How Special Operators Are Taking Artificial Intelligence To War
Data and machine learning will steer missions and predict uprisings before they start.
The U.S. fight against the Islamic State and other extremist threats is increasingly in the hands of elite special operations units who will succeed or fail by their ability to collect, process, and exploit data at the speed of crisis. At the command level, that means reducing the number of analysts required to get data to make sense. On the ground, it means sending much more actionable data to the tip of the spear, and doing so faster and more cheaply. Even the best tech minds in commercial sector don’t produce the sort of product that special operators need, according to special operations intelligence experts.
Today, the cutting edge looks something like this: Imagine the world’s highest-resolution commercial satellite, the WorldView 3 from satellite image provider DigitalGlobe floating 383 miles above the Earth’s surface. It snaps pictures of a residential neighborhood in a bustling African metropolis. One particular city street is free of cars. That’s common in parts of this country but unusual for the blocks surrounding the U.S. Embassy. Less than three hours later, a member of SOCOM takes notice. He’s been living in that country for years, and has inroads with local military units who protect the fragile government from overthrow.
Electricity is spotty, and communications are far from secure, but the operator needs nothing more than a standard laptop running the Chrome browser to pull in the satellite imagery. He reviews shots of the area from the previous several months to confirm that today’s lack of traffic is abnormal.
He opens a layer on his map, and red clouds spread out where open-source intelligence has predicted civil unrest. A third layer shows tweets from the area that mention a recently killed militia member. A flash of circles—green and yellow, large and small—show the locations of the tweeters. One of the accounts is associated with three attempted bombings in other parts of the country, but is new to the capital. So the operator shares the view with his contact in the local tactical police unit, providing the Twitter handles and locations of the people he’s most worried about. A minute later, he opens a topographical layer and finds a flat spot to land a special operations helicopter. The embassy is evacuated just minutes before a car bomb explodes.
Some of these capabilities are already in use by special operators. “Say that an event happened and you needed to figure out how to do an evacuation; you’re coming in during the day and using urban tactics to come in,” Paul Millhouse, DoD and Federal /Civilian Technical Solutions at DigitalGlobe explained to Defense One during a live demo. “In literally 5 seconds, you’re going to have an overlay that will tell you, ‘Here are the areas that you should focus on for landing.’ To put this into perspective, if you were to do this the traditional way, you would have to download gigabytes of imagery, gigs of elevation data, put it on a high-end machine, get expensive software, run it all – four to eight hours’ worth of work.” Milhouse said that the DigitalGlobe tools requires little bandwidth. “Rather than transmitting these input ingredients and performing the analysis on a high-end workstation, we perform the processing where the data is, and only send the resulting 10-kilobyte analysis overlay. All the hard work happens in the background.”
He said the hardest part is the second layer: predicting hotspots where civil unrest might occur. That’s created by a mixture of proprietary software and survey data collected on the ground, and it’s already in use by special operations forces.
The company’s newest offering is perhaps its most ambitious yet. On Tuesday, DigitalGlobe announced Vricon, a new partnership with Saab to create a fully accurate 3D model of the Earth. It’s primarily aimed at the commercial marketplace, but could have relevance for special operations forces. “Identifying safe locations for infiltration or exfiltration, conducting radio-frequency propagation analysis for communications planning, and route planning all require high-resolution elevation data. Our Vricon joint venture will enable SOF operators to reliably make shareable 3D solutions available to enable coalition efforts to counter emerging threats around the world,” DigitalGlobe senior vice president Tony Frazier told Defense One in an email.
Neither DigitalGlobe nor anyone in Silicon Valley can solve some of the biggest challenges that face the special operations forces community: the nation is relying on them to solve too many problems in too many places. But they just might be able to help SOF with data issues.
Special operations forces have two big problems: they need better intelligence extracted from data and they need to be able to collect it and deliver it in in an unclassified setting under challenging communications situations. The solution to both problems involves teaching software to learn to discriminate useful intelligence from raw, unstructured data, a subfield of artificial intelligence or AI.
Today’s operators have more raw data than battlefield bandwidth. “Something DoD needs to get its arms around” is “moving sensor data around the battlefield to the folks who need to exploit it,” said Air Force Col. Matthew D. Atkins, chief of the intel capabilities and requirements division at United States Special Operations Command at a recent industry event in Tampa, Florida. “We do need help solving the data transport problem from a technological perspective,” he told a group of industry representatives at a recent conference in Tampa, Florida. “Every time we roll out a new [high-definition] sensor, a new widget, we crush the data rates.”
