Playing games is a universal human experience. Every country and culture plays games unique and ubiquitous to its people, from the ancient Chinese game Go to modern computer games such as SimCity.

But games are more than a human hobby. They’re a training ground for artificial intelligence (AI) to learn fundamental decision-making skills, handle uncertainty, and tackle complex interactions. AI’s ability to tirelessly explore thousands of possibilities makes it the perfect tool for finding strategies humans might overlook.

Angelos Stavrou, professor in the Bradley Department of Electrical and Computer Engineering and founder of mobile security startup A2 Labs, plans to take AI training to the next level with a multi-agent training exerciser (MATrEx). Through this “matrix,” AI faces a new opponent — other AI — but the implications go beyond a fun game.

“What we want to do is take a network, like the Virginia Tech network, copy it into a virtual environment, and then run simulations of attacks and defenses with logic generated by AI agents,” said Stavrou, who’s also entrepreneurship lead for the Virginia Tech Innovation Campus. “AI can play billions of game variations so they can actually characterize the cybersecurity start of the system, and decide which attacks or defenses are superior.”

Playing the game

In a world plagued by persistent and never-ending cyber threats, improving the security of computer networks by minimizing vulnerabilities is essential for protecting cell phones, the 5G wireless system, and the thousands of data centers in the United States. The process is called “network hardening.” It’s like adding strong locks to a home and reinforcing its walls. 

MATrEx is a $16 million project funded by the Defense Advanced Research Projects Agency as part of Cyber Agents for Testing and Learning Environments. It’s a trailblazing way to train AI in a three-tiered gaming system – simulations, emulations, and a real network. Within the MATrEx, AI agents are grouped into two categories:

  • Red agents, that need to learn the gaming network and how best to break through system defenses for attacks
  • Blue agents that understand the network fully, including defense capabilities

The “games” are scenarios red and blue agents run against each other. Researchers will study what attacks break through, which defenses stop the attacks, what happens during a specific attack, and more.

“The idea is that we’ll create a loop, so the AI can play in a network that will start from a real network, then emulate it and simulate it,” said Stavrou. “The real system is accurate, but it’s slow. The simulation is fast, but it’s not as accurate. So you need to play the game across all these environments, transferring knowledge from the real network to the simulated. Then you transfer strategies from the simulated back to the real network.”

A descriptive graphic of how the red and blue agents interact with the different networks within the MATrEx.
The graphic demonstrates how red and blue agents go through a message bus to receive instructions prior to entering one of the networks for gameplay. Graphic courtesy of Angelos Stavrou.

The distinctive combination of real, simulated, and emulated networks sets MATrEx apart from any other AI training system in the world. Each network provides a training ground – and researcher data set – for the AI.

  • Real network: This is the physical computer or wireless network used by people every day. It has actual switches, routers, cables and devices that connect together to transmit data. In MATrEx, this is a copy of a company’s real network, provided and hosted by the company themselves.

  • Emulated network: This “digital twin” of the real network is a collection of virtual machines that run the actual operating system from the real network, and also mirrors its network configuration.

  • Simulated network: Also called a “state machine,” this network captures the behaviors of the computers within a network, without having to utilize the network’s real operating system. Because of that, actions can run much faster than in real time.

As the AI agents run games through each of these networks, the agents grow smarter and the neworks increase in complexity, providing clients the ability to bulk up or restrategize their real, non-MATrEx networks.

Securing the future

Security is an arms race, and the “good guys,” like researchers in cybersecurity, aren’t the only ones with access to AI agents.

“The work hackers do opens up many more possibilities for attack than it does defense, because attackers only have to succeed once, even if they fail 1,000 times,” said Stavrou. “We’re opening Pandora’s box, but we’re trying to understand the effects of AI in cybersecurity.”

The end for MATrEx is to scientifically assess the red agent capabilities – what new strategies will they develop – and to teach the blue agents to outsmart their attackers. MATrEx will be designed for companies to run the program internally, so their protected data never leaves and all developed agent strategies stay contained. It’s the first training environment that will not only test on traditional network-connected servers, but also 5G.

“There has never been a 5G system connected to a cybersecurity testbed with this level of emulation,” said Stavrou. “5G networks are fairly new, and they’re usually enclosed within big companies. This will allow us to see what will happen when AI agents attack or defend that type of network and connectivity.”

Adam Gorski installs CCI xG Testbed components

Adam Gorski installs CCI xG Testbed components
Adam Gorski installs xG Testbed components for the Commonwealth Cyber Initiative, which has capabilities of securing 5G and next-generation mobile networks and AI assurance. MATrEx will utilize a 5G system, making it a one-of-a-kind cybersecurity test bed. Photo by Anthony Wright for Virginia Tech.

While Virginia Tech provides scientific infrastructure and a connection to the Commonwealth Cyber Initiative xG testbed in Arlington, four additional partners will support the MATrEx through large-scale network simulation and emulation, cyber range setup, agent learning optimization, management, and more:

Over the next four years of the project, Stavrou is looking for more than AI to play games; MATrEx will serve as a cyber training ground where humans can play, train, and learn about cybersecurity with a trained red agent built to win the game.

He also wants to build out an educational component that will support hands-on student learning and the future of the Innovation Campus.

“The goal of this project isn’t just to create a company, but to develop a company-university collaboration,” said Stavrou. “On the Innovation Campus, we’d like to have a space where we can showcase our work, a one-of-a-kind teaching lab for AI, for students to create projects. Really, we want it to be anything you can imagine.”

Contact:

Tags

Share this story