Military and civilian tactical communications programs widely acknowledge the critical need to improve mobility for warfighters and operators in all domains. This includes making network connectivity available anywhere in the world, on every platform, and on the move – while increasing communications speeds to make video, big data, and AI-enabled decision making possible.
To meet these needs, defense communications programs are increasingly turning to wireless infrastructure – mirroring the proliferation of mobile devices for consumer use – that enable a vast array of use cases for mobile command posts, warfighter-worn communications and IoT devices, en-route situational awareness, and more.
A new class of rugged, small wireless network systems is bringing the benefits of secure meshing Wi-Fi to warfighters for both unclassified and classified networks, allowing warfighters to use commercial mobile devices, in dynamic environments.
Join Charlie Kawasaki, CTO for PacStar and Andrew Puryear, Group CTO for Ultra Electronics in an in-depth exploration of requirements, best practices, and state-of-the-art for tactical secure Wi-Fi – providing specific examples of how defense programs can meet these requirements using cost-effective commercial, low SWaP technologies.
Transcript
Bob Ackerman
Good day, everyone. I'm Bob Ackerman, the Editor-in-Chief of SIGNAL Magazine and I'd like to welcome you to our SIGNAL Media webinar titled How to Maximize Mobility at the Tactical Edge with Secure Wireless. In today's webinar, which is sponsored by PacStar and Ultra, two experts will discuss how cost-effective commercial technologies can deliver tactical secure wi-fi to war fighters in dynamic environments. Now, those of you who have participated in past SIGNAL webinars are familiar with our audience interface. But just to remind you and for those who are new, throughout the presentation, webinar attendees are encouraged to submit questions electronically through the ask a question box on the webinar console.
When our experts are finished with their presentations they'll answer as many questions as time permits during this hour-long session. Today's first presenter is Charlie Kawasaki, CTO for PacStar. In his 16 years with the company Charlie has been part of a team that won tactical networking equipment awards for several defense department programs across the services. All told, he brings more than 40 years experience to the table in a number of networking, cyber security and systems integration disciplines. Joining Charlie as a presenter is Andrew Puryear, Group Chief Technology Officer, Ultra Electronics, naval reservist who is a founding member of a navy cyber warfare group specialized attachment. Andrew focuses on innovating emergent, disruptive and convergent technologies. He also advises the commerce department on emerging technologies with dual use applications at home and abroad. So with that, I would like to turn the presentation over to Charlie Kawasaki. Charlie.
Charlie Kawasaki
Thanks Bob. In order to achieve and maintain warfighting overmatch, coordinate deployed forces and enable new capabilities, the US Army, Air Force and Navy are actively looking to new programs such as Joint All-Domain Command and Control or JADC2 to ensure warfighters have maximum situational awareness. These programs will adopt a variety of compute and bandwidth intensive technologies increasing the use of sensor data, big data, analytics, artificial intelligence and machine learning and video to deliver this situational awareness and command and control information that warfighters need.
While the DoD intends to rely heavily on information resources, tactical and expeditionary and networking command post programs widely acknowledge the critical need to improve mobility. This necessitates the implementation of higher capacity secure wireless farther out at the edge of the network and across multiple data transport types. Additionally, our warfighters doctrine increasingly includes coordination with coalition partners, which creates additional requirements for information sharing across nation boundaries in dynamic environments. To enable mobility, wireless and partner interoperability, the National Security Agency or NSA, established a program called Commercial Solutions for Classified, or CSfC. This NSA program requires dual layers of commercial encryption configured correctly to transmit classified data. This program enables a variety of use cases including the ability to use commercial wireless mobile devices for classified communications over wi-fi and the ability to use commercial encryption devices for site-to-site transmission of classified information without the use of Type 1 classified cryptographic equipment.
This capability has the potential to enable a vast array of use cases. In the same way consumers have been able to use mobile devices for literally millions of applications we have just scratched the surface for how mobile devices can be used in defense and secure use cases. Here are just a few examples we've seen of CSfC use cases including for things like command post mobility and vehicle mounted command posts, pilot electronic flight bags, dismounted situational awareness applications, air and route roll-on, roll-off applications ship-to-shore communications, offload of condition-based maintenance information, flight line and depot wi-fi applications, links for UAS and for coalition interoperability.
