Clinical diagnosis and therapy of brain tumors are greatly aided by proper classification of the tumors. Brain tumors can be diagnosed more quickly and accurately if radiologists use deep learning to help the specialist and doctors examine the enormous volume of brain MRI Images. Large datasets are required in training process, and whole of such data must be centralized for be handled by such techniques. It is sometimes impossible to collect and distribute patient data on a centralized data server because of medical data privacy regulations. In this paper, federated learning (FL) is proposed, in which data is non-shareable because of patient privacy issues. Using the FL approach, we have proposed two methods of aggregation; first, which concerns ranking the weight percentage of each client, and Second average weights method. and to evaluate the suggested model, we have compared the performance of the ranking weights percentage method with the average weights of proposed CNN and pre-training (VGG-16) in the FL environment in addition to SVM and VGG-16 . The experiments result was applied on two datasets, it shows our model accuracy result is very effective when using the ranking weight percentage method as compared with other methods, it achieves accuracy (98%) on datasets (BT_large-1c) and achieve (97.14%) on the dataset (BT-large-2c).