Wireless federated learning (FL) is envisioned as a promising paradigm of distributed learning in wireless networks without disclosing users' data privacy. However, radio channel leads to a potential risk of eavesdropping attack when sending the trained model data in wireless networks. To address this eavesdropping attack, in this work, we propose an energy-efficient wireless FL by using artificial jamming. Specifically, the group of wireless devices (WDs) adopt the time division multiple access (TDMA) approach to send their locally trained models to the FL server for model aggregation subject to the eavesdropping attack of a malicious node. When one of the WDs sends its local model, all the other WDs send the artificial jamming signals to interfere with the eavesdropper, which helps increase the secrecy throughput of the targeted WD for uploading its local model. Different from many existing studies using stochastic gradient descent (SGD), we adopt the stochastic average gradient (SAG) method in the local training to improve the convergence of FL and derive the corresponding lower bound of the FL convergence rate via SAG. Furthermore, we formulate an optimization problem that aims at minimizing the energy consumption of the WDs by jointly optimizing different WDs' local training time, their uploading transmission time in TDMA and the transmit-powers for providing artificial jamming, as well as the FL configurations of the local/global iterations. We also propose an efficient algorithm for solving this non-convex optimization problem. Numerical results are illustrated to validate the advantages of our design of wireless FL and the corresponding algorithms. In particular, the results demonstrate that our secrecy oriented energy-efficient FL can significantly outperform the other heuristic FL schemes.