To make this more concrete I'm going to cover several programs that have rolled out CSfC wireless or are piloting it. On this slide you can see a quote from US Army PM tactical network secure wireless detailing their success rolling out CSfC for wi-fi enabled command posts. Their aim was to reduce setup time, reduce the duration that they have no situational awareness and improve command post mobility. By reducing the time it takes to set up and tear down a command post it opens up new options for maneuver as opposed to the past where commanders had little choice but to defend a static position.
Here are a few photographs of that system being set up and used. It really was as simple as unpacking access points, hanging them on tents and turning on wireless infrastructure and getting to work. The wireless command post photos I just showed are using variants of PacStar technology in our secure wireless command post family of products for classified SIPR and CUI NIPRNet data based on the NSA campus WLAN architecture or capability package. In this simplified diagram the network and wi-fi infrastructure includes a red inner VPN gateway. This provides the first layer of VPN encryption as required by the CSfC program office. The red encryption network also requires certificate authorities, intrusion detection systems, firewalls, hardware security modules and continuous monitoring systems to support it.
This configuration can be duplicated once for each network such as NIPRNet, SIPRNet and the mission partner environment or MPE which is also secret. The system then uses an outer or gray network which establishes a second outside layer of encryption using WPA2 EAP-TLS. It also needs supporting CAs and cyber security infrastructure like the red networks. The gray network can be shared between multiple internetworks but is not shown here for clarity. Once the data is encrypted with WPA2 it's considered black and is allowed to be transmitted over the wi-fi radios. When this transmission arrives at the mobile device, it's decrypted with matching equipment. First the wi-fi client software, then a VPN client, not unlike how most enterprise users today have mobile devices set up. Once the data is decrypted twice it can be used by any typical application.
Based on those early successes, the PEOC3T secure wireless program is now proliferating these kinds of solutions to serve numerous other users including ESBs or the Expeditionary Signal Battalion-Enhanced, brigade combat teams and soon mobile command posts. That is driven by the US Army belief that future CONOPS require a substantial increase in mobility and flexibility, leading army to invest in pilots that move the C2 infrastructure onto vehicles, moving away from giant tent-based cities shown in the upper left that take too long to set up and move, that have two large a signature and that limit maneuver options. Instead they intend to move C2 networks onto a wide variety of vehicle platforms such as shown in the lower right, including transport vehicles, armored vehicles such as the AMPV and Stryker and light vehicles such as JLTV.
This means wireless for C2 networks must increasingly move onto all vehicle types. In this concept drawing from US Army's CPI2 program, you can see a CSfC secure wireless hub providing the management and cyber security for an entire meshing network, not unlike the network infrastructure in the previous temp based diagram. It's paired with vehicles transporting server farms for mission command applications. This architecture is being built and tested now as a proof of concept. In this second diagram you can see a secure wireless meshing remote endpoint or SWMR remote vehicle receiving that dual encrypted CSfC transmission over meshing radios. This is the remote vehicle that would be paired with our main command post vehicle. In the case of both vehicles, wi-fi access points can be integrated to provide local wi-fi coverage for both classified and CUI or unclass networks.
In one last use case example I'm showing vehicles from the Marine Corps network on the move, or NOTM program, undergoing full system assessment in this photograph. These are mobile communications vehicles providing network services on the move with secure wireless meshing technology on board. Because they're lightweight vehicles, reduced Size, Weight and Power is critical and you'll see in subsequent slides they benefit from PacStar solutions. Though dual-layered encryption is straightforward in concept, full CSfC implementations must include a breadth of technologies or components including public key infrastructure, encryption gateways and clients, authentication systems, cybersecurity technologies and secure network infrastructure. These components must first have their cryptographic implementations validated to FIPS 140 standards through the crypto module validation program. Next, their implementations for their protocols must be common criteria validated against one or more protection profiles such as IPSEC VPN gateways, TLS protected applications or wi-fi access systems. Additionally, certified public key infrastructure is required for certificate or key management.
These certifications are typically sponsored by commercial vendors and conducted by recognized testing labs. Once that step is complete, the products such as VPN gateways are placed on the NSA's CSfC components list and are eligible for use in CSfC solutions. Organizations or integrators can select from the products on the components list to create solutions that conform to one of NSA's capability packages such as for multi-site transmission, for mobile access transmission or for wi-fi access. Once the integrated solution is assembled and configured and tested, they must be approved by NSA's CSfC PMO. So how does PacStar, now part of Curtiss-Wright Defense Solutions division, fit into CSfC programs?
