Date of Award

Spring 3-2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Cyber Operations (PhDCO)

Department

Computer Science

First Advisor

Yong Wang

Second Advisor

Jun Liu

Third Advisor

Wayne Pauli

Fourth Advisor

Josh Pauli

Fifth Advisor

Joshua Stroschein

Abstract

This dissertation explores functional malware classification using running window entropy and machine learning classifiers. This topic was under researched in the prior literature, but the implications are important for malware defense. This dissertation will present six new design science artifacts. The first artifact was a generalized machine learning based malware classifier model. This model was used to categorize and explain the gaps in the prior literature. This artifact was also used to compare the prior literature to the classifiers created in this dissertation, herein referred to as “Malgazer” classifiers.

Running window entropy data was required, but the algorithm was too slow to compute at scale. This dissertation presents an optimized version of the algorithm that requires less than 2% of the time of the original algorithm. Next, the classifications for the malware samples were required, but there was no one unified and consistent source for this information. One of the design science artifacts was the method to determine the classifications from publicly available resources.

Once the running window entropy data was computed and the functional classifications were collected, the machine learning algorithms were trained at scale so that one individual could complete over 200 computationally intensive experiments for this dissertation. The method to scale the computations was an instantiation design science artifact. The trained classifiers were another design science artifact. Lastly, a web application was developed so that the classifiers could be utilized by those without a programming background. This was the last design science artifact created by this research.

Once the classifiers were developed, they were compared to prior literature theoretically and empirically. A malware classification method from prior literature was chosen (referred to herein as “GIST”) for an empirical comparison to the Malgazer classifiers. The best Malgazer classifier produced an accuracy of approximately 95%, which was around 0.76% more accurate than the GIST method on the same data sets. Then, the Malgazer classifier was compared to the prior literature theoretically, based upon the empirical analysis with GIST, and Malgazer performed at least as well as the prior literature. While the data, methods, and source code are open sourced from this research, most prior literature did not provide enough information or data to replicate and verify each method. This prevented a full and true comparison to prior literature, but it did not prevent recommending the Malgazer classifier for some use cases.

Download

Share

COinS