Poisoned data sets, trojan horses, and ever-changing cyber threats could derail the Army’s plan to broadly adopt AI. But a new 100-day plan aims to root out obstacles and prepare the service to scale third-party models and algorithms.
The plan, released in March by the Army’s assistant secretary for acquisitions, logistics and technology, comes after the Army released a software directive that aims to help the service—and industry partners—deliver capabilities faster by mimicking how commercial developers work.
“The Army recognizes that we’re not going to be doing algorithms and model development and training better than industry; we want industry to do that. And they do that extremely well. And so we want to adopt a lot of that. And one of the obstacles for the adoption is: how do we look at risk around AI?” Young Bang, the Army’s principal deputy assistant secretary for acquisition, told reporters Monday. “What are the issues around poisoned datasets, adversarial attacks, Trojans…. And it’s easier to do if you have developed it in a controlled, trusted environment that, say, the DOD or the Army owns.”
Those efforts will eventually dovetail with the Army’s ongoing work in trying to simplify how it uses AI, particularly for intelligence systems, through a program called Project Linchpin.
“There are a lot of programs right now that are working on models, and training and deploying and testing those out,” on their own as the Army develops the Linchpin environment, Bang said.
Linchpin is designed to be a digital environment with vetted tools, infrastructure, standards, potential use cases and associated data for AI, said Bharat Patel, the lead for the Army’s Project Linchpin and sensor AI program. And having all of that in one place should make it easier for project managers to integrate artificial intelligence into programs that could benefit from it.
Right now, Linchpin is focused on defining those use cases and preparing the data associated with that, said Patel, who is also a product lead for the program executive office for intelligence, electronic warfare, and sensors.
“We are doing some basic model training. But it’s not for performance or anything. We are trying to figure out what is our process, what is our governance, what [are] our standards that will allow us to do this a lot faster. So our purpose is really to kind of learn and to make sure that we are ready to run once the 500-day implementation plan and contracts and all that is in place,” Patel said.
And as Linchpin matures, the plan is to feed it algorithms from other programs and third-parties so program managers can choose and build on capabilities from there. Bang said details on Linchpin’s readiness will come after the 100-day plan, which is likely to be completed this summer—with a 500-day plan to follow.
“Each theater is different,” Patel said. “You can’t think a model for [European Command] is going to work out of the box for [Indo-Pacific Command]. The trees are different, the biosphere is different…that’s why it’s super important to get after the use case. And where that [area of responsibility] is specifically at. So we are looking at that very closely.”
Comments are closed.