Multi-Threat Detection in Optical Networks Using Deep Neural Networks

This paper investigates deep-learning-based multi- threat detection in optical fiber networks using state-of- polarization (SOP) monitoring. Mechanical disturbances such as shaking, tapping, and fiber manipulation introduce charac- teristic variations in the stokes parameters, which were captured through a high-speed polarimeter in a controlled robotic- arm testbed. Eight polarization fingerprints, representing both normal and critical events, were collected to emulate realistic physical-layer disturbances. Three neural network models - CNN, LSTM, and GRU - were trained to classify these events, and their performance was evaluated using test accuracy, train- ing time, inference time, memory usage, and CPU consumption. A computational load index (CLI) was defined to quantify the overall computational cost based on normalized training time, inference latency, CPU load, and memory usage. The results show that GRU provides the highest accuracy, LSTM offers the best balance between accuracy and complexity, and CNN yields the lowest computational load for lightweight deployment. The combined accuracy–complexity assessment demonstrates that SOP-based sensing integrated with deep-learning models can support fast and reliable detection of physical-layer threats in optical transport networks, enabling proactive protection and improved system resilience.