Abstarct: Chest images play a crucial role in diagnosing various diseases, including COVID-19. Given that the number of patients in each hospital is limited, resulting in an insufficient number of labeled images for training a network, and the fact that sharing images between hospitals is not feasible due to privacy concerns, training an independent network for each hospital is not viable. Federated learning has been employed as an innovative and effective solution to address these challenges. Federated learning is a machine learning method that creates a global model, allowing all hospitals to benefit from it without the need to share images. Each device trains its local model on its own data and then only returns the model parameters (not the raw data) to the central server. This method not only preserves data security and privacy but also uses less bandwidth. The method proposed in this thesis has the capability of allowing each client in the suggested network to have a different model. Instead of sending large and sensitive images to central servers, only the raw output of the neural networks is transmitted as the data transfer metric. Furthermore, considering that some clients in the network might contain malicious data that could negatively impact the network's performance, a technique for detecting malicious clients has been used to ensure that the network's performance remains stable and that the training process continues for healthy clients. This method not only addresses the aforementioned challenges but also shows at least a 9% increase in accuracy compared to previous methods, indicating significant progress in this area.