Well, first we make three major types of products that apply. First is our small form factor rugged hardware, popular in many US and partner tactical networking programs. Second is IQ-Core Software, a single pane of glass management system for tactical networking equipment. This is deployed widely, particularly across large numbers of programs at US Army PEOC3T. Lastly are our PacStar integrated solutions, combining our hardware, IQ-Core Software and an extensive suite of third-party technologies into complete CSfC solutions. But before I dive more into detail about our solutions, I'm going to turn this over to Andrew of Ultra to talk about their technologies and future directions enabling secure wireless and networking for the warfighter. Andrew, your turn.
Andrew Puryear
Thank you Charlie. Hey, thank you, everyone for joining. Good morning, good afternoon. Next slide please. I wanted to start off with a quick introduction of Ultra before jumping in to talk about solving user needs with emergent and disruptive technology. Next slide please. So at Ultra our mission is to innovate today for a safer tomorrow and we all really believe in that mission at Ultra. Our focus is on solving customer and user problems in defense, security and critical detection and control markets.
We have over 4,500 employees primarily in the US, Canada, UK and Australia, although we have employees elsewhere across the globe as well. Next slide. Yes, next slide, thank you. We see the overarching theme of Ultra is delivering information advantage and that's the key theme of what I'm going to talk about today. In the maritime domain we build sensors and systems such as sonars, radars etc for anti-submarine warfare decision advantage. In intelligence and communication we build command, control and intelligence systems, tactical com, cyber and crypto systems and specialist RF systems to enable that information advantage. In forensics technology we provide law enforcement agencies from all over the world with tools that they need to prevent and solve crime. In the next slide I'll provide a little more detail on the intelligence and communication strategic business unit. Next slide please.
Intelligence and communication is actually laser focused on enabling and delivering that information advantage. Cyber and crypto communications have to be secure to deliver information advantage. Specialist RF access to the electromagnetic spectrum and electromagnetic dominance is essential to information advantage and so on. Next slide please. So I wanted to say also just a few words about our technology roadmap. This roadmap really represents the nexus of our users most critical challenges both now and in the future. We start with those users and then we look forward to our foundational technology that Ultra has and then we understand how we can add advanced emergent disruptive technology to go and innovate and solve for those needs.
So with the next slide I'm showing the technology roadmap. This is just one view of our technology roadmap - one that is primarily focused on our intelligence and communication strategic business unit but not exclusively. We build an array of technologies to enable the integrated cognitive battle space - battlefield.
These technologies include intelligent comms, network frameworks supported by ultra waveform and radios, cloud to edge crypto, next generation Type 1 high grade crypto with modern medium independent key management infrastructure, plug and play technology for rapid solution deployment and government-compliant cloud enabled application delivery and so on. So the next slide I will start off talking about cool technology. Charlie, if you can, please build this slide out. There's two additional clicks. So I want to start at the high level - artificial intelligence and machine learning, talk about what is the current state of the art and what are the game-changing advances in artificial intelligence/machine learning, then bring that down and actually apply that AI/ML to operational networks and then finally I'll wrap up with how we're actually deploying those AI/ML capabilities to Ultra's offerings. Next slide please.
So for decades, researchers have used gameplay as a way of measuring advances for artificial intelligence and machine learning applications. On the left there I'm showing AlphaStar which is a AI/ML instance that plays Starcraft, which has long been seen as somewhat of a grand challenge for artificial intelligence machine learning. For a long time people thought that AI/ML would not be able to beat the best players in the world at Starcraft. The reason for that is because Starcraft has, you know, three fairly challenging things for past generations of AI/ML. So the first challenge for artificial intelligence and machine learning that they had to address was imperfect information. When you play this game you don't know everything that's going on in the entire play space, so imperfect information is a real challenge historically. It's got very high dimensional state space, virtually infinite number of movements so there's no way to computationally compute what is the optimal solution at each stage. And third is you have to balance many competing objectives both long term and short term. When you combine all those things together it's a very challenging problem for past generations. AlphaStar as you're seeing here has actually beat the best players in the world.
On the right I'm showing AlphaZero. So I think people remember circa 1996, 1998 when Deep Blue beat Kasparov in in chess and things have actually progressed quite a bit since then in gameplay. 2005 was actually the last time that humans won against the top performing computer under normal chess conditions, so not since 2005 has a human won. There have been many draws but humans have not won. Now chess engines on your iPhone can beat the strongest players in the world which is pretty amazing. So AlphaGo, which is the previous generation of AlphaZero, learned to play by studying humans. It studied millions and millions of, well, thousands and thousands of games by humans. In contrast, AlphaZero learned to play exclusively from playing itself. It figures out everything by itself - in every step is a little creative leap.
