Theoretical Analysis and Performance Evaluation for Federated Edge Learning with Integrated Sensing, Communication and Computation

Edge artificial intelligence (AI) provides potential solutions for ubiquitous intelligent access and services in future 6G networks, by integrating sensing, communication and computation (SCC) at network edges. On the other hand, federated edge learning (FEEL) is identified as one of the key candidates for edge AI due to its advantages on distributed learning and privacy protection. Nevertheless, SCC-assisted FEEL (SCC-FEEL) is still an open issue to meet the development of 6G networks, which is thus considered in this work. Particularly, each device derives training data samples by sensing the surrounding environment, and aggregates local training models to an edge server through an efficient over-the-air computation (AirComp) technique. Moreover, it is challenge currently to evaluate the SCC-FEEL performance theoretically and quantitatively, which is essential for efficient resource management. To address these issues, convergence analysis is theoretically conducted, which reveals that either the sensing data size or the AirComp aided communication error would influence the SCC-FEEL model training performance. Based on the convergence analysis, we derive an average training error (ATE) metric that can evaluate the SCC-FEEL performance. Numerical results validate the effectiveness of our proposed SCC-FEEL, and verify the feasibility of ATE.

I. Introduction

With the explosive requirements of ambient intelligence, industrial internet of things (loT), and vehicle-to-everything (V2X), there is a paradigm shift for 6G networks from simple information transmission to multi-functional platform that integrated the abilities of both sensing, communication and computation [1], [2]. Meanwhile, edge artificial intelligence (AI) is also emerged as a promising technology to enable pervasive intelligence at the network edge by leveraging massive generated data under limited communication resources [3]–[5].