Title: BumbleBee: Secure Two-party Inference Framework for Large Transformers
Abstract: We introduce BumbleBee, a two-party private inference system designed for efficiency and speed. Key contributions include optimized matrix multiplication protocols reducing communication costs by 80-90%, new protocols for non-linear activation functions improving speed and reducing costs by 80-95%. BumbleBee outperforms previous systems, Iron and BOLT, by a significant margin in both speed and communication efficiency.
