Vertical Federated Learning for Failure-Cause Identification in Disaggregated Microwave Networks
This addresses fault management for telecom operators and vendors in collaborative network architectures, offering a privacy-preserving solution, though it is incremental as it adapts existing VFL methods to a new domain.
The study tackled failure-cause identification in disaggregated microwave networks by applying Vertical Federated Learning (VFL) methods like SplitNNs and FedTree, achieving F1-Scores within at most a 1% gap compared to centralized scenarios while ensuring minimal sensitive-data leakage.
Machine Learning (ML) has proven to be a promising solution to provide novel scalable and efficient fault management solutions in modern 5G-and-beyond communication networks. In the context of microwave networks, ML-based solutions have received significant attention. However, current solutions can only be applied to monolithic scenarios in which a single entity (e.g., an operator) manages the entire network. As current network architectures move towards disaggregated communication platforms in which multiple operators and vendors collaborate to achieve cost-efficient and reliable network management, new ML-based approaches for fault management must tackle the challenges of sharing business-critical information due to potential conflicts of interest. In this study, we explore the application of Federated Learning in disaggregated microwave networks for failure-cause identification using a real microwave hardware failure dataset. In particular, we investigate the application of two Vertical Federated Learning (VFL), namely using Split Neural Networks (SplitNNs) and Federated Learning based on Gradient Boosting Decision Trees (FedTree), on different multi-vendor deployment scenarios, and we compare them to a centralized scenario where data is managed by a single entity. Our experimental results show that VFL-based scenarios can achieve F1-Scores consistently within at most a 1% gap with respect to a centralized scenario, regardless of the deployment strategies or model types, while also ensuring minimal leakage of sensitive-data.