Master of Engineering Science
Electrical and Computer Engineering
This thesis explores intelligent automation solutions, particularly Automated Machine Learning (AutoML) and Digital Twins (DTs), to enhance network management and security in Fifth-Generation (5G) networks. The first topic surveys Zero-touch network and Service Management (ZSM) systems, implementing an online AutoML pipeline to predict application throughput in Fourth Generation (4G) and 5G networks. Simulation results demonstrate the superiority of AutoML over traditional ML approaches, enabling ZSM to adapt to changing traffic patterns and, consequently, improving service quality and operational efficiency. The second topic implements an AutoML pipeline for multi-class network attack classification, incorporating real-time model updates using synthetic data within the DT environment. Uniform sampling for data generation is shown to offer faster recovery, lower overhead, and enhanced privacy. As for the generative models, variational autoencoders show lower overhead than generative adversarial models. However, they exhibit suboptimal performance when dealing with limited training samples, which can be attributed to their objective functions.
Summary for Lay Audience
In this thesis, the focus is on intelligent automation, which is like having a smart assistant that uses advanced technology to perform tasks faster and better with minimal human effort. The following topics are covered.
Zero-touch Network and Service Management (ZSM)
ZSM is a smart system that manages the internet and services without manual intervention. It ensures smooth streaming during crowded times, like movie nights, without user worry.
The study explores using online Automated Machine Learning (AutoML) for dynamic networks in ZSM. AutoML uses artificial intelligence to create computer programs automatically. In this case, AutoML is used to predict the application throughput in 4G and 5G networks. The throughput in a streaming application, for example, determines how fast video data can be downloaded/streamed to the user's device. This helps in optimizing network resource allocation, especially during heavy traffic.
The model is monitored and updated to handle changing data patterns. For instance, during a surge in streaming due to a popular music album release, the network becomes busier. The current model needs to learn from the new data to make better decisions and handle the increased demand for streaming.
AutoML for Network Attack Classification
The aim is to detect and classify network attacks for network protection. Real-time updates are crucial as new attacks emerge, preventing them from going undetected. Some Internet of Things devices with limited resources may struggle with these updates, so a digital twin is used to hold a virtual copy of the model and learn from synthetic data without compromising real data privacy.
One method of updating the security model is through uniform sampling, which collects statistics from the physical world and transmits them to the twin. From this information, data is generated and used to adjust the model's settings, which are sent back and loaded into the physical model. Uniform sampling results in fast recovery, low overhead, and intact privacy. Generative models, such as Variational AutoEncoders (VAEs) and Generative Adversarial Networks (GANs), are also explored. They learn from existing data and create new similar content. VAEs have lower overhead than GANs, but limited training data affects their recovery performance.
El Rajab, Mirna, "Intelligent Automation Solutions for Network Management and Security in 5G Networks: A Study on AutoML and Digital Twins" (2023). Electronic Thesis and Dissertation Repository. 9584.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Available for download on Saturday, August 24, 2024