M.S. student in Artificial Intelligence, Sungkyunkwan University IRIS Lab, advised by Jong Hwan Ko
Under review at ICLR 2027
Removes a speaker from a zero-shot TTS model in closed form — no retraining — so repeated removal requests compose exactly instead of degrading the model.
Industry project with SEMES · IRIS Lab, Sungkyunkwan University
Semiconductor inspection has almost no defect images to train on. I built a framework that synthesizes them at the feature level — fusing continuous memory representations of normal and anomalous features, then guiding defect placement with a detector's reconstruction residuals — and studied training under severe class imbalance. Evaluated on MVTec AD.
Code·built on CRAD (ECCV 2024)
Industry–academia project with Samsung Research · IRIS Lab, Sungkyunkwan University
Studied direction-of-arrival estimation for multi-channel audio, and trained AASIST-based detectors of synthetic speech on ASVspoof 5 and ASVspoof 2019, measuring how far they generalize to deepfake and TTS systems never seen in training.