Deep Learning-Based Optical Spectrum Analysis for Lightpath Distance Evaluation

Fiber Optics are the backbone of communication, and their security is essential in the present era. In recent years, Machine Learning and Deep Neural Networks have been employed in a wide range of applications for monitoring fiber optic communications. In this paper, we present two machine learning-...

ver descrição completa

Detalhes bibliográficos
Autor: Rivera Torres, Ronald
Tipo de documento: dissertação
Data de publicação:2022
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositório:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglês
OAI Identifier:oai:upcommons.upc.edu:2117/376702
Acesso em linha:https://hdl.handle.net/2117/376702
Access Level:Acceso aberto
Palavra-chave:Neural networks (Computer science)
Optical communications
Neural networks
AutoEncoders
Variational AutoEncoders
Optical Spectrum
Optical networks
Optical performance monitoring
Optical signals
Xarxes neuronals
codificadors automàtics
codificadors automàtics variacionals
espectre òptic
comunicacions òptiques
xarxes òptiques
monitorització del rendiment òptic
senyals òptics
Xarxes neuronals (Informàtica)
Comunicacions òptiques
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Telecomunicació òptica
Descrição
Resumo:Fiber Optics are the backbone of communication, and their security is essential in the present era. In recent years, Machine Learning and Deep Neural Networks have been employed in a wide range of applications for monitoring fiber optic communications. In this paper, we present two machine learning-based security systems, a Deep Neural Network (DNN) and AutoEncoders (AE) that use spectrum samples to forecast lightpath length and identify anomalies. As well, we tested the robustness of these systems against an attack use case, using Variational AutoEncoder (VAE) to generate synthetic spectrum samples, with the aim of deceiving the defense systems. The proposed DNN distance predictor correctly estimates distances and anomalies with an error of less than 56 km with a confidence interval of 95%. On the other hand, the proposed AutoEncoder, successfully classifies examples of the same distance with a 99% accuracy. When both methods were tested in our use case attack, they retained good performance and were able to accurately diagnose irregularities. Based on the findings, we can conclude that both defense mechanisms are suitable and reliable for detecting anomalies in fiber optic transmissions. Forming the basis for further research on developing monitoring systems for fiber optics communications based on spectrum samples.