Digital Twin-enabled Federated Learning in Mobile Networks: From the Perspective of Communication-assisted Sensing

Abstract

With the continuous evolution of emerging technologies such as mobile network, machine learning (ML), 5G, etc., digital twins (DT) bursts out great potential by its capacity of data analysis, data tracking, data prediction, etc, building a bridge between the physical and information world. Meanwhile, mobile network is moving towards data-driven paradigm, the issue of data privacy and data security seem to be a bottleneck. As a result, federated learning (FL) and mobile network are deeply converging. However, the mobile network is time-varying and the parameters of FL-empowered mobile network is huge and continue to increase with exponential growth of wireless terminals, result in the failure of traditional modeling. In the mobile networks, DT is conducive to prototyping, testing, and optimization, enabling mobile networks to be modelled more efficiently in a virtual environment and thus providing guidance for practical application. To this end, a communication-assisted sensing scenario is considered in this paper with FL in DT-empowered mobile networks. More specifically, two communication-assisted sensing architectures are proposed to improve communication efficiency of mobile network, namely, centralized architecture of federated transfer learning (FTL) and decentralized architecture of FTL. For centralized architecture of FTL, feature extraction of sensing information is conducted by FL between partial nodes and central server while the remaining nodes are used to train the fully connected layers at the central server. Considering data safety during the communication between sensing nodes, a decentralized architecture is designed based on FTL and Blockchain, where the feature extraction module is obtained by the fusion of sharing model (by Blockchain) and local model. The performance of proposed schemes is evaluated and demonstrated by the simulations.