AI and Data Analytics for Mission Support on the Border - Part 1

In this first episode of a two part series, learn how CBP is deploying Artificial Intelligence to support the border security mission and how industry is collaborating to create new solutions and respond to new technology requirements.

Embark on a journey into the high-tech frontier of border security as we discuss the transformative power of AI.  Venture further into the digital realm where cybersecurity stakes are at an all-time high, as we explore the necessity of zero trust architecture with our distinguished panel. 

Dive into the discussion about safeguarding our networks with robust encryption, even in the face of adversarial machine learning. This episode brings to light how government agencies and industry leaders are teaming up to reinforce our digital defenses and ensure that our security keeps pace with evolving threats.

Featuring:
Ed Mays, at CBP's Office of Information and Technology,
David Aguilar, former CBP Commissioner 
Joshua Gough, from CBP’s Planning, Analysis and Requirements office, and 
Jay Meil, Chief Data Scientist at SAIC’s Artificial Intelligence Innovation Factory

This discussion took place at the Annual HSDF Border Security Symposium in Washington DC on December 12th, 2023.

Follow HSDF THE PODCAST and never miss latest insider talk on government technology, innovation, and security. Visit the HSDF YouTube channel to view hours of insightful policy discussion. For more information about the Homeland Security & Defense Forum (HSDF), visit hsdf.org.

Transcript

Announcer: 0:03

Welcome to HSDF the podcast, a collection of policy discussions on government technology and Homeland Security brought to you by the Homeland Security and Defense Forum. In this first episode of a two-part series, learn how CBP is deploying artificial intelligence to support the border security mission and how industry is collaborating to create new solutions and respond to new technology requirements. Featuring Ed Maze at CBP's Office of Information and Technology, david Aguilar, former CBP Commissioner, joshua Gow from CBP's Planning Analysis and Requirements Office, and Jay Meal, chief Data Scientist at SAIC's Artificial Intelligence Innovation Factory. This discussion took place at the annual HSDF Border Security Symposium in Washington DC on December 12, 2023.

David Aguilar: 0:58

The way I see this panel, it's the who, the requires, how it's done, and the area of AI, ml, the requirements guru here and, by the way, I'm going to push this because he's a Border Patrol agent but he is actually doing what he does for the entirety of CBP, some of which was just covered a few minutes ago, includes the requirements analysis that he does in his area and, of course, jay from industry, representing all of you as to how that might be done in order to incorporate into the operations.

David Aguilar: 1:35

So we had a little prep session a day before yesterday and there was almost way too much stuff to talk about. But let's begin with one of the things that I think is critically important here. It is what is known as the accuracy and validity potential of AI and ML. We're in the beginning stages, but it's moving very rapidly. It's evolving very rapidly, giving us a tremendous amount of capabilities and therefore a very powerful tool. So let me begin with the DAC here. What do you see as some of the challenges coming forth on the concern of accuracy and reliability as we move forward on AI and ML implementation?

Ed Mays: 2:18

So great question. I think the real challenge is going to be with data. How do we collect it, what does it mean? How does it fit? One of the things that we always have a challenge with is how do we link data, how do we ensure that it's reliable, how do we ensure that it's accurate? And you'll see that in terms of what CVP is doing, we've hired a chief data officer to start helping us get our arms and minds around the challenge of data. Where is that data processed? Is it processed on the edge? Is that data moved and processed somewhere in a back-end system? And making sure that any changes to that data are actually changed under that change exists, under some governance, because it's really critical. Without accuracy and data integrity, all things fall apart long-term.

David Aguilar: 3:15

So governance, integrity and utility, bringing value to the operator? Yes, sir, absolutely. And those are all great points. Now to you, josh. We always think of ourselves as an organization that is going to use the capabilities, and that's absolutely correct. But the bigger the pool of players, the more powerful especially AI and ML is going to become. So that to me translates into interagency international, binational, tri-national, multinational, some of those challenges because of the differences in privacy, interest sensitivities come into play. How can we start taking a look at that, moving forward? Thank you.

Joshua Gough: 4:06

I think when you get as many players like that involved with something like this, with all the I don't want to say competing interests, because I think everybody has the same general interest in solving this, I think it gets down to being able to collaborate and use that term and you hear that term a lot. But I think there's something that everybody is now starting to functionally understand is that there's a bit of a collaboration tax that exists on you when you do something like that and collaboration is actually slow and expensive. So being able to do it and be able to facilitate those things right is an art and it takes a kind of a specific person to do it. But I was listening to some of the other speakers that were talking before this and one of the things they mentioned were the requirements. We talk about requirements operational, functional, technical, be able to document those things or going off the objectives.

