Yurong Hao

We propose a noval fine tuning framework that uses reinforcement learning and adapter compression to selectively shield model layers in TEEs, which achieves strong model protection with 2–3× efficiency gains and no accuracy loss.