Data-Mining Work Offers Insight into Detecting Multiple Hackers
Dr. Murat Kantarcioglu
Security efforts to combat hackers usually focus on one method of attack, but computer scientists at UT Dallas have developed a strategy more effective at tackling various types of attacks.
Dr. Murat Kantarcioglu, professor of computer science in the Erik Jonsson School of Engineering and Computer Science and director of the Data Security and Privacy Lab, and research scientist Dr. Yan Zhou have created a data-mining model that can identify various adversaries, or hackers.
Data mining, the process of analyzing big sets of data and organizing it into useful information, is used in all corners of industry, Kantarcioglu said.
“One area where adversaries commonly come into play is spam filtering,” he said. “In the early days, we would try to figure out whether an email was spam or legitimate by looking at the words contained within the body of the message. Adversaries, in this example, were anyone modifying emails to try and deceive the data-mining process.”
These adversaries come in an array of types. Some aim to send spam content to email users, while others want to clog networks by making the resource unavailable. Some spammers have the capability to modify spam and legitimate emails, while other have little to no access to such emails.
Kantarcioglu said it’s impossible to implement a filter that uses a single method to counter every possible type of spammer, motivating researchers to develop an “adversarial learning framework” that accounts for different types of hackers.
The team of researchers based its model on game theory principles, primarily for its resemblance to a two-player game. To tackle this particular challenge, they used a nested Stackelberg game framework, which is designed to handle both malicious data corruption and unknown types of adversaries.
He said that past work in the field has focused on developing data-mining models that are resilient against only one type of adversary. Computer scientists mimic adversarial behavior by looking at the end results, or the data extracted from a system. Researchers receive this input and determine whether they’re dealing with a malicious entity.
Kantarcioglu’s research presents a new, multitiered framework that simultaneously looks for adversarial data transformations and an optimal strategy to combat those changes. Zhou said these transformations are performed by the hacker to find the best way to modify data maliciously, allowing them to evade detection.
This mixed strategy is more reliable in situations when the data-mining applications are confronted by unknown adversaries.
Zhou said future applications of the research may expand the idea to include hackers working in tandem.
“In the current work, we assume the adversaries are independent of each other and their actions have no impact on each other's decisions,” Zhou said. “In the future, we will consider problems where there are multiple collaborative adversaries.”
Their work received the Best Application Paper Award when it was presented last spring at the 20th Pacific Asia Conference on Knowledge Discovery and Data Mining. The research was recently published in the conference’s report, Advances in Knowledge Discovery and Data Mining.
Media Contact: The Office of Media Relations, UT Dallas, (972) 883-2155, [email protected].