It’s a problem that’s only going to grow as data-collection devices improve and special operations forces expand the sorts of data they use. That will include imagery from expensive high-flying drones like the 47-foot Northrop Grumman Global Hawk and from novel sources like ProxDynamics’ tiny PD 100 Black Hornet , used for years by British Special Forces in Afghanistan. More and more of it will come from body-worn sensors and intelligence-gathering equipment mounted on trucks and at bases.
“Eighty percent of our portfolio is geared toward the air. This is also where our spending and investment has been,” Atkins said, referring to manned and unmanned aircraft built for intelligence, surveillance and reconnaissance, or ISR. “We need to reduce our reliance on airborne platforms. As a result… we’ll be putting our efforts into ways to expand ground-based and maritime-based ISR. That’s not only to gain dominance in those domains but also to buy down the dependendency on airborne, which is the most costly,” he said.
Ground-based data and intelligence poses different problems if you’re an analyst trying to make sense of data than does drone-based ISR. “When you’re up 15,000 feet, your signal environment is dense. When you are man-portable and in a street, there’s a variety of devices to contend with, so you don’t need this wide band. You have a narrow band so your ability to sense and understand is a lot more localized. The trick is to get it back to a node where they can take advantage of it. A lot of it is stuff you’ll process audibly…Making the kit as user-friendly for that guy is important, but also getting it back to [Joint Operations Center] or a [Combined Air Operations Center] and then nest it into the bigger picture.”
Meanwhile, U.S. Special Operations Command is trying to find ways to do data processing, exploitation, and dissemination, or PED, with fewer people, Atkins said. “PED troubles us the most,” he said. “It’s the most human- intensive.” He said U.S. Special Operations Command is interested in “investments to buy down that manpower burden.” The goal is to do what now requires 500 analysts in a dark room with just one or two in a forward base.
Among the particular challenges of data for SOF is its variety. “It’s pretty much anything,” Atkins said The commercial world is full of big data analysis tools that wring insight data in large volumes or at great speed. But data researchers will tell you that big data comes in three flavors, volume, velocity, and variety.
“We’re the variety folks,” Atkins said. “We gather strange things off targets, pocket litter, yearbook pictures. It’s like, how [does SOF] make sense of that? How do you enrich it? How do you tie it to a geospatial location? So that’s really the challenge…I go out to Silicon Valley and stump that all the time.”
That variety problem creates a very particular need for machine learning and artificial intelligence. Virtually every enterprise that operates on a global scale uses AI or machine learning for something, whether to optimize product delivery (Amazon) or create better online interactions (Facebook). Investors buy satellite imagery from DigitalGlobe, then uses complex artificial intelligence to count the cars in parking lots on Christmas Eve, identify their owners’ tendencies based on make and model, and deduce the effects on future earnings statements for companies like Home Depot.
But the market doesn’t care if it’s off by a car or two. Navy SEALs need their AI to be far more precise. “When we’re trying to personally ID an individual, there is zero margin for error... when we’re trying to count kids that might be in a compound before we go assault it, that’s something that you have to default back to—not just one human but five humans because of the margin for error.”
While AI has become a hot field in Silicon Valley, there’s just no commercial outfit that’s meeting what special operations needs. Atkins is seeking to import more useful data from the field, use AI to turn it into intel actionable by policy folks and the soldiers in mid-mission, do it all at low data and energy rates and—perhaps most importantly—in a way that allows special operators to immediately share it with the international partners working alongside them. All of that is in the near future.
The United States is leaning heavily on the special operations community to serve as the nation’s on-the-ground response to the Islamic State. Some missions, like the recent raid that resulted in the death of a top ISIS commander and the capture of a treasure trove of important data, will look perfectly executed. Others, such as hostage extractions, offer much higher levels of complexity. The difference between a successful operation and one that fails, producing headlines and tears, such as the botched special operations rescue attempt for U.S. hostage Kayla Mueller that took place in February, is often a matter of intelligence.
When asked about how to create intel to save hostages during such attempts, in situations where the United States is not willing to commit human assets to ensure mission success, Atkins acknowledged that there was no technological answer. “I’ll be honest,” he said, “we don’t have that solution.”
NEXT STORY: Building An Unkillable Robot