Differential privacy is a widely used privacy protection mechanism that provides strong privacy guarantees for data analysis. Typically, the Laplace mechanism is an essential method to implement differential privacy. However, the randomness nature of noise produced by the Laplace mechanism inevitably not only leads to low data quality but also result in weak performance of relevant data mining tasks. To overcome the aforementioned challenges, this paper attempts to enhance the efficiency of data mining procedures by a noise redistribution stage, which aims at minimizing the interval frequency changes between perturbed data and raw data via reallocating the noise generated by Laplace mechanism to raw data. To guarantee the optimality, we formulate the noise redistribution stage as a mixed integer programming model (Lap-MIP). Numerical experiments reveal that the performance of data mining tasks is significantly improved after embedding the noise redistribution stage in the classification framework, compared to directly adding the Laplace noise to raw data.