Abstract: When users’personal sensitive data are used to build deep learning models, some sensitive data will be " remembered " by the network, resulting in privacy leakage. This paper proposes DP-Adam by introducing differential privacy into the Adam algorithm, drawing inspiration from the DP-SGD algorithm which incorporates differential privacy into SGD. The proposed approach develops a power-function-based adaptive privacy budget allocation mechanism to achieve more rational privacy budget distribution, thereby better balancing privacy preservation and model accuracy. The experimental results show that the DP-Adam algorithm in this paper has better privacy budget than the traditional DP- SGD algorithm. The experimental results show that this paper’s DP-Adam algorithm has better accuracy than the traditional DP-SGD algorithm under the same privacy budget, about 5% higher in the case of low privacy budget and about 2% higher in the case of medium-high privacy budget. Comparative analysis between the power-function-based allocation algorithm and the exponential-function-based allocation algorithm demonstrates that the former exhibits superior adaptability.