Joshua Gough: 4:55

But one of the other speakers had mentioned this discussion of a standard and one of the things that we've started to experiment with in CBP is using an actual ISO standard 21838 for anybody. That's ISO standard kind of nerdish out here. It's basic, formal ontology, it's applied ontology and essentially what it does is it applies context to the machines. So lexicon people say you need a lexicon. Lexicons are for human beings. That helps us understand words and definitions. Lexicons do not help the computers. Using something like that ISO standard to define the words that we care about is going to help the machine understand what it is we're talking about, helps translate our reality as human beings into the machine. Using a standard like that is one method I'm not saying it's the method, but that's one method for achieving this interagency, collaboration. If we can get everybody to agree on what a standard way to go about doing this is, it makes the collaboration that much faster, that much easier and I think that much more efficient and cost effective.

David Aguilar: 6:02

Jay standards. What part does industry play, should play or can play? Earlier we talked about understanding the requirements of the operators. So what's industry's part in standardizing to the degree possible?

Jay Meil: 6:21

So I think it's earlier Josh and I were talking, and it's really about understanding the operator requirements, Understanding what's your mission, because they differ from division to division and that's for CBP or any of the services. There's going to be different challenges and different problems. There's some underlying similarities in just how we move data, how we tag data, how we define data, but very specific nuances to each particular division. It's important that industry understands those needs and doesn't come try a one size fit all capability. I want to go back to the data standards for a minute, because I think there's two ways to handle this.

Jay Meil: 7:00

This is something that across all of our services DOD, the IC, DHS everyone's struggling with. How do we get a common taxonomy or ontology? This goes back to also what Dr Mays was saying. If we're collecting on the edge and we're dealing with heterogeneous data and it's all disparate and there's multiple sensors and multiple systems, how do we make sure that the computers understand that if three different sensors in a particular area are picking up something that is standardized so that you can come back and the machines understand that those are all interrelated events? I think part of industry is understanding the mission. The other part is working collaboratively across academia, the government and the operations to come up with those standards, to then inform the machines on how to make better decisions.

David Aguilar: 7:53

Informing machines. I just got a picture in my head of a terminator Inform the machines, informing the machines, the variances of machines and capabilities, scalability that we have now because of legacy systems and because of the evolution of IT systems, also machines and so forth. How do we look at integration with existing systems, legacy systems and especially evolving systems? That last week we were we can have to go. We talked about more is law and long is law which is reduced time frames, I t evolution from eighteen months, literally four to five, six months at a time. So, legacy, existing and evolving it, capability, the integration process.

Ed Mays: 8:45

Well, I think that's a huge challenge, first of all, and it in. You know, if it was easy, everyone would do it right and they would do it well. And obviously Look around, specially across, you know, all services, the whole government is a challenge. But when I will tell you and I like to just take a second look back at what jay mentioned earlier was that whole machine to machine sort of intercommunication. We have the challenges now.

Ed Mays: 9:13

Right, we actually have, you know, non human inter. You know, interchanges with machines and we use certificates right, and you'll see, that's a. It's one of the key building blocks of zero trust. So those sorts of integrations are hard. And I will also tell you one of the things that, with respect to that is you know, I'm also kind of, you know, initiated a lot of cloud work inside of cbp.

Ed Mays: 9:42

When people hear cloud and they hear this sort of work, they hear magic. Right, I tell everyone it's solid engineering, it's solid work and it's one of those things where it takes a lot of creativity in collaboration. So those are the things across systems that have to happen. Everyone all understands, you know, you gotta have a common architecture, you gotta have a common understanding of data, but you also. You know, no matter what happens with the I n thanks. You also have to have really good people that understand what they're supporting, how they're supporting it and why delivery is important. I think those are the key takeaways for me knowing what you're doing, knowing that in state is and then building that roadmap to get there.

Joshua Gough: 10:33

I think, I think I think that mace is right on any solution set that industry comes comes up with. One of the things that I'm generally interested as the requirements person is how does this trace back to the problem set of the challenge that exists in the field? And one of the things I really want to just sort of push out there right now is if you have a solution I, this dd like a I that exists out there, whatever that means also think magic. Yeah, yeah, that's right, it's, what is it? Technology is different from it. But what I would encourage folks in industry to do Is to not come at the organization with a solution and tell me it's gonna solve world hunger, but comment, comment with the solution in your back pocket. Talk to me about problem sets, talk to me about the challenges that the operators face, because they're different between the ports of entry. They're different at the ports of entry, they're different for the pilots, they're different in the trade space, but but CBP is always willing to talk about this. The problem that ends up happening, just because we've been conditioned this way, is that the solutions come at us and we have to make decisions as of what problems they fix, and sometimes that's square peg round hole and I found some people with some big sledgehambers that are willing to put it through that round hole. But that's, that's not the easy way to do it, it's not the efficient way to do it, it's not the cost-effective way to do it.