It's like millions of mini discoveries one after another to build up its creative thinking and in fact AlphaZero learned to play three different games that I'm showing here. In that the graph on the right, ELO score is a measure of performance, how well a player plays and then training steps is the x-axis and you see by the time we get to about 200,000 training steps that we're beating the best players in the world with this particular AlphaZero instantiation. Google Duplex on the bottom - I never like to talk about artificial intelligence and machine learning without talking about human machine teaming - making sure that humans and machines are able to interact intelligently is critical to successfully deploying AI/ML.
And so Google Duplex is working on much more natural interactions with computers and so this is actually a very smooth interface where you're able to in this example Google Duplex is able to set up appointments for you but often times the users who say I'm getting an appointment to get a haircut, users aren't even able to tell that they're interacting with a computer for these cases. Next slide.
So talk a little bit about advances in creativity for the AI state of the art. What I'm showing here - all these examples are something called generative adversarial networks or GAN networks. And they're phenomenal in producing on the right there I'm showing progressive GAN. Those are synthetic celebrity faces. What was done is they trained a model based on many, many thousands of celebrity faces collected and then asked that the system to produce synthetic fake people basically and they look very realistic and this is actually, this image here is a few years old. They can actually do a much better job now. Feature wise transformations for CycleGAN on the bottom there. So you can choose to enhance features such the type of brush stroke or we would like this zebra to now look like a horse or summer to look like winter - very very powerful. Perhaps my favorite though is bottom left where we show you can describe an image that you would like generated and AI/ML will generate that image.
So a small yellow bird with a black crown and a short black pointed beak there on the left and then it was fed into the algorithms and then the picture above it was generated . Not perfect in all cases and you can see some artifacts but still very interesting. So how would we apply this to technical communications? Well, data is a new oil and access to data will oftentimes limit our ability to build these models that we need. With synthetic data generation, future generations of tactical networks will be able to bootstrap from a very small number of observations and then use that to generate bigger data sets which we can then use to train our artificial intelligence machine learning models. Next slide please.
So access to data is a key and recurring theme to data science and so to deploying artificial intelligence machine learning. One way to get around the need for massive amounts of data is something called transfer learning and that's what I'm showing there on IMPALA on the right. What that is, is I think those are, there's actually 57 different Atari games that IMPALA knows how to play. Not all of them were shown there but the way that transfer to learning works is you train the model on a specific game and then, and that takes an enormous amount of learning, a lot of time, right, and a lot of data but then you're able to take that model and teach it to play other games with far less data. And so this is a way to bootstrap and minimize the amount of in-theater data that we need to observe to hone our models to apply them to specific tactical environments.
So you can imagine training the model in a space where we have much access to data, right, you know say just outdoors and a CONUS environment right, but then you go into theater where it's slightly different. You can add just incremental amount of data to that to train your models to get where you need to be. So next slide please. Artificial intelligence revolution - why now? There have, artificial intelligence has been around since the 1950s as decision machines, decision trees. There have been a couple of valleys of despair - first in the 1960s, 70s and then again in the 1980s where the hype cycle built up and then the hype was not realized and so funding was lost, interest was lost. We've been sustaining this hype cycle for quite a while and I believe it will continue and the reason that artificial intelligence machine learning is finally starting to fulfill the hype that has been set out for is it's really a combination of three things. Massive compute, massive amounts of data - 14 million labeled images in ImageNet and algorithmic innovation. So algorithmic innovation includes things like the transfer learning that we just talked about. To really point out how much compute has increased, on the graph on the right there I'm showing the amount of compute cycles that's used to train a single model versus time.
And what you're seeing there is actually pretty incredible. So Moore's Law doubles every 18 months. What this is showing is it doubles every 3.5 months. So over the span of this graph, Moore's Law would have increased by a factor of 12 while this actually increased by a factor of 300,000 and so the compute applied to machine learning problems is incredible and has increased substantially. Next slide please.
So I've divided up artificial intelligence use cases into sort of two broad categories. Inference, information, knowledge generation and then on the next slide when we get to is autonomy, which is where you add decision making into the mix. So classification. So the image that I'm showing there is actually where an algorithm is able to identify malignant cancer much better than the best radiologist. Synthesis - this is back to creativity. This was actually the first AI-generated artwork that sold at Christie's for half a million dollars on auction. Anomaly detection - certainly something critical for deploying this in a tactical environment.
