Balancing Privacy and Explainability in Federated Learning

Abstract

As we advance towards future generations of communication networks, such as 6G, Artificial Intelligence (AI) and Machine Learning (ML) will assume an increasingly pivotal role in network optimization, management, and operation. In this context, the pursuit of reliability in ML models has emerged as a highly active area of research. Explainable AI (XAI) is an essential tool to unravel the underlying mechanisms of network behaviour, enabling a deeper understanding of the decisions made by black-box models in future-generation networks. This paper investigates the impact of Federated Learning (FL) on explainability by evaluating the correlation between the feature importance derived from models trained using four different FL approaches and the model trained using a conventional single-host approach. The analysis utilizes two publicly available datasets, two XAI metrics, and two correlation metrics. The findings reveal a high correlation between FL and single-host trained models in three of the four FL approaches, suggesting that some FL approaches failed to capture similar patterns.