Utility Maximization of Multi-Federated Learning in Edge Computing with Personalized Privacy Preservation

Edge intelligence enables mobile users to benefit from real-time inference services based on deep neural networks (DNN). Federated learning (FL) provides a solution for using DNN training while protecting privacy. FL over mobile edge computing (MEC) can aggregate models at the edge and process them in parallel, providing far more real-time results for realworld applications. However, edge nodes have limited computing capacities and bandwidth, and not all user equipments (UEs) can be selected to upload their trained local models. In addition, private information of users can still be leaked while attackers analyze the uploaded model parameters, and users' privacy requirements vary. Thus, we proposed a novel optimization framework - Federated learning with personalized differential privacy over MEC based on deep reinforcement learning. We use deep reinforcement learning (DRL) to maximize the total utility, i.e., the overall accuracy of all global FL models, by choosing UEs for uploading their updated local models due to limited bandwidth on access points (APs) and computing resource capacities on cloudlets (edge servers). Then, we inject differential private noise into local models to enhance privacy and satisfy users' personalized privacy requirements while guaranteeing model accuracy. We finally evaluate the performance of the proposed approach through experiments. Experimental results show that the proposed approach outperforms the comparison counterparts significantly, using public accessible datasets.