Advances in AI will enable new capabilities and make others far more affordable – not only to the U.S., but to adversaries as well, raising the stakes as the United States seeks to preserve its hard-won strategic overmatch in the air, land, sea, space and cyberspace domains.
The Pentagon’s Third Offset Strategy seeks to leverage AI and related technologies in a variety of ways, according to Robert Work, former deputy secretary of defense and one of the strategy’s architects. In a forward to a new report from the market analytics firm Govini, Work says the strategy “seeks to exploit advances in AI and autonomous systems to improve the performance of Joint Force guided munitions battle networks” through:
- Deep learning machines, powered by artificial neural networks and trained with big data sets
- Advanced human-machine collaboration in which AI-enabled learning machines help humans make more timely and relevant combat decisions
- AI devices that allow operators of all types to “plug into and call upon the power of the entire Joint Force battle network to accomplish assigned missions and tasks”
- Human-machine combat teaming of manned and unmanned systems
- Cyber- and electronic warfare-hardened, network-enabled, autonomous and high-speed weapons capable of collaborative attacks
“By exploiting advances in AI and autonomous systems to improve the warfighting potential and performance of the U.S. military,” Work says, “the strategy aims to restore the Joint Force’s eroding conventional overmatch versus any potential adversary, thereby strengthening conventional deterrence.”
Spending is growing, Govini reports, with AI and related defense program spending increasing at a compound annual rate of 14.5 percent from 2012 to 2017, and poised to grow substantially faster in coming years as advanced computing technologies come on line, driving down computational costs.
But in practical terms, what does that mean? How will AI change the way defense technology is managed, the way we gather and analyze intelligence or protect our computer systems?
Charlie Greenbacker, vice president of analytics at In-Q-Tel in Arlington, Va., the intelligence community’s strategic investment arm, sees dramatic changes ahead.
“The incredible ability of technology to automate parts of the intelligence cycle is a huge opportunity,” he said at an AI summit produced by the Advanced Technology Academic Research Center and Intel in November. “I want humans to focus on more challenging, high-order problems and not the mundane problems of the world.”
The opportunities are possible because of the advent of new, more powerful processing techniques, whether by distributing those loads across a cloud infrastructure, or using specialty processors purpose-built to do this kind of math. “Under the hood, deep learning is really just algebra,” he says. “Specialized processing lets us do this a lot faster.”
Computer vision is one focus of interest – learning to identify faces in crowds or objects in satellite or other surveillance images – as is identifying anomalies in cyber security or text-heavy data searches. “A lot of folks spend massive amounts of time sifting through text looking for needles in the haystack,” Greenbacker continued.
The Air Force is looking at AI to help more quickly identify potential cyber attacks, said Frank Konieczny, chief technology officer in the office of the Air Force chief information officer, speaking at the CyberCon 2017 in November. “We’re looking at various ways of adjusting the network or adjusting the topology based upon threats, like software-defined network capabilities as well as AI-based analysis,” he said.
Marty Trevino Jr., a former technical director and strategist for the National Security Agency, now chief data/analytics officer at intelligence specialist at Red Alpha, a tech firm based in Annapolis Junction, Md. “We are all familiar with computers beating humans in complex games – chess, Go, and so on,” Trevino says. “But experiments are showing that when humans are mated with those same computers, they beat the computer every time. It’s this unique combination of man and machine – each doing what its brain does best – that will constitute the active cyber defense (ACD) systems of tomorrow.”
Machines best humans when the task is highly defined at speed and scale. “With all the hype around artificial intelligence, it is important to understand that AI is only fantastic at performing the specific tasks to which it is intended,” Trevino says. “Otherwise AI can be very dumb.”
Humans on the other hand, are better than machines when it comes to putting information in context. “While the human brain cannot match AI in specific realms,” he adds, “it is unmatched in its ability to process complex contextual information in dynamic environments. In cyber, context is everything. Context enables data-informed strategic decisions to be made.”
Looking at cybersecurity another way, AI can also be used to rapidly identify and repair software vulnerabilities, said Brian Pierce, director of the Information Innovation Office at the Defense Advanced Research Projects Agency (DARPA).
“We are using automation to engage cyber attackers in machine time, rather than human time,” he said. Using automation developed under DARPA funding, he said machine-driven defenses have demonstrated AI-based discovery and patching of software vulnerabilities. “Software flaws can last for minutes, instead of as long as years,” he said. “I can’t emphasize enough how much this automation is a game changer in strengthening cyber resiliency.”
Such advanced, cognitive ACD systems employ the gamut of detection tools and techniques, from heuristics to characteristic and signature-based identification, says Red Alpha’s Trevino. “These systems will be self-learning and self-healing, and if compromised, will be able to terminate and reconstitute themselves in an alternative virtual environment, having already learned the lessons of the previous engagement, and incorporated the required capabilities to survive. All of this will be done in real time.”
Seen in that context, AI is just the latest in a series of technologies the U.S. has used as a strategic force multiplier. Just as precision weapons enabled the U.S. Air Force to inflict greater damage with fewer bombs – and with less risk – AI can be used to solve problems that might otherwise take hundreds or even thousands of people. The promise is that instead of eyeballing thousands of images a day or scanning millions of network actions, computers can do the first screening, freeing up analysts for the harder task of interpreting results, says Dennis Gibbs, technical strategist, Intelligence and Security programs at General Dynamics Information Technology. “But just because the technology can do that, doesn’t mean it’s easy. Integrating that technology into existing systems and networks and processes is as much art as science. Success depends on how well you understand your customer. You have to understand how these things fit together.”
In a separate project, DARPA collaborated with a Fortune 100 company that was moving more than a terabyte of data per day across its virtual private network, and generating 12 million network events per day – far beyond the human ability to track or analyze. Using automated tools, however, the project team was able to identify a single unauthorized IP address that successfully logged into 3,336 VPN accounts over seven days.
Mathematically speaking, Pierce said, “The activity associated with this address was close to 9 billion network events with about a 1 in 10 chance of discovery.” The tipoff was a flaw in the botnet that attacked the network: Attacks were staged at exactly 57-minute intervals. Not all botnets of course, will make that mistake. But even pseudo random timing can also be detected. He added: “Using advanced signal processing methods applied to billions of network events over weeks and months-long timelines, we have been successful at finding pseudo random botnets.”
On the flipside however, must be the recognition that AI superiority will not be a given in cyberspace. Unlike air, land, sea or space, cyber is a man-made warfare domain. So it’s fitting that the fight there could end up being machine vs. machine.
The Harvard Artificial Intelligence and National Security study notes emphatically that while AI will make it easier to sort through ever greater volumes of intelligence, “it will also be much easier to lie persuasively.” The use of Photoshop and other image editors is well understood and has been for years. But recent advances in video editing have made it reasonably easy to forge audio and video files.
A trio of University of Washington researchers announced in a research paper published in July that they had used AI to synthesize a photorealistic, lip-synced video of former President Barack Obama. While the researchers used real audio, it’s easy to see the dangers posed if audio is also manipulated.
While the authors describe potential positive uses of the technology – such as “the ability to generate high-quality video from audio [which] could significantly reduce the amount of bandwidth needed in video coding/transmission” – potential nefarious uses are just as clear.