Federated learning, as a machine learning paradigm that protects data privacy, requires participants to collaborate in training models without sharing raw data. Therefore, an efficient and secure data exchange mechanism is crucial. This paper proposes a federated learning data-sharing mechanism based on communication protocols, which combines communication protocols with federated learning algorithms to achieve key functions such as model training, data transmission, model aggregation, and privacy protection in a distributed system architecture. The mechanism innovatively adopts techniques such as protocol optimization, parallel chunked transmission, encryption authentication, and differential privacy noise injection, ensuring efficient model training, secure and reliable data exchange, and effective privacy protection. Comprehensive testing demonstrates that this data-sharing mechanism can meet the basic requirements of federated learning, to some extent protecting the privacy of participant data, and providing an effective technical solution for privacy computing and other fields. However, there is still room for optimization in communication efficiency and privacy protection algorithms.