The rapid popularization of the Internet of Things has enabled the edge environment to obtain massive data resources. Usually, these data resources need to be further utilized with the help of cloud servers. But this will bring huge communication overhead and privacy security issues. In order to meet this challenge, we propose a safe and efficient data processing scheme based on cloud-edge-end collaboration for federated learning, which can use edge data resources for model training while ensuring low communication overhead and a reliable privacy and security environment. This scheme builds blockchains in sub-regions in the edge environment to ensure the safety and reliability of federated learning end devices. In addition, the solution adopts a layered computing strategy to parallelize the cloud-edge network and the edge-end network. At the same time, the cache pool mechanism is used to alleviate the communication congestion of asynchronous federated learning and the problem of outdated parameter versions, which improves the overall efficiency of the system. Finally, simulation experiments show that our scheme is very suitable for the cloud-edge-end environment, tolerates the heterogeneity and unreliability of terminal devices, and improves the efficiency of global model construction.