Semester

Spring

Date of Graduation

2019

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Katerina Goseva-Popstojanova

Committee Co-Chair

Vinod Kulathumani

Committee Member

Vinod Kulathumani

Committee Member

Roy S. Nutter

Committee Member

Stacy Prowell

Committee Member

Yanfang Ye

Abstract

Although malware detection is a very active area of research, few works were focused on using physical properties (e.g., power consumption) and multimodal features for malware detection. We designed an experimental testbed that allowed us to run samples of malware and non-malicious software applications and to collect power consumption, network traffic, and system logs data, and subsequently to extract dynamic behavioral-based features. We also extracted code-based static features of both malware and non-malicious software applications. These features were used for malware detection based on: feature level fusion using power consumption and network traffic data, feature level fusion using network traffic data and system logs, and multimodal feature level and decision level fusion.

The contributions when using feature level fusion of power consumption and network traffic data are: (1) We focused on detecting real malware using the extracted dynamic behavioral features (both power-based and network traffic-based) and supervised machine learning algorithms, which has not been done by any of the prior works. (2) We ran a large number of machine learning experiments, which allowed us to identify the best performing learner, DC voltage rails that led to the best malware detection performance, and the subset of features that are the best predictors for malware detection. (3) The comparison of malware detection performance was done using a comprehensive set of metrics that reflect different aspects of the quality of malware detection.

In the case of the feature level fusion using network traffic data and system logs, the contributions are: (1) Most of the previous works that have used network flows-based features have done classification of the network traffic, while our focus was on classifying the software running in a machine as malware and non-malicious software using the extracted dynamic behavioral features. (2) We experimented with different sizes of the training set (i.e., 90%, 75%, 50%, and 25% of the data) and found that smaller training sets produced very good classification results. This aspect of our work has a practical value because the manual labeling of the training set is a tedious and time consuming process.

In this dissertation we present a multimodal deep learning neural network that integrates different modalities (i.e., power consumption, system logs, network traffic, and code-based static data) using decision level fusion. We evaluated the performance of each modality individually, when using feature level fusion, and when using decision level fusion. The contributions of our multimodal approach are as follow: (1) Collecting data from different modalities allowed us to develop a multimodal approach to malware detection, which has not been widely explored by prior works. Even more, none of the previous works compared the performance of feature level fusion with decision level fusion, which is explored in this dissertation. (2) We proposed a multimodal decision level fusion malware detection approach using a deep neural network and compared its performance with the performance of feature level fusion approaches based on deep neural network and standard supervised machine learning algorithms (i.e., Random Forest, J48, JRip, PART, Naive Bayes, and SMO).

Embargo Reason

Publication Pending

Download

Share

COinS