Talk Title: Federated Learning (FL) for Next-Generation Edge Intelligence: Over-the-Air  Federated Learning (AirFL) beyond Spectral Efficiency

With emerging edge intelligence applications deployed over geo-distributed IoT terminal devices, an upsurge of interest has been gained for FL, where multiple devices collaboratively train or infer AI models w/o explicit exchange of local data. To improve spectral efficiency, over-the-Air computing (AirComp) is well known to enable multiple devices’ simultaneous access to uplink channels for analog model aggregation. AirComp based FL (referred to as “AirFL”), however, encounters many challenges in practice, for instance, deteriorated signal alignment due to unaffordable channel estimation, poor generalization caused by data heterogeneity, and vulnerability to privacy attacks. In this talk, first, we advocate non-coherent AirFL (NCAirFL) that proves to achieve the same convergence rate as the ideal baseline (FedAvg), leveraging binary dithering codes and non-coherent detection. Next, by revisiting the unique role played by channel noise in AirFL, we introduce meta-learning based personalized FL (meta-pFL), which pretrains a shared hyperparameter that can be efficiently fine-tuned for new tasks and/or devices, and identify fundamental trade-offs between convergence and generalization. Furthermore, to harness the inherent channel noise for preserving differential privacy (DP), in an AirFL system equipped with a multi-antenna access point (AP), we jointly optimize receive beamforming and power allocations to characterize the convergence-privacy trade-offs, revealing insightful conditions, in which user-level DP is achievable as a perk without compromise to training or reliance on any artificial noise (AN).