Towards Understanding Device Impact in Heterogeneous Federated Learning

The development of the Sixth-generation Network (6G) involves integrating Artificial Intelligence (AI) to meet performance goals and enable new applications. Federated Learning (FL) emerges as an efficient method to build native AI in 6G. However, FL faces challenges due to the heterogeneity of devices, such as variations in computational power, which can hinder efficient training and reduce overall model accuracy, especially in the expected 6G ecosystem. This work introduces a novel method leveraging Explainable Artificial Intelligence (XAI) and correlation metrics to quantify device contributions in FL without compromising performance. Two FL approaches, FedAvg and HeteroFL, were analyzed regarding the impact of the devices in the aggregated model when using them. Results demonstrate that while FedAvg achieves uniform contributions across devices (Pearson Correlation Coefficient (PCC) of 0.7 for every device), HeteroFL shows a decline in contributions with increased network reduction, achieving 0.45 PCC for devices training 66% of the network and 0.25 PCC for those training 33%. These findings provide critical insights into optimizing FL in resource-constrained environments.