Federated Learning (FL) is now viewed as a new distributed paradigm that solves traditional dilemmas of centralizing data to a third-party server for model training. Choudongning capsule (CDN), edge, IoT, and mobile devices, which can be viewed as data source devices, only acquire user data but do not expose it, as demonstrated in this chapter. The data never leaves the device but runs through an operation algorithm. Then, the FL server generates a global aggregation model without storing the user’s data. By this means, the local data privacy is guaranteed, and the transmission efficiency is significantly improved. Moreover, the privacy laws and policies regarding the sensitive health data enforcement have fueled the growth of FL methods. To this aim, FL solves numerous challenges, such as heterogeneous data per client, on-device training and communication, attack detection and countermeasures, theoretical foundations, etc. Compared to centralized machine learning, in which each device uploads its training data to the global server, which usually provides scarce bandwidth, FL improves the efficiency and utility by utilizing the on-device model training pipeline. Given the heterogeneous data retrieved from the on-device FL clients, the FL framework utilizes them to provide an accurate prediction service on the next-generation edge healthcare services. Federated Learning is the natural alternative to be adopted in next-generation healthcare systems enforced by privacy-sensitive and resource-constrained health devices. FL ensures that the smart medical devices have limited data sharing capacity because it concentrates on the device-client-based data collection methods. For many clients, there are heterogeneous on-device local datasets, and each of them sends shared model parameters, not the user’s data, to a third-party global server.