Machine Learning Analysis of Polarization Signatures for Distinguishing Harmful from Non-harmful Fiber Events

Secure and reliable data transmission in optical networks is essential for supporting high-speed internet services. Optical fibers, the enabler of global connectivity for millions of users, are vulnerable to various potentially harmful events including mechanical failures, like fiber cuts, and malicious physical layer attacks, such as eavesdropping. These incidents can degrade network performance, breach privacy and integrity through unauthorized access to the transmitted data, and cause significant financial and data loss. It is, therefore, crucial to detect and classify the malicious events. Continuous monitoring of polarization state changes, combined with application of machine learning algorithms, enables detection of deviations in the polarization patterns caused by the harmful events. In this study, we introduce a method that detects and identifies potential harmful events in optical networks. By using a Histogram Gradient Boosting classifier within our machine learning framework, we achieve $97.94 \%$ detection accuracy of the harmful and non-harmful events.