Federated Learning with Integrated Over-the-Air Computation and Sensing in IRS-assisted Networks

Abstract:
Network intelligence is a key feature to be implemented in future generation networks. Distributed learning schemes such as federated learning (FL) are considered as promising solutions to it while preserving users’ privacy and security and saving communication burden. However, in the case of traditional cross-device FL where massive participating devices are considered, only few clients can participate in each communication round of FL training for efficiency. Over-the-air computation (OTA) is well-suited for alleviating this issue by allowing simultaneous transmission to directly obtain the computation results by the superposition property of the wireless channels. The performance of OTA which is significantly impacted by weak channel links could be saved by adapting the wireless environment by means of an intelligent reflecting surface (IRS). Network intelligence systematically requires being aware of the environment. The sensing functionality needs therefore to be included efficiently. Integrated communication and sensing (ISAC), which consists of using the communication signal for achieving sensing at the same time, could introduce additional interference to an FL system. Therefore, an FL system design with integrated OTA and sensing in IRS-assisted networks is investigated. A maximization problem of the number of participating clients has been proposed to optimize the learning performance. A solution based on alternative optimization and difference-of-convex (DC) programming is given. Simulation results confirmed the impact of including the sensing functionality and the role of IRS to the system design.