The artificial intelligence training of the terminal clients is mostly limited by the insufficient computing power and energy shortage of the client itself, as well as the client’s offline and malfunction, which lead to the terminal intelligent training consuming a large amount of system time. A trusted server is introduced in this scenario, which allows the training tasks of terminal clients to be offloaded to the edge server, to solve this problem. Firstly, a joint optimization model of computation offloading and resource allocation for federated training in mobile edge networks is established. Then the optimization problem is transformed into a Markov process, and the DQN algorithm is applied to obtain the optimal decision. Large amounts of simulation results show that the proposed algorithm reduces the training delay of terminal clients effectively.