FedRAN: Federated Mobile Edge Computing with Differential Privacy

In this paper, we propose FedRAN, a mobile edge, federated learning system that incorporates differential privacy to improve the privacy integrity of sensitive edge information, preventing adversarial entities from exploiting the network interactions within a federated ecosystem to access private edge data, while tapping into the vast amounts of data generated from distributed endpoints. We deploy and evaluate FedRAN in a real controlled radio-frequency LTE environment, as opposed to a simulated one. We show that FedRAN's distributed model outperforms locally-constrained models on the MNIST handwritten digits dataset. Additionally, we explore a variety of differential privacy settings, in an effort, to enable a privacy preserving, large scale mobile edge computing ecosystem. To our knowledge, our work is the first evaluation of a federated learning system within a controlled radio-frequency LTE environment.