An Experimental Framework for Studying Non-IID Data in Federated Learning for Network Telemetry

The increasing complexity of emerging 5G and 6G network environments has intensified the need for data-driven automation under heterogeneous and dynamic conditions. Federated Learning (FL) is a promising paradigm in this context. This paper presents an experimental framework to generate realistic Non-Independent and Identically Distributed (Non-IID) datasets through controlled execution of a distributed service and telemetry collection, aiming to improve the applicability of FL in network automation. Using Apache Cassandra as a representative cloud-native application, we construct datasets exhibiting temporal and structural heterogeneity. We statistically characterize these datasets and evaluate their impact on regression models and horizontal federated learning using a Wide & Deep architecture. Results show that while horizontal federation improves generalization compared to direct cross-dataset transfer, its performance degrades under pronounced structural Non-IID conditions, highlighting both its potential and limitations.