Google Scholar I am a PhD student in EE at KAIST working on Multi-Modal Models and Large Language Models. As AI’s role expands across various fields, building trustworthy AI is crucial. I am particularly interested in AI security, as vulnerabilities compromise reliability and fairness.
Additionally, as humans rely on various senses for judgment, I focus on research in Multi-Modal Learning and AI Security. My goal is to help bridge the understanding between AI and humans.
My work explores the following, but not limited to:
- Multimedia Forensics: detecting forgeries, synthetic image, and deepfakes
- AI Security: adversarial attack, jailbreaking and their defense
- Multi-Modal Learning: multi-modal large language models, efficient multi-modal learning
dpenguin [at] kaist.ac.kr
Publications
Conference Papers
AIM: Anchor Identity Features, then Match for Multimodal Large Language Model Unlearning
Wonjun Lee*, Jaehyuk Jang*, Kangwook Ko*, Hee-Seon Kim, Changick Kim (* indicates equal contribution)
The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP findings), 2026
Where Identity Lives: Localized, Retain-Free Identity Unlearning in Multimodal Large Language Models
Kangwook Ko*, Jaehyuk Jang*, Wonjun Lee*, Hee-Seon Kim, Changick Kim (* indicates equal contribution)
The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP findings), 2026
Constraining to Generalize: Subspace Tuning for Few-shot Generalization of Audio-Language Models
Jaehyuk Jang, Kangwook Ko, Wonjun Lee, Changick Kim
The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP findings), 2026
Safety in Batches? Understanding and Mitigating Safety Failures in Batch Prompting
Kihyun Kim, Hee-Seon Kim, Wonjun Lee, Changick Kim
arXiv preprint, 2026
SELFI: Selective Fusion of Identity for Generalizable Deepfake Detection
Younghun Kim, Minsuk Jang, Myung-Joon Kwon, Wonjun Lee, Changick Kim
arXiv preprint, 2025
Benign-to-Toxic Jailbreaking: Inducing Harmful Responses from Harmless Prompts
Hee-Seon Kim, Minbeom Kim, Wonjun Lee, Kihyun Kim, Changick Kim
arXiv preprint, 2025
Optimization-based jailbreaking that induces safety misalignment from benign conditioning prompts.
Others
Research Experience
Roen Surgical
KAIST EE Externship Program Second Cohort
SAMSUNG Electronics CE/IM Division Mobile Communications Unit
SUMMER INTERNSHIP
SAMSUNG Electronics DS Division Foundry Business Unit
SAMSUNG TALENT INTERNSHIP PROGRAM (STIP)
Patents
A Universal Image-Generation Framework for Jailbreaking Large Vision–Language Models by Bypassing Safety Alignment
- Participated as a Key Developer in Patent Development
- Participated through the Project with ETRI
Method and System for Generating Universal Adversarial Perturbations Using High-Sensitivity Components of Vision Encoders in Large-Scale Vision-Language Models
- Participated as a Key Developer in Patent Development
- Participated through the Project with IITP
Stone size estimation method
- Participated as a Key Developer in Patent Development
- Participated through the KAIST EE Externship
Projects
Automatic Evaluation Benchmark Generation from LLM Log Data
Development of AI Technology with Robust and Flexible Resilience Against Risk Factors
- Team Leader
Penetration Security Testing of ML Model Vulnerabilities and Defense
Scene Text Recognition with Visual Contexts
Practical Adversarial Attacks of AI Facial Recognition Technology Using Physical Patterns
Bypass Techniques for Identifying Vulnerabilities in CAPTCHA