Abstract: In order to give consideration to privacy protection and data utility at the same time, a privacy protection publishing method of smart meter based on specific information upper bound is proposed. A privacy protection release training function based on the upper bound of specific information was introduced. In order to minimize the KL divergence between the real distribution and approximate distribution of sensitive variables, a relaxation formula suitable for the original problem of deep learning framework was introduced. In addition, deep cyclic neural network was used to capture and utilize the time correlation of smart meter signals. Experiments show that the proposed method can reduce the distortion of specific privacy targets and effectively protect privacy.