Enabling Efficient Federated Learning via Ternary Quantization over Fading Channels

Federated Learning (FL) performance degrades when noise and fading affect uplink and downlink transmissions between users and the Fusion Center (FC). This research addresses a gap by investigating the impact of these channel effects on global machine learning (ML) models. We propose a method to improve communication and computation efficiency in FL using ternary quantization and a doping ratio. We simulate an additive noise channel with Rayleigh fading in both uplink and downlink transmissions. Two algorithms, Stochastic Gradient Descent with Momentum (SGDM) and Gradient Descent Backpropagation (GRDBP), are used to train local ML models. ML model parameters are then quantized with three values and power-adjusted before transmission. The doping ratio explores how combining user's local ML model with the global ML model at different ratios affects re-learning. Our simulations show that a doping ratio of 0.5 or higher significantly improves the system's resistance to additive noise and channel fading, leading to higher accuracy. Further, the combination of GRDBP, ternary quantization, and a high doping ratio is shown to be particularly effective under these challenging channel conditions. This technique enables FL systems to achieve faster, more accurate performance while maintaining user data privacy.