To address the significant risks of privacy leakage, various differential privacy techniques have been incorporated into deep learning models. However, these privacy-preserving methods often result in noticeable performance degradation. To balance privacy and utility, this paper proposes the Differentially Private with Decoupled Self-Knowledge Distillation (DPDSD) method, which effectively transfers high-usability knowledge to privacy-preserving deep networks during training. Specifically, DPDSD employs a teacher-student module designed with differential privacy: intermediate checkpoints serve as the teacher network, focusing on acquiring high-usability knowledge, while the student network emphasizes privacy protection via differential privacy stochastic gradient descent. Moreover, we decouple the distillation loss into two components: the loss for the target class and the loss for the non-target classes. This process adaptively adjusts targets by merging ground-truth labels with intermediate checkpoint predictions, enabling the model to progressively enhance its informativeness throughout training. Simultaneously, through refining knowledge from the teacher, the student achieves better performance while ensuring data privacy. Finally, extensive experiments on three public datasets have demonstrated that DPDSD can effectively improve model performance in the case of ensuring rigorous data privacy.