Abstract: Aiming at the problem of effectively protecting user privacy data,  we proposed a privacy-preserving logistic regression training scheme based on two-party secure computation  to complete the  joint modeling work of multiple data parties.  Firstly, the scheme  optimized the  generation process of the multiplicative triplet to reduce the time required in  the offline phase. Secondly, we replaced  the activation functions that were difficult to calculate in secure multi-party computation with approximate functions. 
Finally, we vectorized the proposed protocols  and accelerated the local matrix computation by using  CUDA (compute unified device architecture).  The experimental results of using different datasets to test the privacy-preserving logistic regression  performance in both local and wide area networks show that the scheme can enable the model to converge in a  short time and  increase the possibility of solving privacy-preserving machine learning related problems in real world.

Key words: secure multi-party computation, privacy-preserving machine learning, garbled circuits, oblivious transfer