Date of Award

Spring 3-2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Cyber Operations (PhDCO)

Department

Computer Science

First Advisor

Wayne E. Pauli

Second Advisor

Michael Ham

Third Advisor

Tom Halverson

Fourth Advisor

David Bishop

Abstract

The growing sophistication of attacks and newly emerging cyber threats requires advanced cyber threat detection systems. Although there are several cyber threat detection tools in use, cyber threats and data breaches continue to rise. This research is intended to improve the cyber threat detection approach by developing a cyber threat detection framework using two complementary technologies, search engine and machine learning, combining artificial intelligence and classical technologies.

In this design science research, several artifacts such as a custom search engine library, a machine learning-based engine and different algorithms have been developed to build a new cyber threat detection framework based on self-learning search and machine learning engines. Apache Lucene.Net search engine library was customized in order to function as a cyber threat detector, and Microsoft ML.NET was used to work with and train the customized search engine.

This research proves that a custom search engine can function as a cyber threat detection system. Using both search and machine learning engines in the newly developed framework provides improved cyber threat detection capabilities such as self-learning and predicting attack details. When the two engines run together, the search engine is continuously trained by the machine learning engine and grow smarter to predict yet unknown threats with greater accuracy. While customizing the search engine to function as a cyber threat detector, this research also identified and proved the best algorithms for the search engine based cyber threat detection model. For example, the best scoring algorithm was found to be the Manhattan distance. The validation case study also shows that not every network traffic feature makes an equal contribution to determine the status of the traffic, and thus the variable-dimension Vector Space Model (VSM) achieves better detection accuracy than n-dimensional VSM.

Although the use of different technologies and approaches improved detection results, this research is primarily focused on developing techniques rather than building a complete threat detection system. Additional components such as those that can track and investigate the impact of network traffic on the destination devices make the newly developed framework robust enough to build a comprehensive cyber threat detection appliance.

Download

Share

COinS