中央研究院 資訊科學研究所

Privacy-preserving deep learning computing has become popular these days as it can help protect both user data and deep neural network (DNN) model parameters at the same time with cryptographic techniques. In particular, significant efforts have been made to leverage secure two-party computation schemes for preventing user/model data from being disclosed during DNN inference. In this talk, I will share our experiences in building a compiler framework called TONIC that can automatically convert DNN models to one of two secure two-party computation language versions: ObliVM and ABY. Additionally, I will present the pre-computing scheme called POPS to accelerate the inference time by shifting the required communications from the time of execution to the time prior to execution.