Deep Learning for Detection of Harmful Events in Real-World, Noisy Optical Fiber Deployments

Optical network infrastructure underpins global communication networks. It is exposed to various physical layer breaches, such as fiber cuts or eavesdropping via fiber bending, that may violate privacy or disrupt services. Analyses of State of Polarization (SOP) variations induced by external events, combined with Machine Learning (ML) techniques, can contribute to early identification of events and categorization of potential threats. However, real-world deployment of automated threat detection and mitigation faces many challenges, including the inconsistencies between controlled laboratory settings, often used for dataset collection for ML training, and real-world, noisy environments. In this paper, we study the detection of external disturbances in real-world fiber installations by analyzing the induced changes in the SOP of optical signals. We develop a suite of Deep Learning (DL) models, including One-Dimension (1D) Convolutional Neural Network (CNN) and fully-connected dense layers, for the detection of harmful events in noisy environments comprising a shorter (300 m) and a longer (21 km) fiber link installation, corresponding to Fiber to the Home (FTTH) and metro-scale optical paths, respectively. The proposed approach employs an optical analyzer to capture SOP changes resulting from mechanical or acoustic vibrations, as well as eavesdropping attempts. Upon careful tuning of the DL models' hyper-parameters, 98.57% accuracy is obtained for the shorter, and 92.26% for the longer link installation.