Robust regression as a benchmark for regression based federated learning

Abstract

This research evaluates the effectiveness of robust regression techniques, specifically Theil-Sen and Repeated Median Regression (RMR) as benchmarks for regression- based Federated Learning (FL) across a range of simulated sensor configurations. FL is a privacy-preserving paradigm that enables collaborative model training across distributed nodes without centralizing the raw data. Standard regression techniques such as linear regression, which is not robust against outliers and heterogeneity are commonly employed in real world FL environments. In contrast, RMR demonstrates the highest resilience and lowest error rates, with Theil-Sen also showing strong per- formance as a more computationally efficient alternative. RMR achieves the overall highest performance gains, with the best efficiency improvement with respect to the linear federated regression. The findings of the study support the use of robust regres- sion as a dependable alternative in distributed, real-world sensor networks.