Collaborative Research: CISE MSI: RDP: CNS: Towards the Resilient NextG Network Design for Federated Learning over Mobile Devices

Next generation (NextG) network systems are envisioned to be complex, ubiquitous, and smart, which are likely to consist of millions of heterogeneous mobile devices to connect everything digital, enable machine-to-machine communications, and support a variety of critical machine learning (ML) paradigms, including the most popular federated learning (FL) over mobile devices. However, stakeholders in many intelligent mobile applications/services are resource constrained in terms of spectrum, energy, computing, etc., which poses many challenges to FL inspired applications/services. This project targets to develop a novel NextG network with high degrees of resiliency to address those challenges, in particular, when there may be massive bursty workloads, insufficient spectrum availability, limited computational and storage capability on edge, and privacy concerns of the training data on mobile devices. The anticipated project outcomes will enrich the knowledge of wireless systems and machine learning technologies and provide multidisciplinary training especially for underrepresented students. Additionally, the findings and innovations will be shared across the 23-campus California State University (CSU) system, where 90% of campuses are minority-serving institutions. Outreach activities including high school internships and summer undergraduate training programs can provide early exposure to research in science and engineering, fostering interest and encouraging more female and minority students to pursue careers in these fields. This project aims to address the resilient issues of FL over mobile devices via a novel holistic NextG network design across network architecture, local mobile devices, and accessing networks. (1) From the networking system's perspective, to support FL over large-scale heterogeneous mobile devices, serverless computing is exploited at the edge to resiliently and efficiently provide ML computing as a service. (2) From the local mobile devices' perspective, to resiliently protect local training data privacy against inference attacks in FL, an energy-efficient piggyback differential privacy (DP) design is proposed by jointly considering DP amplification from gradient quantization and sparsification, and free Gaussian noises from wireless channels. (3) From the accessing networks' perspective, to improve the spectrum accessing resiliency, network scalability, and spectrum efficiency, a multi-bit over-the-air computation (M-AirComp) based spectrum accessing design is proposed, which can enable efficient transmission of FL model updates even with limited spectrum availability, reducing the total energy consumption for mobile devices. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.