How do we know when there is something novel in the electromagnetic space that we need to react to, right? We've got to be able to identify and react to those anomalies. Prediction, optimization, data mining. In a subsequent slide I will tie these all back together into tactical and the tactical networking. Next slide please.
And so autonomy. Embedded expertise, larger scale operations. That can be used both for good and bad. You can certainly imagine larger scale operations being used to increase spearfishing attacks where an agent scrapes the web for everything that it can possibly find out about you. It finds out you've got children who play soccer, right, and so it crafts a spearfishing attack targeted specifically at an individual that says hey, your son's soccer match has been canceled, it needs to be rescheduled please click on this link to reschedule.
And so there's information in there that would lead you to believe that it's a valid email because it knows things about you. You click on it - now you're infected, right. So nation states APTs have been able to do this for a long time but they haven't been able to do it at scale so that's where the danger of AI/ML comes in. Superhuman precision and reliability. So the example I like to give here is X-47B - an experimental aircraft that was able to land on an aircraft carrier with precision that actually far exceeds what a human could ever achieve. So next slide.
And build it out for me Charlie. So next we're going to take it down and apply AI/ML to operational networks. Next slide. So traditionally communications and networks folks, myself included, have focused on measures of quality of service that include throughput latency etc. But to be honest that's not what users care about and even worse, those quality of service metrics aren't always even proxies for things that our users care about, right? Network utility management, data prioritization are similarly coarse and don't dynamically apply to mission objectives to evolving information needs or the commander's intent. So with mission-focused quality of service, what that is is we must ensure that critical information finds a path to the right user at the right time in congested, contested and highly dynamic electromagnetic conveyor environments using secure control of all available communication networking resources.
That's really the essence of mission-focused quality of service. Artificial intelligence is a key enabler for mission focus quality of service. This is an extremely high dimensional space with high uncertainty and the need to balance short and long term objectives exactly the same as the example that I gave up front with the Starcraft example. AI will enable joint management of both networks and information which is critical to getting to mission focus quality of service. Next, ease of use during the fog of war.
Too often we've seen manual static configuration of individual tactical networks associated with limited internet working capabilities. These networks take trained operators to set up a provision and they're often manual and prone to process error, right? The issues are compounded by the growing emphasis on all-domain warfare and the complexity of controlling these heterogeneous networks. AI can address these challenges by introducing the concepts of operationality, diversity, rapid adaptability to the orchestration of network of networks and then finally secure resilience. By jointly optimizing the information flows with all degrees of freedom to a com system, modulation type, constellation size, antenna pattern, network topology with an eye on that mission-focused quality of service, we can ensure that we're not just resilient but we're also secure.
And I'll get, I'll show some examples in the next slides. Next slide, Charlie. So this is an example of some research that we're doing at our resilient machine learning institute in Montreal, Canada. The challenge here, just like I articulated on the previous slide, is to deliver the most critical information to the right place at the right time in congested, contested electromagnetic environments. You can see there on the left, I've highlighted the artificial intelligence machine learning strengths that we're utilizing to actually execute this sort of overarching challenge, right? Prediction - how do we predict and optimize critical information. Optimization - how do we jointly optimize every layer, from the physical layer to the topology to the application layer to maximize that mission focus quality of service. Classification - how do we classify interferes and jammers for ECCM? Autonomy, synthesis - all these things are coming together to deploy this sort of critical capability. Next slide please.
So that this slide is all about turning data into information because it's not enough to just view a network as being optimized and in the absence of data priorities, right? So that the challenge here is, you know, users are swimming in sensors and drowning in data and the bandwidth just doesn't exist to get all that data out to all the users. So our solution is to have a distributed AI agent that predicts what information will be needed when and where and then dynamically push that data to the places where it needs to be so that you can get that profound insight or that actionable information and to pull the information to places where we can run analytics to be able to get those insights. So I've got a couple of slides. Next slide please, where we show single layers of this joint optimization. So on this chart here I'm showing adaptive network topology and contested environments. So that this is a very high fidelity simulation. We've simulated path loss, multipath, etc and the multi-layer learning algorithms optimize this network topology. Next chart please.
