Resource Consumption Analysis of Distributed Machine Learning Models for 6G Security

Abstract

Communications networks have become increasingly complex environments due to the massive increase in the number of communicating nodes and diverse services with unique requirements. Therefore, machine learning has become extremely important from the access to backhaul and core networks, as well as various technologies required for the smooth operations of different tasks and services within those networks. The overall complexity of networked environments and increasing volumes of data further complicates the network security landscape. Machine learning with its various techniques and tools, thus, has become vital for network security. In 6G network security, the promises of machine learning are vast, from preventive measures to detection to response and remediation. However, machine learning requires a huge amount of resources mainly due to the fact that machine learning operates on data and data volumes are consistently rising. This work studied and investigated the resource consumption of machine learning techniques used for network security to provide insights into the potential resource implications of deploying machine learning in 6G security.

The thesis explored a wide range of state-of-the-art resource-efficient Machine learning based security solutions to find out the key resources consumed by those solutions and the key enablers of resource efficiency for those solutions. In particular, the thesis focused on investigating the resource consumption of distributed learning for 6G networks in terms of computing, memory, bandwidth, energy, latency, and human resources. Distributed machine learning is highly relevant to the context of 6G, as it can meet the future 6G requirement of processing substantial amounts of data generated from numerous devices while preserving data privacy and security. The thesis presents an experimental and comparative analysis of the Federated Learning (FL) and Split Learning (SL) based network security solutions, which are the two most popular distributed learning in terms of resource consumption fingerprinting. The finding shows that both models perform well, while Federated Learning appears to have a slight edge over Split Learning in terms of precision and F1 score. However, the differences are quite small. In terms of resource consumption fingerprinting, we observed that both of them have their advantages and shortcomings. In terms of CPU usage, SL had higher CPU usage, while FL had higher peaks and variability. In terms of memory usage, FL was more memory efficient than the SL. Finally, SL was more time and power-efficient and had lower CO2 emission.