Leveraging machine learning for security related decision making
The need to ensure the primary functionality of any system means that considerations of security are often secondary. Computer security considerations are made in relation to considerations of usability, functionality, productivity, and other goals. Decision-making related to security is about finding an appropriate tradeoff. Most existing security mechanisms take a binary approach where an action is either malicious or benign, and therefore allowed or denied. However, security and privacy outcomes are often fuzzy and cannot be represented by a binary decision. It is useful for end users, who may ultimately need to allow or deny an action, to understand the potential differences among objects and the way that these differences are communicated matters. In this work, we use machine learning and feature extraction techniques to model normal behavior in various contexts and then used those models to detect the degree that new behavior is anomalous. This measurement can then be used, not as a binary signal but as a more nuanced indicator that can be communicated to a user to help guide decision-making. We examine the application of this idea in two domains. The first is the installation of applications on a mobile device. The focus in this domain is on permissions that represent capabilities and access to data, and we generate a model for expected permission requests. Various user studies were conducted to explore effective ways to communicate this measurement to influence decision-making by end users. Next, we examined to the domain of insider threat detection in the setting of a source code repository. The goal was to build models of expected user access and more appropriately predict the degree that new behavior deviates from the previous behavior. This information can be utilized and understood by security personnel to focus on unexpected patterns.
Li, Purdue University.
Artificial intelligence|Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our