Analog Over-the-Air Federated Learning With Real-World Data
Zihan Chen, Zeshen Li, Jingyi Xu · 2022
Federated edge learning enables intelligence services to be deployed at the edge of future wireless network. To address the limited spectral resource and constrained scalability, analog over-the-air federated learning is recently proposed to achieve fast aggregations with enhanced spectral efficiency, privacy and concurrent reduced access latency, via exploiting the superposition property of wireless waveforms. From a practical aspect of real-world data, noisy label quality commonly exists in local data of diverse clients in a heterogeneous wireless network, which would possibly result in performance degradation. In this work, we are the first to explore the overall effects of the interference in the random channel and the noisy labels in local datasets. A comprehensive framework to generate distributed data with discrepancies in data heterogeneity and quality is introduced. We then illustrate the overall effects of both noisy data and channel interference. The numerical results indicate that the two-sided effect also exists in this real-world system. More importantly, the potential positive effect brought by the uncertainty and randomness, could be further investigated for better utilization of the channel interference to improve the robustness of analog over-the-air federated edge learning with real-world data.