This year we're actually demonstrating automated network topology adaption where we're adapting the antenna patterns with LPI low probability of intercept, LPD low probability detect waveforms. Again, this is optimizing every single degree of freedom so that we can maximize the mission focused quality of service. Next slide please. Here we see a jammer - the red blob that is wandering around there trying to disrupt communications and the multi-layer learning algorithms are optimizing these for anti-jam capabilities. So it's both optimizing - you don't see it here but it's optimizing antenna patterns, topology, connectivity, with the intent of maximizing the mission focus quality service. Next slide please. In here we're showing a cognitive anti-jamming capability. So there will be a jammer that pops up in the center graph here. The artificial intelligence machine learning will attempt to learn the jamming pattern if there is any jamming pattern or if it's a random pattern that is repeated, it will learn that. I don't know if you've played tic-tac-toe on the New York Times web page that's backed by artificial intelligence machine learning algorithm but it starts off pretty good - you win about half of them but very quickly converges to where it beats you every single time. It has learned what you're going to do and that's the intent here is to try to intelligently learn what the jammer is going to do and to always be one step ahead of that jammer. Next slide please. And build it out, yeah. So we talked about artificial intelligence machine learning and we talked about how AI/ML can be applied to operational networks. In the next slide we will be applying that and we'll talk about how those algorithms are actually applied to current products.
So I'll wrap it up with one of Ultra's current offerings today, one of quite a few offerings in the mobile communication space. I see this platform as one instantiation, one place for rapid capability insertion of all those algorithms that we showed on previous charts. In its current form factor it's NIST certified level 2, DODIN APL certified so highly certified for government use. It's proven entrusted by the U.S. Navy and it's self-organizing self-healing and interoperable. Most importantly to me as a CTO is, you know how does this solve our customers problems today and how are we using it as a platform to solve their problems in the future as well. Next slide Charlie. So this is an example of the network topology for this particular system. So at the top you see the access wired network. So this is where you can plug it in and then from there it communicates out to the mesh network so that they're both mesh nodes and their access point nodes that then communicate to edge nodes. This is all fully self configurable and resilient so if a particular node drops out it reconfigures itself optimally to get the right data to the right place at the right time. Back to you Charlie.
Charlie Kawasaki
Thanks Andrew. So now I'm gonna talk a little bit about how PacStar can bring all these technologies together to create turnkey deployable tactical solutions. First, in terms of small form factor and rugged hardware, all else being equal communications equipment can never be too small, too light, or too power efficient. In contrast to legacy data center style 19-inch rack mount equipment, new generations of equipment designed for tactical and expeditionary use are becoming available with enterprise grade networking and security technologies on board. Additionally, with network function virtualization, many required technologies can be co-located on a single server platform such as the tactical one you see here. For example, new tactical equipment as compared to legacy 19-inch rackmount equipment on average is around 10 times lighter than a typical 1RU server, 12 times smaller and consumes 18 times less power. So we have a complete family of small form factor hardware that's great to deploy a lot of the technologies that Andrew just talked about.
First, on the left-hand side you'll see that our family includes routing and switching modules with embedded Cisco technology including our new top of the line 10 Gig switch. And we have a broad array of server types based on Intel Xeon processors with up to 120 Gigabytes of RAM, 16 CPU cores and up to 122 Terabytes of storage on our storage server. We also have NVIDIA GPU-enabled servers that can do a lot of the kinds of things that Andrew is talking about supporting AI and also video processing applications at the tactical edge. We also have a line of radio gateways for radio for IP, radio relay, radio crossbanding and voice management. And you'll see much more in a minute. I'm going to show you the PacStar 464 Wi-Fi AP that we've developed in collaboration with Ultra. And we have a variety of adapter sleds for things like Type 1 encrypters, hardware security modules and tactical radios - handy when we can't physically modify those devices but we need to integrate them into our systems.
So here's a little bit of a closer look at the PacStar 464 Secure Wireless Access Point. And what this does is it takes the certified and proven Ultra Electronics wi-fi protect technology and it packages it into a PacStar 400-Series module that's compatible with the rest of our product family and I'll show you how we can combine that to create full solutions. So what's different about this solution is that our product does things like support snap-on radio batteries from standard tactical batteries will slide right into our chassis providing UPS. It supports wide range AC and DC and it has gone through extensive environmental qualification with a minus 20 to 70c environmental operating temperature as well as a wide battery of MIL-STD-810/461 EMI certification by independent labs. And this thing only weighs about two and a half pounds. It's about the size of a paperback book.
I'm pleased to share with you that as of this morning the NSA NIAP program has added this to their certification list and you can find this online and it's now eligible to be used as a wi-fi access component in CSfC solutions. So here's just a little bit closer look at the front panel of this device and you can see we've got the six antenna connectors for the two built-in radios and then we have some indicators, we also have a lights out switch as well so if you want to run this in dark mode and some WAN connectivity as well.
