Publications

1. SelGrad: Selective Gradient Projection for Efficient Safety Alignment Against Harmful Fine-Tuning Wei Cheng Chiu, Shao-Jui Wang. IEEE Transactions on Dependable and Secure Computing (TDSC) — Under Review, 2026. A selective gradient-projection method for safety alignment that defends large language models against harmful fine-tuning, preserving alignment while keeping fine-tuning efficient.

Competitions & Technical Reports

• NICS Privacy-Enhancing Technologies (PETs) Competition — Award. Differentially-private synthetic-data generation built on MP-SPDZ (secure multi-party computation) with MWEM-style mechanisms.