Towards ai-empowered wireless networks: from edge to core
Wireless multi-hop networks have been widely exploited for deploying cost-efficient network backbones including wireless community mesh networks, high-speed urban networks, global wireless internet infrastructure, battlefield networks, and public safety/disaster rescue networks. Federated learning (FL) is a distributed machine learning technology for next-generation AI systems that can collaboratively improve a shared global model while preserving user privacy and encompassing users at a larger-scale. FL systems are designed to be used on Single-hop wireless networks which consists of edge servers that are connected to the high-speed internet core. Enabling FL over wireless multi-hop networks can democratize AI and make it accessible in a cost-effective manner. However, our preliminary study found that FL over wireless multi-hop networks possess significant challenges due to the presence of multiple noisy interference-rich wireless links which not only slows the learning process due to the underlying communication delay but also leads to nomadic model updates. The inherent bottleneck for FL over wireless multi-hop networks are (1) One-size-for-all model deployment and training, where each edge device process the same number of local iterations for model updates (2) Model-based optimization is not feasible for multi-hop FL, since FL performance metrics cannot be formulated as a closed-form function for network control parameters such as packet forwarding decision and transmission power for each router. In this thesis, we proposed a novel Artificial Intelligence (AI) empowered wireless network systems that can guarantee stability, high accuracy, and faster convergence speed by taming communication latency, system heterogeneity, and statistical heterogeneity. Towards this goal (1) We have developed a Programmable Wireless Network Operating System (WINOS), which allows the user to implement AI routing solutions (2) We have developed a novel Hierarchical Synchronous FL system architecture that can maximize communication efficiency in addition to tolerating stragglers for non-blocking executions (3) We have implemented a naturally federated application, Gait based user authentication and recognition mechanism using millimeter wave, which is one of the privacy-preserving biometric authentication mechanism (4) We have developed a novel domain adaptation solutions that helps in applying FL system trained in single domain to different domains that possess spatial variations. Finally, our experimentation result shows that AI enabled wireless networked systems are extremely efficient in handling heterogeneity and communication latency, surpassing the traditional network systems performance.