Processing Systems (NeurIPS), 2026, under review.Under Review
-
Bolster Hallucination Detection via Prompt-Guided Data AugmentationProceedings of the Fortieth AAAI Conference on Artificial Intelligence (AAAI), 2026.
-
Untargeted Closed-Box Attack Against Healthcare Image Retrieval via Rank ManipulationIEEE Transactions on Consumer Electronics, 2025.JCR Q1, IF: 10.9
-
Transferable Adversarial Face Attack with Text Controlled AttributeProceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI), 2025.
-
ELF: An End-to-end Local and Global Multimodal Fusion Framework for Glaucoma GradingProceedings of the IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2023.
-
Locality Preserving Multiview Graph Hashing for Large Scale Remote Sensing Image SearchProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.
-
Asymmetric Scalable Cross-Modal HashingProceedings of the IEEE International Conference on Image Processing (ICIP), 2023.
-
A Single-Target License Plate Detection with AttentionInternational Workshop on Advanced Imaging Technology (IWAIT), 2022.
Services
Reviewing
- Reviewer: CVPR 2023-2025; ACM MM 2023-2024; NeurIPS 2024-2025; ICASSP 2023-2025; AAAI 2024-2026
- Reviewer: IEEE Transactions on Information Forensics and Security; Information Fusion; IEEE Journal of Biomedical and Health Informatics
- Area Chair Reviewer: IJCNN 2025-2026
Awards
- Outstanding Class Cadre, 2024 - 2025
- Peng Cheng Director Scholarship, 2023 - 2024
Patents
- A keyness-aware fine-tuning based adversarial training method and device, Application No.
2026100261860, status: pending
Powered by Jekyll and Minimal Light theme.