Abstract
Increasingly researchers are turning to machine learning techniques such as artificial neural networks (ANN) to address communication network research challenges in the areas of enhanced security, quality of service, visibility and control. Central to each is the need to classify packets. Determining an effective architecture for the artificial neural network is more difficult because traditional techniques such as principal component analysis (PCA) show reduced effectiveness. Presented are the techniques for preprocessing datasets and selecting input traffic features for the multi-layer perceptron (MLP) architecture. This methodology achieves classification accuracy above 99%.
An investigation into neural network architectures revealed the optimal structure and parameters for communication packet classification. This work also studies optimization algorithms with completely balanced datasets and provides performance criteria for training time and accuracy.
The application of MLPs to security challenges is also investigated. Port scans are a persistent problem on contemporary communication networks. Sequential MLPs are investigated to classify packets and determine TCP packet type. Following classification, analysis is performed in order to discover scan attempts. Neural networks can be used to successfully classify general packet traffic and more complex TCP classes at rates that are above 99\%. The proposed methodology achieves accurate scan detection without having to utilize an intrusion detection system.
In order to harness the power of Convolutional Neural Networks (CNNs), the conversion of packets to images is investigated. Additionally, a sequence of packets are combined into larger images to gain insight into conversations, exchanges, losses and threats. The use of this technique to identify potential latency problems is demonstrated. This approach of using contemporary network traffic and convolutional neural networks has success rate for individual packets exceeding 99%. Larger images achieve the same high level of accuracy. Finally, neural network ensembles are researched that reach 100% accuracy for packet classification.
Ensembles are also studied to accurately predict Mean Opinion Score for voice traffic and explored for their use in combating adversarial attacks against the source data.
Library of Congress Subject Headings
Neural networks (Computer science); Perceptrons; Packet switching (Data transmission)
Publication Date
5-2020
Document Type
Dissertation
Student Type
Graduate
Degree Name
Computing and Information Sciences (Ph.D.)
Department, Program, or Center
Computer Science (GCCIS)
Advisor
Andres Kwasinski
Advisor/Committee Member
Victor Perotti
Advisor/Committee Member
Minseok Kwon
Recommended Citation
Hartpence, Bruce, "Neural Network Architectures and Ensembles for Packet Classification: Addressing Visibility, Security and Quality of Service Challenges in Communication Networks" (2020). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/10486
Campus
RIT – Main Campus
Plan Codes
COMPIS-PHD
