Differential privacy, homomorphic encryption, and multi-party security computing can protect exchanged data during federated learning. However, they fail to focus on accuracy, training time, or communication traffic while ensuring security. Therefore, we proposed a federated learning scheme for hierarchical protection and multiple aggregation. Firstly, the client contribution is calculated and compared with the threshold. The local model update is discarded, disturbed, or encrypted according to the comparison result. Secondly, the server adjusts the selection weights of clients and aggregates the disturbed and the encrypted local model updates respectively. Finally, the client gets the final global model by decryption and aggregation. Compared with other schemes, the scheme shows the highest accuracy of 86.15 %, the second lowest communication traffic of 170,261,072 bytes, and the training time is only 3112 s. It not only ensures security but also reduces the training time and communication traffic while improving accuracy.