Sehwan Kim

Hello! I am a Ph.D student in Electrical and Computer Engineering at Seoul National University, advised by Jungwoo Lee. Before my doctoral studies, I completed my B.S. in Electrical and Computer Engineering at Seoul National University.

From September 2026 to March 2027, I will be a Research Intern at Sony AI in Tokyo, Japan, conducting research on Generative and Protective AI for Content Creation.

If you'd like to discuss my work or potential collaborations, feel free to reach out!

Email  /  LinkedIn  /  Scholar

Research (Research Statement)

My research focuses on developing robust and trustworthy AI systems. Currently, I am particularly interested in understanding and mitigating memorization in diffusion models, as well as developing unlearning and inference-time mitigation techniques for preventing privacy leakage and NSFW content generation. My past research contributions include:

• Explainable AI (XAI): Designing interpretable models and visual reasoning frameworks for understanding deep neural networks.
• Bias mitigation: Developing algorithms to detect and reduce spurious correlations in vision and language models.
• Trustworthy generation: Improving the robustness of text-to-image diffusion models by addressing memorization and preventing unintended content reproduction.

This work proposes STEP-KTODER, which decomposes generated programs into testable functions and combines execution-derived function-level correctness labels with outcome-level KTO to improve code preference optimization.

This work proposes leveraging pruning-induced confidence dynamics to identify bias-conflicting samples and train sparse neural networks that mitigate spurious correlations without requiring group annotations.

This paper introduces a training-free diffusion-based dataset distillation framework that boosts both fidelity and diversity through adaptive sampling and repulsion regularization.

This work proposes adjusting the initial noise sample in text-to-image diffusion models to promote earlier escape from memorization basins, thereby reducing training-data replication while preserving prompt alignment.

A method is introduced to reshape a pretrained generative model’s latent space into a fair and disentangled representation, enabling counterfactual explanation and improved fairness.

This work introduces a bias-label-free debiasing method that upsamples samples with high target–biased-model disagreement to reduce spurious correlation dependence.

We propose injecting instance-dependent multi-label noise into remote sensing image datasets, creating a more realistic and challenging noise setting.

Template based on Jon Barron's website.