Federated learning (FL) has emerged as a representative paradigm for providing distributed AI services, allowing decentralized clients to jointly train models while keeping raw data localized. Although FL reduces privacy risks associated with data transmission, it does not fully prevent information leakage through shared model updates. To provide robust privacy protection, differential privacy (DP) is employed, and the integration of DP with FL has become a mainstream approach for achieving rigorous privacy guarantees in distributed learning. Differentially private stochastic gradient descent (DP-SGD) is widely used, but the addition of noise often leads to notable accuracy loss in practical, especially under strict privacy budgets. To address these challenges, we propose Differential Privacy with Dynamic Threshold and Gradient update (DP-DTG), a novel DP-based privacy preservation mechanism for distributed services. It introduces dynamic gradient clipping with nonlinear threshold decay and adaptive momentum updates, jointly reducing unnecessary noise and improving the balance between privacy and model utility. Experiments on MNIST and CIFAR-10 datasets demonstrate that DP-DTG consistently outperforms representative baselines in accuracy under equivalent privacy constraints.