<p>Efficient communication is very important for federated learning, but in the actual training process, multiple rounds of communication are required between the server and the clients to train the global model, which leads to a large amount of communication costs. Knowledge distillation is widely used in deep learning to reduce model size for accelerating the training progress. However, most of the existing knowledge distillation algorithms applied in federated learning require a lot of calculation locally or customize the personalized model for the participating clients, which may add some unnecessary costs. To address this challenge, we propose a new federated distillation algorithm, FedMKD, which introduces multi-teacher knowledge distillation based on confidence perception in traditional federated learning. Each client participating in the training is used as the teacher side, and the server side is used as the student side. The different weights of the teachers are calculated according to the confidence algorithm, and the model knowledge contained by each teacher is directly distilled and taught to the student model. Experiments show that the model trained by our algorithm has a certain degree of reduction in average running time, parameter number, and memory size compared with other federated distillation algorithms and also ensures a certain degree of accuracy, which prove that the FedMKD algorithm is more lightweight while ensuring considerable model accuracy.</p>