So what we do is we typically take our modules and we connect them together and we have a variety of packaging technologies that allow us to create backpack solutions or we can mount them into our smart chassis that have built-in UPS, mount those on vehicles. We also have 19-inch rack mount solutions that can mount our modules directly into standard 19-inch rack mounts and we have lots of different transportable transit case technologies that we can use for deployable applications. And we typically work with partners then to combine these with further accessories and create full programmatic solutions for large command posts as you can see here, plenty of vehicle mount applications and resulting in complete program solutions.
So as part of our CSfC solutions we integrate IQ-Core software which I introduced previously to enable system management. We include what we call IQ-Core Network Communications Manager or NCM. This provides a comprehensive suite of nodal management interoperating with a wide variety of heterogeneous networking equipment popular in the DoD. It's also an on-platform solution and it provides setup configuration and management tools.
We also offer IQ-Core Remote Operations and Management, or ROAM, which enables remote and distributed network management even in disconnected intermittent and limited environments, which is critical for our army customers. And importantly for today's discussion, we also offer what's called IQ-Core Crypto Manager, or CM, which automates PKI, VPN and device management functions for CSfC solutions. We've combined all these together to create an award-winning family of CSfC and CUI wireless solutions for wi-fi access, mobile device access and multi-site access including with multiple vehicle and meshing solutions and we have numerous field-deployed solutions today. So that concludes my prepared remarks. Now we're going to turn it back over to Bob Ackerman for live Q&A.
Bob Ackerman
Okay, well thank you Charlie. Again, I want to remind everyone the ask a question box is your portal joining in this discussion. We've already got some questions in the queue. We'll be happy to take more. First question. Andrew, I think this one is your neighborhood. Securing a communications channel is usually a significant focus. What methods target ensuring that some communication gets through and not jammed for example?
Andrew Puryear
Yeah, it's a great question and it's actually one that spans many, many dimensions. So all the way from what's most important is not that some data gets through but the right information gets to the right place at the right time. So I think the first technique we like to employ is prioritizing which data you get where execute commanders all the way down through the physical layer where you're dynamically routing data through the network in selecting waveforms that exhibit ECCM sort of capabilities. So back here we showed you that we're learning patterns for jammers. We're learning those to ensure that we can select for that really critical piece of information a time slot, a frequency slot, a time and frequency slot where we can get the data right. Really, the answer is you know, the entire communication would be oriented around just what you said, right, how do we overcome adversarial jamming in a contested environment that we get the right information in the right place every time. And so by learning the entire typography all the way down the physical layer we're able to achieve that.
Bob Ackerman
Thank you. Okay, well, thank you. Charlie, I think I will toss this next one at you because it refers to PacStar. Which hypervisor providers have been certified to run on PacStar's compute platforms?
Charlie Kawasaki
Thanks Bob. Actually the list is really long. So the most popular one that we've deployed, you know, thousands and thousands of copies of, is VMware ESXi. Most of the CSfC certified solutions that we that we deliver today are certified assuming that VMware ESXi is already on board. But we've also demonstrated with lots of other top commercial providers, Hyper V and a bunch of the hyper-converged folks as well. KVM is supported. The PacStar servers are very compatible, you know. They're Intel, they use Intel MICs. We've really been able to have great success with writing all those types of technologies on our platform.
Bob Ackerman
Yeah, okay. Thank you. Next one, well. I'm going to play ping-pong here. This one is for Andrew. For AI operating in contested environments, have you considered the impacts of adversarial or counter AI?
Andrew Puryear
It's a very good question. Thank you. So perhaps it's best to start off with a bit of disambiguation because there's several contexts with which adversarial is associated with AI. So adversaries using AI for malicious purposes such as spear phishing at scale, something that we talked about earlier, we don't do that. Then there's adversarial generative networks or GANs, which we do that but that's not really the context of this question. So finally there's a huge field of research right now in academia and industry around adversarial AI or counter AI and this is where someone tries to game the algorithms. We've seen a lot of this right, so folks can make toy turtles that are 3D printed such that they're with high confidence misclassified as rifles. They can also create a sticker that you put on a stop sign that, you know, if I were reading, it looks like stop war, that's the sticker says war on it. But it's actually intended to game AI systems to classifying a stop sign as a 60 mile per hour sign. Given electromagnetic environments for AI, the AI will be applied to for tactical comms, we do expect adversaries to try to game the algorithms, whether it's a data poisoning attack, an invasion attack or an adversarial example, we're investigating how to prevent someone from hacking AI to ensure that we can deliver the capabilities that are promising.
