Analysis of centralized to federated learning-based anomaly detection in networks with explainable AI (XAI)

Abstract

Future networks are expected to be fully autonomous using highly advance machine learning (ML) and artificial intelligence (AI). These techniques are deployed in different context in the network such as various network optimizations, security management and decision making. Various stakeholders are focused on using interpretable ML models instead of existing black-box models. To enable edge computing capabilities and with decentralization of network architecture Federated Learning (FL) has shown promising results. Without the human expert in the loop, anomaly detection system should perform to provide required guaranteed service performance. There are growing number of concerns over the impact on FL decisions due to data scarcity and accuracy issues. It is vital to integrate explainable AI (XAI) for FL to future development in wireless networks. Therefore, we provided a comprehensive analysis on using XAI in centralized and federated learning-based anomaly detection in networks. For this analysis deep neural network was considered as the black-box model. Experiments were done with two datasets UNSW-NB15 and NSL-KDD using SHapley Additive exPlanations (SHAP) as XAI method. Specifically, we showed that FL explanation changes with FL client anomaly percentage from centralized learning explanation.