Minimizing Energy Consumption for Decentralized Federated Learning Using D2D Communications

Federated learning (FL) is a promising distributed machine learning technique for building inference models over wireless networks due to its ability to maintain user privacy and reduce communication overhead. In this paper, we consider minimizing the energy consumption of a device-to-device (D2D) network while maintaining the convergence rate of FL subject to its time constraint. In the considered D2D network, each device has limited transmission range and is connected partially to other devices in the network. A group of devices can form a cluster and one of these devices is judiciously selected as a local aggregator (LA) to aggregate the local models of other devices in the cluster. Leveraging the nature of D2D communications, we exploit the devices that are located at the conflict zones of LAs. As such, the LAs can disseminate their local aggregated models among them. Towards this goal, a joint optimization problem, considering scheduling the devices to the LAs and computation frequency allocation of the devices, is presented. In order to solve this NP-hard problem, an iterative solution is devised. Particularly, we decompose it into two sub-problems, namely, LAs selection and device scheduling sub-problem and computation frequency allocation sub-problem. By solving theses sub-problems iteratively, a FedD2D (federated learning with D2D communications) scheme is proposed. MATLAB simulations are conducted to verify the effectiveness of the proposed FedD2D scheme over FL conventional schemes.