On the trustworthiness of federated learning models for 5G network intrusion detection under heterogeneous data

The rapid advancement of 5G networks is reshaping wireless communications through ultra-fast speeds, low latency, and seamless connectivity. This shift is accompanied by emerging technologies such as edge computing, network traffic management, resource allocation, and network slicing, which distribute data processing across the network. These trends amplify the need for privacy-preserving and decentralized learning methods, particularly in security-critical applications where transmitting raw data to a central server may be infeasible or undesirable. Federated Learning has emerged as a promising paradigm to meet these demands by enabling model training across distributed data sources. In this study, we explore the trustworthiness of Federated Learning models compared to a centralized counterpart, under various heterogeneous client data distributions. The distributions are generated via label-based symmetric Latent Dirichlet Allocations, where the Dirichlet concentration parameter α controls the degree of class imbalance across clients. We use traffic flow data from a 5G Network Intrusion Detection task to design centralized and federated Artificial Neural Network architectures and extract feature importance scores using the Integrated Gradients algorithm. Our findings, based on the top-10 features, show that federated models trained on slightly-skewed ( α = 10 ) and mildly-skewed ( α = 2 ) data achieve trust scores closely aligned with the central model. The α = 10 model achieves an average importance score of 2.2 (7.2% lower than the central model at 2.37), with 91.3% feature overlap. The α = 2 model scores 2.44 (3% higher), with identical overlap. In contrast, highly-skewed models ( α = 0 . 1 , 0 . 5 ) show diminished trustworthiness, scoring 1.62 and 1.74 (31.6% and 26.6% lower), with overlaps of 80% and 82.5%, respectively. These results highlight the impact of client data heterogeneity on model trustworthiness and underscore the sensitivity of federated models to high levels of data heterogeneity.