Joshua Gough: 11:47

Having industry understand what it is we face as operators in the field but not not even just the operators. There's a mission support side of this whole thing. There's budget finance. There's there's the acquisition process, the requirements process. There's a big strategic planning process that backs up the things that we buy so we can show the American taxpayers you're getting the level of security that you're paying for and it works. And here's how we know all of that. All of that can be backed up with done by these things. But but again, that's just the big point. Just piggybacking on what Dagmiz was saying Please talk to the organization about problems and challenges. I encourage you to do that because that the demand signal that will come out of us is going to be truthful, because it's going to be beneficial to the guys and gals that are out there in the field.

Jay Meil: 12:29

And I'll just so I'll add that this, this whole, it's real big problem, right, it's, it's a challenge. You know, legacy, current and future architectures, and I think that we need to start moving towards Modular, open systems, architectures, open standards. You know, in the past we built Technologies or we built applications, which was great. They do really highly functional, high fidelity things. But the problem is, you had, you know, you came at it with I need to build a software that does a certain thing, and that software requires a database where data is in a Certain format and that requires some type of a service layer and some piece of hardware and then a place that someone can come and interact with. And we see these systems all over, but they don't communicate with each other. So that's the legacy side and and sometimes the current side. So it's, how do we sort of turn that on its head, flip that paradigm and, instead of focusing on I call them single applications or single software stacks, how do we move towards a modular open system or a single software experience? So what do I mean by that? The operators themselves see it as a single experience and they're interacting with the machines. It doesn't feel like they're jumping from system to system or screen to screen, but instead what's behind that, what powers?

Jay Meil: 13:44

That is essentially a common data layer Right, where all of the heterogeneous data can move into as necessary you know, promoting fabrics and meshes depending on how you want to architect it a common analytic framework that does the transformations, it does the, the normalization.

Jay Meil: 14:01

So another way to look at the ontologies, if you jump back for a second, is it's very hard to get everybody the same standard. You know to call something, a particular something right, or define it, and so you could use a top-down approach where machine learning Can actually do that normalization for you. But if we can get to a point where we have that common data layer, we have a common analytic framework, and then we can use some type of you know, use that that analytic framework like a chassis and now I can put lots of applications in micros, microservices, across it. So the operator themselves sees this integrated system, but they're actually accessing different things and by separating the software from the data, which is the big key here Now we get extensibility. We can talk about the future. I can pull old applications out and bolt new applications in without breaking all of the pipelines beneath it. So you get this, this interoperability.

David Aguilar: 14:52

So something that kind of ties into what you just said, jay. Earlier we had a couple of panels that talked about Denied areas, contested areas, degraded areas of communication. So some of the things that we hear about are, for example, secure private networks operating independently out there in big bends, for example, and then the data center up here in Ashburn, quickly moving to the cloud, integrating all three of those, because that's what it would take for an operator to be operating in the big bend of the boot hill or northern border somewhere To to to have the capability to, by way of a secure private network, have the capability to to operate and Utilize a IR machine learning, independent of the cloud, independent of the data center. But when required backhaul, what needs to be backhauled in order to get the, the universal benefit how is industry going to address those?

Jay Meil: 15:53

So I think that it's very good question I think industry needs to start thinking about. We're all doing big machine learning right. We're starting to really deal with heavy weight models that do things like facial comparison or large language models that we always talk about, or some type of object tracking To a degree because storage and compute are commoditized in the Cloud or because we can have these big data centers. It seems like actually running inference on these are easy, but the problem is that the edge we have to assume that we're going to have no connectivity or very little connectivity. We're going to have a thin line essentially. So what we need to start thinking about in industry is how do we take these really high fidelity models and how do we take them to the edge rather than bringing the data back? So that means having to start thinking about things like size, weight and power constraints. How do we do lightweight deployment of these models? It doesn't take a lot of computing power, so that what we're transferring back is very small amounts of data.

Jay Meil: 16:52

So the way I look at it because it's easy to visualize is if you have a drone that is doing video coverage of an area, even in a perfect network environment, you don't want to be backhauling the amount of data for a live video feed to go get it processed inside an operation center.

