This Federated learning has received extensive attention in recent years due to its potential in mitigating risks of intruding privacy. However, most of existing methods don’t provide adequate consideration to the heterogeneity of client. Therefore these parameters are treated equally and aggregated with the same weight. As a result, the overall performance may be degraded because of straggling clients. In this paper, we propose an aggregation algorithm considering the accuracy of local model of heterogeneous clients. The server evaluates the accuracy of the uploaded local model with a benchmark dataset, and then updates model parameters according to the accuracy ratio. Experimental results show the model accuracy of our proposed method performs better than existing methods. At the same time, in the presence of noisy users, it can eliminate their effect by assigning lower aggregation weights to them.