On the Impact of Model Compression for Bayesian Federated Learning: An Analysis on Healthcare Data

Bayesian Federated Learning (FL) policies enable multiple nodes to collaboratively train a shared Machine Learning (ML) model while accounting for the uncertainty of its predictions. This is accomplished by estimating the global posterior distribution in the model parameter space. Currently, Bayesian FL strategies are impaired by large communication costs that need to be reduced to provide more sustainable training platforms. This letter investigates the impact of compression strategies in centralized Bayesian FL setups, where a Parameter Sever (PS) is tasked to supervise the learning process. The goal is to study how compression affects the ability of Bayesian FL systems to provide high-quality, yet well-calibrated ML models. The analysis is carried out in the healthcare domain, where the prediction reliability is particularly critical, focusing on a medical imaging task. Numerical results show that applying aggressive compression policies highly reduces the ability of Bayesian FL systems to provide accurate and reliable ML models. On the contrary, light compression stages maximize accuracy and calibration at the cost of larger communication overheads.