<p>Traditional applications of differential privacy (DP) in federated learning (FL) usually use a fixed privacy budget. However, this approach has significant limitations when handling data with varying sensitivity between different nodes. This article introduces a novel dynamic privacy budget allocation (DPBA) algorithm to address these shortcomings and improve the effectiveness of DP in FL. The DPBA algorithm considers factors such as node data sensitivity, participation in budget allocation, participation frequency, and the global model convergence state. By dynamically adjusting the privacy budget, DPBA not only increases the flexibility of privacy protection, but also enhances efficiency within the FL framework. This is achieved by allocating the privacy budget based on the data characteristics of each node and the training progress of the model, thus providing stronger protection for sensitive data while minimizing privacy leakage and performance degradation. This paper details the design principles and implementation steps of the DPBA algorithm and validates its applicability and effectiveness through extensive experiments on various datasets. The results show that DPBA not only provides robust privacy guarantees for individual nodes but also substantially improves classification accuracy compared to a fixed-budget approach. These findings underscore the promise of DPBA for federated learning and offer new insights into applying differential privacy in increasingly complex data settings.</p>