边缘计算下基于Transformer的安全高效自然语言处理

riven by the encoder-decoder architecture, Transformer has extended its applications into natural language processing (NLP). With the increase of model parameters, inference services in cloud and edge platforms have emerged. However, when users utilize these Transformer models, they need to expose their private inputs. Some efforts have beenmade to address the privacy issues for Transformer-based inference services, but their inference efficiency has been notably low. To this end, we propose an efficient three-party transformer-based NLP system in edge computing, TFCrypt, by combining additive secret sharing and function secret sharing.This paper devises an efficient secure embedding protocol by generating the offset tuple in the offline phase using function secret sharing and computing the embedding item during the interaction phase using dot products. In addition, a number of protocols are designed to optimize the inference efficiency. Finally, the paper conduct a series of experiments in both LAN and WAN environments, showing that this paper improves the processing efficiency by 1.53 ~ 33.75 times in LAN and 1.85 ~ 33.97 times in WAN compared to existing schemes.