Jay Meil: 17:10

What, instead, you should be able to do is, on that platform or on sensors near that platform or on a forward operating area near that platform, you should be able to run the machine learning inference and say, in this five second period or in these three frames or whatever it is, I've identified what you're looking for and send only that information back. And if you think about having to do that in a large scale, over hundreds of miles or different areas of operation, then you have to start thinking because again, if you're on a denied or degraded network about prioritization and criticality, so that's something else machine learning can do so not only identify a particular object, but you might have 100 objects being identified in multiple places. How do you prioritize what's most important for backhauling those small messages? I think industry really needs to look at these edge architectures and look at how we not only do the inference at the edge but also how we get the right information at the right time back to the operators.

Ed Mays: 18:06

So, if I could interject, I like to say that we need to have better conversations with industry.

Ed Mays: 18:15

In our CTO office, with our CTO, sunil Madhugari, and some of his team that are here in the room last year we did something that we called the Douglas Analytics Project.

Ed Mays: 18:25

Now, it was kind of nation at best, but it taught us a whole lot and I think we've got an edge computing paper that's coming out or that's already out and a data paper. But a lot of that we learned along the way and we would have loved to have a little more advice. I mean, we actually did looking at sensors, processing at the lowest, at nearest to that where the data was collected, all the way to pushing data back to a big back end, major CSP and using their AI tools. But we definitely need help, and I look around the room and I think that's one of those things we'd love to hear from you about your ideas of how we could do it better. Sometimes we come at things I know I do from an academic perspective but there's lessons learned that industry has garnered from all your interactions with all your customer base from around the country and internationally. We'd like to leverage some of that to make ourselves better, quicker.

Joshua Gough: 19:28

Chief, I just want to just provide some operational context for this. When you have an agent in the field or an officer at the port or a pilot in the air, they're going through a very quick cycle. So this is just as you develop these solutions or you think about these solutions, think about the cycle that they go through. The initial part of this is they detect a threat. They either see it or something alerts them that something is there. The next step is they identify it by identifying. They say is it a car? Is it a person? Is it a bird? What is the thing that's out there?

Joshua Gough: 20:00

And then there's a classification aspect of this thing how threatening is this item of interest? Then it gets into what we call a track mode. Track is nothing more than redetection, over and over and over and over and over again. And when an agent deploys to a detected threat, he's cycling through that constantly. He's wanting to get updates on, and these things start to update themselves and they start to get a little more fidelity and it makes determinations as to how he or she responds, because going after a 58-year-old grandmother with a two-year-old child in their arm is different than seven backpackers with AK-47s strapped to them.

Joshua Gough: 20:40

You're gonna change the way you do business based on what it is you've identified and how you've classified those threats. So when you think about edge computing, let's not forget the machine reasoning side of this thing. If I can turn this into my partner instead of simply a tool Some way, somehow, if I can get this to partner with me no different than another human or my dog or my horse, that's that's gonna win it for these guys and gals out in the field. That's that's really wanna. Just just putting that, putting out there so you get an understanding of what agents and officers and pilots, even folks in trade, even you know an IPR violation is still a detection, identification, classification and attract where's this thing in space and time and knowing where that thing is up to the minute, being being, say, updated a thousand times a second. That's what's gonna make us that much better. This idea of what I refer to as human machine teaming. I'm not building skydive, chief, I promise you I think that's great.

Jay Meil: 21:37

I so human machine team, is a huge Passion with what we're trying to do and I think you're absolutely right. That contextual reasoning piece is so helpful to help the operator because they might be dealing with not just one thing but multiple things and they need to understand again how to prioritize what they're going to do, how they're going to spread the limited resources across to complete the mission. And I really think what it does is twofold. I think with the machines do, if we can build this utopia, human machine teaming it's not about replacing operators or analysts or anyone. It's about augmenting right their workflows, so it's about reducing their cognitive load and what they're having to take in and do the reasoning on themselves. And it's also about speed decision right as well as force multiplication, because one operator can do significantly more. When a machine is doing what machines do best, which is looking for anomalies, following tracks, finding a signal in the noise, whatever it might be, they tip in queue and now the operator can act accordingly.

Joshua Gough: 22:36

And we certainly don't want the machines making decisions for us.

Jay Meil: 22:38

Don't want them making decisions, but I would I would, I would venture the I would.

Joshua Gough: 22:42

I would just say this, chief, like when we talk about a partnership between humans and and making this, you know, partner, I think what I the simplest way I've found to put it in somebody. Somebody is going to correct me this afternoon, I guarantee it. But humans have an order loop. You observe, your and you decide and you act. When we talk about machine teaming, I don't want my machines deciding things for me. I want them to have a modified order loop where they observe, orient, trigger and then act to tell me something that I need to know, that I wouldn't have known before.

Announcer: 23:12

Thank you for tuning in. You can follow hsdf the podcast on any major podcast platform. Visit hsdforg to learn more about the Homeland Security and Defense Forum.