A federated learning method, apparatus, and system are disclosed. A first node obtains data distribution information of a plurality of second nodes based on a target data feature required by a training task; the first node selects at least two target second nodes from the plurality of second nodes based on a target data class required by the training task and the data distribution information of the plurality of second nodes; and the first node indicates the at least two target second nodes to perform federated learning, to obtain a federated learning model that is in the training task and that corresponds to the target data class. In this way, when participants have a plurality of data distributions, a trained model is prevented, as much as possible, from being affected by data poisoning.