Failure Localization in Disaggregated Optical Networks: Application of Vertical Federated Learning on Heterogeneous Data

We investigate the problem of failure localization in disaggregated optical networks and multi-operator environments, where end-to-end management requires coordination and data sharing between operators. We leverage Vertical Federated Learning (VFL) to enable collaborative failure localization while minimizing exposure of risk-sensitive data. We propose two VFL solutions, a neural network-based solution (SplitNN) and a gradient-boosted tree-based solution (GBDT-VFL), addressing collaborative and privacy-preserving failure localization using homogeneous and heterogeneous types of data in multi-operator and disaggregated multi-vendor scenarios, respectively, and test them on two independent optical network datasets. Our experimental results show that the proposed approaches provide comparable Accuracy in failure localization (differences are within 3%) compared to baseline scenarios of centralized failure management in which a global model is used to perform failure localization.