Over The Air Federated Learning in the Presence of Impulsive Noise

Data-driven applications in the next-generation wireless communication networks lead to data privacy and security constraints. Privacy-preserving federated learning (FL) in wireless systems has emerged as a solution for this issue, where machine learning (ML) models are trained collaboratively by only sharing ML model updates over the wireless medium. While addressing bandwidth constraints in wireless communications, over-the-air FL (OTA-FL) is preferred since the edge users can simultaneously transmit model updates to a central entity over power-controlled wireless uplink channels. We show that the OTA-FL framework is highly susceptible to impulsive noise, which is prevalent in several wireless systems. We study the convergence of OTA-FL techniques in the presence of impulsive noise and propose Sk-OTAFL, where we omit the impulsive noise-contaminated model updates in FL. Using the MNIST and CIFAR10 datasets, we demonstrate the vulnerability of the OTA-FL systems to different strengths of impulsive noise with respect to ML model accuracy while varying the number of clients and level of data heterogeneity. We also demonstrate that the proposed Sk-OTAFL strategy effectively mitigates the impact of impulsive noise.