A Bayesian-Based Aggregation Approach to Radio Outdoor Heatmap Construction Using Federated Gaussian Process

To enhance the prediction accuracy and efficiency of the wireless outdoor heatmap, we propose a novel federated Gaussian Process (GP) approach combined with Bayesian Model Averaging (BMA). Traditional centralized GP models need extensive communication between distributed sensors and a central server, which leads to inefficiencies, increased computational costs, and potential privacy risks. Additionally, the GP model's log-likelihood function is optimized using the entire dataset, which makes it incompatible with standard federated aggregation techniques such as Federated Averaging (FedAvg). To overcome these challenges, our approach enables each sensor to process its data locally with GP algorithms by mapping Received Signal Strength (RSS) to corresponding locations. At the central server, BMA predicts pseudo-labels from limited global data to create a pseudo-labeled set for knowledge distillation. This allows the central server to train a global student GP model that updates parameters rather than averaging local models directly as in FedAvg. The global model is then sent back to sensors for further iterations. We evaluate our approach using real-world RSS data from the National Science Foundation (NSF) funded Platform for Open Wireless Data-driven Experimental Research (POWDER) at the University of Utah. The experiment results demonstrate that the proposed federated GP model significantly outperforms existing methods, including the federated GP schemes using FedAvg and classical Alternating Direction of Multipliers Method (cADMM) and federated Neural Network (NN)-based scheme.