Federated learning (FL) has emerged as a key technology for enabling next-generation, privacy-preserving AI at scale. It allows a large number of edge devices, such as mobile phones, to collaboratively train a shared global model while keeping their data local, thus preventing privacy breaches. Deploying FL over wireless multi-hop networks, including wireless community mesh networks and satellite-based wireless internet, can not only enhance AI experiences for urban mobile users but also democratize AI, making it accessible affordably to everyone, including those in low-income communities, rural areas, underdeveloped regions, and disaster zones. This project aims to democratize AI by developing a novel wireless multi-hop FL system that guarantees stability, high accuracy, and fast convergence by systematically addressing communication latency, system heterogeneity, and statistical heterogeneity.

The project's intellectual merit includes:

(1) Novel FL system architectures. We developed innovative FL architecture, e.g., Sync-DFL, Async-DFL, and DHA-FL, which leverage layered federated computation, semi-asynchronous model aggregation, and regularized objective functions. These designs significantly improve system scalability, communication efficiency, and stability, enabling reliable FL operation in challenging environments. We also employed multi-agent reinforcement learning to optimize FL system performance by maximizing accuracy while minimizing convergence time under the computing constraints of edge devices.

(2) Adaptive deep learning strategies on edge devices. We devised high-performance, computationally efficient, and robust deep learning and federated training strategies for resource-constrained edge devices. These include a novel model regularization method (GradAug), a resource-aware model adaptation framework (MutualNet), and communication-efficient and robust federated learning algorithms that systematically exploit local learning generality (FedAlign), parameter-efficient fine-tuning (FedPEFT), diffusion models (FedDiff), and prefix-tuning with parallel attention (FedPEFT).

(3) Wireless multi-hop FL experimental framework. We designed and prototyped the EdgeML testbed and simulator, the first experimental framework for FL over multi-hop wireless edge computing networks. This platform facilitates rapid prototyping, deployment, and evaluation of novel FL algorithms in both real-world and simulated wireless environments, along with RL-based system optimization methods.

The project's Broader Impacts

This project advances the understanding of the strong synergy between distributed edge computing and wireless networking, bridging the gap between the theoretical foundations of distributed deep learning and its practical applications. The proposed FL frameworks and testbeds are the first to demonstrate the effectiveness of FL within low-cost multi-hop wireless edge computing networks, thereby making AI more affordable and accessible to underserved populations in low-income communities, underdeveloped regions, and disaster-affected areas. The privacy-preserving nature of the developed FL systems makes them suitable for diverse applications, including smart healthcare, smart cities, intelligent power systems, and AI-driven finance. Furthermore, this project provides valuable interdisciplinary training opportunities for graduate and undergraduate students through research activities and relevant courses developed and offered by the PIs.

Last Modified: 12/27/2024
Modified by: Pu Wang