This application provides a federated learning method, apparatus, and system, so that a server retrains a received model in a federated learning process to implement depersonalization processing to some extent, to obtain a model with higher output precision. The method includes: First, a first server receives information about at least one first model sent by at least one downstream device, where the at least one downstream device may include another server or a client connected to the first server; the first server trains the at least one first model to obtain at least one trained first model; and then the first server aggregates the at least one trained first model, and updates a locally stored second model by using an aggregation result, to obtain an updated second model.