Membership Inference
I study how training-data membership can be detected in fine-tuned language models, especially diffusion language models.
Ph.D. Student · AI-ISL Lab · Yonsei University
Yeachan Jun
I am a Ph.D. student at the Graduate School of Artificial Intelligence, Yonsei University, advised by Prof. Albert No. My research studies privacy and security issues in modern language models, with a focus on membership inference for diffusion language models and the conceptual foundations of machine unlearning in LLMs.
About
Researcher profile
I am currently a Ph.D. student at the AI-ISL Lab, Yonsei University. My work focuses on privacy risks, auditing methods, and evaluation problems that arise as language models move beyond standard autoregressive generation.
Recent projects include single-pass membership inference for fine-tuned diffusion language models and position work on how the term “machine unlearning” is used in the context of LLMs. Previously, I graduated early and at the top of my class from the Department of Mathematics at Kwangwoon University, with a GPA of 4.44/4.5.
Diffusion Language Models
My current work investigates privacy signals and efficient auditing methods for non-autoregressive language generation.
Machine Unlearning in LLMs
I am interested in clarifying what unlearning should mean for LLMs and how claims about unlearning should be evaluated.
Publications
Selected work
Recent papers on privacy, membership inference, diffusion language models, and machine unlearning in LLMs.
JUMP: Single-Pass Membership Inference on Fine-Tuned Diffusion Language Models
A membership inference study for fine-tuned diffusion language models, focusing on efficient single-pass privacy signals for identifying training-data membership.
Diffusion Language ModelsMembership InferencePrivacy AuditingFDGM Workshop
Position: The Term “Machine Unlearning” Is Overused in LLMs
A position paper arguing for more precise use of the term “machine unlearning” in LLM research and clearer evaluation of what unlearning methods actually achieve.
Machine UnlearningLLMsEvaluationPosition Paper
Research
Research directions
Single-Pass Privacy Auditing
Designing efficient membership inference signals that reduce the cost of auditing fine-tuned diffusion language models.
Privacy in Diffusion Language Models
Understanding how non-autoregressive generation changes the privacy risks and attack surfaces of language models.
Unlearning Terminology and Evaluation
Clarifying the conceptual gap between removing information, changing behavior, and claiming machine unlearning in LLMs.
Education
Academic background
Yonsei University
Ph.D. Student, Graduate School of Artificial Intelligence
Mar 2024 – Present · AI-ISL Lab · Advisor: Prof. Albert NoKwangwoon University
B.S. in Mathematics
Mar 2018 – Aug 2023 · GPA: 4.44/4.5 · Early graduation · Summa cum laudeHonors
Awards & honors
President's Award, LX Korea Land and Geospatial Informatix Corporation
First Prize, The Startup Competition for Big Data in Spatial Convergence · Sep 2024
Dean's List Academic Honors
Multiple semesters, including first-place distinctions in the department.
Contact
Get in touch
For research discussion, collaboration, or questions, feel free to contact me by email.