Daniel Balasubramanian, a senior research scientist at Vanderbilt’s Institute for Software Integrated Systems, will lead a four-year $6.89 million grant from the Defense Advanced Research Projects Agency (DARPA) to create realistic network environments used to train cyber agents to counter advanced and persistent cyber threats.
Estimates have placed the cost of global cybercrime as high as $9.5 trillion in 2024 according to Cybersecurity Ventures, and a single data breach can have costs in the millions of dollars. As the scale and complexity of cyberattacks grows, network operators need the ability to understand and defend against these attacks in real time.
“Anything we can do to reduce and lessen the severity of such attacks has a huge economic benefit,” Balasubramanian said. “Beyond the financial impact, safeguarding these networks is crucial to our national security.”
The new DARPA grant, “Reinforcement Against Malicious Penetration by Adversaries in Realistic Topologies (RAMPART),” is a joint effort among a Vanderbilt team at ISIS and its partners: Jack W. Davidson, a recognized expert in network security and professor of computer science at the University of Virginia, and Paul Roysdon, chief solutions architect of artificial intelligence and machine learning at Leidos.
The goal is to build a state-of-the-art learning environment where teams can learn to defend against incoming attacks as well as search for system vulnerabilities.
“Our approach to training cyber agents uses a combination of model-based representations and deep reinforcement learning (RL) to develop autonomous agents that are capable of defending a network against attacks” Balasubramanian said. “Augmenting human operators with automated agents will be crucial to defending against the next generation of cyberattacks.”
Vanderbilt team members include Xenofon Koutsoukos, Thomas R. Walter Professor and chair of the Department of Computer Science; Kevin Leach, computer science assistant professor; Himanshu Neema, senior research scientist; and Sandeep Neema, computer science professor and chair of the ISIS Executive Council.
The researchers will model small- to medium-sized computer networks and use these models to train autonomous agents to defend the network against cyberattacks. In the long term, they hope to scale this to enterprise-scale networks, where an autonomous agent can defend a network consisting of thousands of nodes.
“The task of developing such agents is daunting at multiple dimensions, but we believe RAMPART is a unique approach that is up to the various challenges,” Balasubramanian said.
Contact: brenda.ellis@vanderbilt.edu