Date on Master's Thesis/Doctoral Dissertation

8-2017

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Kantardzic, Mehmed

Committee Co-Chair (if applicable)

Elmaghraby, Adel

Committee Member

Elmaghraby, Adel

Committee Member

Yampolskiy, Roman

Committee Member

Lauf, Adrian

Committee Member

Hieb, Jeff

Author's Keywords

streaming data; concept drift; adversarial learning; machine learning; cybersecurity

Abstract

While understanding of machine learning and data mining is still in its budding stages, the engineering applications of the same has found immense acceptance and success. Cybersecurity applications such as intrusion detection systems, spam filtering, and CAPTCHA authentication, have all begun adopting machine learning as a viable technique to deal with large scale adversarial activity. However, the naive usage of machine learning in an adversarial setting is prone to reverse engineering and evasion attacks, as most of these techniques were designed primarily for a static setting. The security domain is a dynamic landscape, with an ongoing never ending arms race between the system designer and the attackers. Any solution designed for such a domain needs to take into account an active adversary and needs to evolve over time, in the face of emerging threats. We term this as the ‘Dynamic Adversarial Mining’ problem, and the presented work provides the foundation for this new interdisciplinary area of research, at the crossroads of Machine Learning, Cybersecurity, and Streaming Data Mining. We start with a white hat analysis of the vulnerabilities of classification systems to exploratory attack. The proposed ‘Seed-Explore-Exploit’ framework provides characterization and modeling of attacks, ranging from simple random evasion attacks to sophisticated reverse engineering. It is observed that, even systems having prediction accuracy close to 100%, can be easily evaded with more than 90% precision. This evasion can be performed without any information about the underlying classifier, training dataset, or the domain of application. Attacks on machine learning systems cause the data to exhibit non stationarity (i.e., the training and the testing data have different distributions). It is necessary to detect these changes in distribution, called concept drift, as they could cause the prediction performance of the model to degrade over time. However, the detection cannot overly rely on labeled data to compute performance explicitly and monitor a drop, as labeling is expensive and time consuming, and at times may not be a possibility altogether. As such, we propose the ‘Margin Density Drift Detection (MD3)’ algorithm, which can reliably detect concept drift from unlabeled data only. MD3 provides high detection accuracy with a low false alarm rate, making it suitable for cybersecurity applications; where excessive false alarms are expensive and can lead to loss of trust in the warning system. Additionally, MD3 is designed as a classifier independent and streaming algorithm for usage in a variety of continuous never-ending learning systems. We then propose a ‘Dynamic Adversarial Mining’ based learning framework, for learning in non-stationary and adversarial environments, which provides ‘security by design’. The proposed ‘Predict-Detect’ classifier framework, aims to provide: robustness against attacks, ease of attack detection using unlabeled data, and swift recovery from attacks. Ideas of feature hiding and obfuscation of feature importance are proposed as strategies to enhance the learning framework's security. Metrics for evaluating the dynamic security of a system and recover-ability after an attack are introduced to provide a practical way of measuring efficacy of dynamic security strategies. The framework is developed as a streaming data methodology, capable of continually functioning with limited supervision and effectively responding to adversarial dynamics. The developed ideas, methodology, algorithms, and experimental analysis, aim to provide a foundation for future work in the area of ‘Dynamic Adversarial Mining’, wherein a holistic approach to machine learning based security is motivated.