Vertical Federated Learning for Failure Localization in Partially Disaggregated Optical Networks

Machine Learning (ML) for failure management in optical networks has recently gained noteworthy attention. Even though real field-collected data is crucial for ML-based failure management, it is challenging to access data in emerging disaggregated optical networks, where multi-vendor equipment co-exist, and the end-to-end network management requires coordination between operators that manage different network segments. Due to data confidentiality issues, network operators tend not to share business-critical data, which sets a barrier to utilizing ML-based approaches. To overcome this issue, we propose a Vertical Federated Learning (VFL) approach based on Split-Neural-Network (SplitNN) for failure localization. We consider different deployment scenarios for ML-based solutions in a collaborative and privacy-preserving manner. Our experiments show that, depending on the VFL client and server model architectures, the proposed approaches provide very similar accuracy compared to a baseline scenario of a single operator managing the whole network (differences are mostly within $1 \%$ of accuracy), while minimizing the exposure of risk-sensitive data.