Diffusion models with continuous stochastic differential equations (SDEs) have shown superior performances in image generation. It can be used as a deep generative prior to solve the inverse problem in MR reconstruction. However, the existing VP-SDE can be treated as maximizing the energy of the MR image to be reconstructed and may lead to SDE sequence divergence. The VE-SDE based MR reconstruction is not consistent with actual diffusion process. In addition, both VE- and VP-SDEs-based models suffer from a time-consuming sampling procedure, resulting long reconstruction time. In this study, a new SDE focusing on the diffusion process in high-frequency space is designed specifically for robust MR reconstruction based on diffusion models. Experiments on the publicly fastMRI dataset show that HFS-SDE based reconstruction method outperforms the parallel imaging, supervised deep learning, and existing VE- and VP-SDEs-based methods in terms of reconstruction accuracy. It also improves the stability of MR reconstruction and accelerates sampling procedure of reverse diffusion.