Spectrum Breathing: A Spectrum-Efficient Method for Protecting Over-the-Air Federated Learning Against Interference

Federated Learning (FL) is a widely embraced paradigm for distilling artificial intelligence from distributed mobile data. However, the deployment of FL in mobile networks is compromised due to the exposure to interference from neighboring cells, besides a communication bottleneck caused by the uploading of high-dimensional model updates. Existing interference mitigation techniques require multi-cell cooperation or at least interference Channel State Information (CSI), which is expensive in practice. To address these challenges, we propose Spectrum Breathing, which cascades stochastic-gradient pruning and spread spectrum to suppress interference without bandwidth expansion. The cost is higher learning latency by exploiting the graceful degradation of learning speed due to pruning. We synchronize the two operations using a common parameter, Breathing Depth, and develop a martingale-based approach to convergence analysis of the Over-the-Air FL with Spectrum Breathing (AirBreathing FL). Given the receive SIR and model size, the optimization of the tradeoff between pruning and interference-induced error yields the scheme for controlling the breathing depth that can be adaptive to channels and learning process. Experiments show that AirBreathing FL with adaptive breathing depth can obtain close-to-ideal performance in scenarios where traditional over-the-air FL fails to converge in the presence of strong interference.