Master's Thesis

Anomaly Detection in State of Polarization Signals Using Machine Learning for Securing Fiber-Optic Networks

Final Thesis 4.63 MB Appendix 4.1 MB

Author of thesis: Joshua Nzube Nwokoye, B.Sc.

Acad. year: 2025/2026

Supervisor: Ing. Adrián Tomašov, Ph.D.

Reviewer: doc. Ing. Tomáš Horváth, Ph.D.

Abstract:

The increasing reliance on fiber optic networks for critical infrastructure necessitates robust Physical Layer Security mechanisms. This thesis investigates the use of the State of Polarization (SoP) as a highly sensitive, passive indicator for intrusion detection. SoP is a fundamental optical property inherently sensitive to external birefringence perturbations, such as macrobending, microbending, or physical tapping. The core goal is to leverage machine learning to distinguish the stochastic, normal SoP drift (caused by Polarization Mode Dispersion) from localized, rapid deviations indicative of tampering.

The research is structured in two major phases. The semestral phase focuses on the acquisition and meticulous preprocessing of SoP time series data, including Stokes vector normalization, detrending, and advanced feature engineering (e.g., Wavelet and time frequency analysis). The diploma thesis implements deep neural network architectures, primarily Recurrent Autoencoders (RAEs), for unsupervised anomaly detection. These models are trained exclusively on statistically characterized baseline SoP behavior. Evaluation demonstrates that this framework achieves high sensitivity and accuracy in detecting subtle, localized physical disturbances by quantifying the reconstruction error, thereby offering a robust and scalable solution for proactive optical network security.

Keywords:

Anomaly Detection, State of Polarization (SoP), Fiber Optic Security, Deep Autoencoders, Recurrent Neural Networks, Unsupervised Learning, Time Series Analysis, Physical Layer Security, Polarization Monitoring, Birefringence Perturbations, Stokes Parameters, Polarization Mode Dispersion.

Date of defence

09.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaAznamka

Grading

A

Process of defence

Student presented the results of his thesis and the committee got familiar with reviewer's report. Student defended his Diploma Thesis. and answered the questions from the members of the committee and the reviewer

Language of thesis

English

Faculty

Fakulta elektrotechniky a komunikačních technologií

Department

Department of Telecommunications

Study programme

Communications and Networking (MPAD-CAN)

Composition of Committee

prof. Ing. Zdeněk Smékal, CSc. (předseda)
doc. Ing. Ivo Lattenberg, Ph.D. (místopředseda)
doc. Ing. Lukáš Malina, Ph.D. (člen)
Ing. Štěpán Miklánek, Ph.D. (člen)
Ing. Jiří Přinosil, Ph.D. (člen)
Ing. Adrián Tomašov, Ph.D. (člen)
Ing. et Ing. Petr Musil (člen)
Ing. Filip Wagner (člen)

Supervisor’s report
Ing. Adrián Tomašov, Ph.D.

The diploma thesis investigates the possibility of multimodal time series anomaly detection, particularly semi-supervised anomaly detection of Stokes vectors from optical polarization in fiber-optic cables. The thesis is well-structured and provides a great introduction to the domain. It is well-written and formally correct; however, the quality of some result figures would benefit from being rendered in vector graphics. The student found and cited suitable literature relevant to the topic. Furthermore, the thesis demonstrates that combining various metrics into a single model is not only possible but also beneficial. The student was highly active in discussions during the semester and brought their own ideas to the model's implementation. Considering all of the above, I award 91 points. Points proposed by supervisor: 91

Grade proposed by supervisor: A

Reviewer’s report
doc. Ing. Tomáš Horváth, Ph.D.

This master's thesis deals with a topic that is very important in an interdisciplinary way: optical communications, physical layer security, signal processing, and machine learning. The motivation is clearly described by proposing a system that can detect any kind of physical disturbance on optical fiber links such as bending, tapping or tampering by measuring changes on the State of Polarization.

The thesis gives a thorough theoretical background on security aspects in optical fibers, polarization monitoring, Stokes parameters, anomaly detection, and unsupervised learning algorithms. Its experimental work is really outstanding as it includes the acquisition, preprocessing, representation based on FFT/MFCC based signals, automatic labeling, and evaluation of the proposed anomaly detection methods.

The experimental setup is clear and pragmatic. By comparing the LOF and VAE models it is shown that stable polarization state could be differentiated from transition/anomaly events. Especially the LOF method results obtained are really good in term of precision, recall, F1-score and ROC AUC.

The work is a well written, technically relevant thesis and the author shows that he/she can mix optical network issues and machine learning approaches to tackle those problems. The results achieved were convincing and the contribution will be useful for the real time detection of any disturbance occurring on physical layer of fiber optic systems. I recommend this thesis for defense with grade A and 92 points. Topics for thesis defence:
  1. The LOF model outperformed the VAE in the reported experiments․ What do you think was the main reason for this difference? When is a VAE or other deep-learning model better than LOF?
  2. The evaluation is mainly focused on the normalized S2 signal. How strong do you expect the detection to be if we consider the three normalized Stokes components i.e. s1‚ s2‚ and s3 as the input of the model?
Points proposed by reviewer: 92

Grade proposed by reviewer: A

Responsibility: Mgr. et Mgr. Hana Odstrčilová