Federated learning has emerged as a promising solution to address privacy concerns in traditional machine learning models, which attracts significant attention from researchers. In this paper, we propose a federated learning optimization framework based on adaptive differential privacy (ALDP-FL) that dynamically adjusts privacy protection strategy to tackle key challenges in federated learning, such as accuracy degradation, difficulty in privacy quantification, and high communication overhead. We incorporate a local update mechanism with an adaptive gradient clipping strategy, while adding noise into the gradient transformation between clients and the central server. The impact of varying privacy levels on ALDP-FL model is analyzed on MNIST and CIFAR10 datasets. Simulation results demonstrate that ALDP-FL significantly enhances computational efficiency and reduces communication costs. Furthermore, the adaptive clipping threshold strategy maintains superior classification accuracy even under stringent privacy constraints and provides a balance between privacy and utility.