Reliable and communication-efficient federated learning for future intelligent edge networks

The thesis focuses on advancing efficient and robust federated learning (FL) to enable embedded intelligence in 6th generation (6G) networks and innovative edge computing. A major challenge stems from statistical heterogeneity arising from divergent user data distributions, rendering conventional one-size-fits-all models ineffective. We propose user-centric policies yielding specialized models tailored to each user's objectives. Users are clustered by similarity to facilitate collaborative training of shared models, mitigating communication overhead. Another focus is integrating remote Internet of Things (IoT) devices into intelligent edges using unmanned aerial vehicles (UAVs) as FL orchestrators. We jointly optimize UAV trajectories and device scheduling based on a proposed proxy metric of training performance. Finally, we investigate the communication overhead of FL wireless networks. To address this, we leverage sparse random networks to approximate target networks with over-parameterized random networks through pruning rather than direct optimization, substantially improving communication efficiency.