Detection of Optical Network Breaches through ML-Based State of Polarization Analysis

Doctoral thesis, 2026

Optical fiber networks form the backbone of modern communications, yet they are vulnerable to physical-layer disturbances ranging from benign environmental vibrations to malicious threats like fiber tapping. This dissertation addresses the urgent need for advanced monitoring solutions of environmental disturbances by leveraging the State of Polarization (SOP) of light as a sensitive, non-intrusive indicator of fiber events. We develop a Machine Learning (ML)–based framework that continuously analyzes SOP variations to detect and classify physical-layer anomalies.  Our approach encompasses Supervised Learning (SL)  for classification of known events, including Deep Learning (DL) architectures that automatically extract complex polarization features in challenging real-world conditions, as well as Semi-supervised Learning (SSL) and Unsupervised Learning (USL) techniques for detection of novel anomalies without reliance on fully labeled data.

In controlled laboratory settings, the proposed methods distinguished mechanical vibrations, eavesdropping-induced fiber bends, and other perturbations with high accuracy (exceeding 97% in multi-class classification). Field trials on live, metro-scale fibers further demonstrated robust performance, detecting intrusion attempts and accidental disturbances with minimal degradation in performance, despite real-world noise. Notably, this work provides the first validation that polarization-based sensing remains effective in standard coherent communication systems: ML models accurately detected disturbances on Dual-Polarization 16-Quadrature Amplitude Modulation ( DP-16QAM) data-carrying channels with accuracy comparable to that obtained on unmodulated Continuous Wave (CW) probes.

Results confirm that ML-driven SOP-based analysis can rapidly flag fiber taps and other physical intrusions, and distinguish harmful events from harmless fluctuations with high confidence. By validating the proposed intelligent SOP-based monitoring framework over diverse real-world conditions, including different fiber types, network configurations, and signal modalities, this work demonstrates that polarization-based fiber monitoring is practically viable for deployment in real-world operational optical networks. The findings establish SOP analytics as a powerful and non-intrusive tool for enhancing the security and resilience of modern optical communication infrastructure without disrupting normal traffic.