Bob Ackerman
Thank you. Thank you. Um, next question I think is yours, Charlie. I'll direct it to you. Does this attempt to hide the wi-fi signals under emissions control?
Charlie Kawasaki
Well actually I'm not sure I'm the best person to answer that and let me explain why. When we're deploying things like CSfC dual-encrypted solutions that we have tended to incorporate those encryption technologies on standalone independent gateways. And that way by the time the data leaves our system it's transport independent and can go over you know, all the different sorts of tactical radios, commercial radios, military radios irrespective of their frequencies. Now I think Ultra has a lot more experience with this kind of techniques like spread spectrum and things like that. Andrew, do you want to chime in on that?
Andrew Puryear
Yeah, yeah. So thank you Charlie. For wi-fi, wi-fi is a very specific waveform, right? So the way that we try to remain low probability detection for wi-fi is through specific antenna patterns and controlling the power transmitted so that we minimize the probability of intercept, probability of detection. If we're unconstrained to use other waveforms then yes, we will use spread spectrum or either TDMA or FDMA to enable an additional layer of protection against detection.
Bob Ackerman
Thank you. Okay. Uh, well Charlie, I'm going to toss the next one at you as well. See if this one sticks. You mentioned your system has an UPS built in. What is it?
Charlie Kawasaki
Well Bob, I figured people might be interested in that so I brought along actually a physical sample of one of our products and what you can see here is a bunch of our modules including the new ultra base PacStar 464. That slides into a chassis and it has tactical radio battery connectors on the side. So what you can do is grab your PRC-152 radio battery and just snap it right onto the side of this thing and power it up. So I don't know if you can see it but I've got blinky lights going and as this thing boots. It's those same radio battery connectors - oh these by the way are hot swappable - we've got two so you can keep them running. It also has battery charging built right in. You can take the same module and slide it into our chassis which has MIL-STD 2590 lithium-ion batteries built right in the back. And so there's a lot of different options that you have for, you know, vehicle mounted solutions or grab and go or backpack solutions and they're all using standard military radio batteries so that our customers don't have to worry about a whole other battery supply chain.
Bob Ackerman
Okay. Andrew, I think this one is in your area. AI. Doesn't AI typically require large cloud compute resources? How could AI be deployed to the edge? That's an issue I think a lot of people are facing.
Andrew Puryear
That's a great question. AI definitely has a problem. In the quest to build more powerful algorithms, researchers are using ever greater amounts of data and compute power and relying on these centralized cloud services to actually implement those. This not only generates, quite honestly, an alarming amount of carbon emissions but also limits the ability of, limits our ability to deliver AI capabilities to the edge, bring AI into the fight. So a counter trend of what's called tiny AI is actually changing that. Researchers at Ultra and elsewhere are working on new algorithms that shrink existing deep learning models, convolutional neural networks et cetera, without losing their capabilities. Emerging trend also and in tandem with that is, so that's the algorithm part, the compute side is an emerging generation of specialized AI chips that promise to pack more computation power into tighter physical spaces so that we can train and run AI on parlous energy . These are both advances that we're working to make available to users.
Bob Ackerman
Okay, thank you. Charlie, I'm coming back at you with another one. How does this technology work with CUI, SBU or unclassified networks?
Charlie Kawasaki
Yeah, so one of the things that we almost discovered by accident was that we set out to create a whole bunch of CSfC solutions and these are the dual layer solutions. To do that you have to certify the encryption functions on your hardware, whether it's a VPN or a TLS protected application or a wi-fi controller with WPA2 and uh, but you need two of them. And so we went through FIPS, we went through common criteria, we made sure these were all STIG compliant and then what we discovered is that those exact same requirements are in place for the NIPR and the CUI and the CUI networks. So you have to be FIPS compliant, you have to be Common Criteria. So we discovered, just cut our CSfC solution in half and you have a pre-certified CUI or NIPRNet solution. So one of our customers, for example, U.S. Army, did exactly that and they have SIPR and NIPR running in the same case and they took it through and uh are using the identical technology for all the different networks so they don't have to worry about, you know, using the Type 1 over here for this classified network and something else for another classified network.
Bob Ackerman
Okay. Toss another one your way. CSfC looks complicated to put together. It looks like you need lots of certifications. How can my organization tackle such a complicated project?
Charlie Kawasaki
Yeah, well, yeah it is complicated to put together and we've been spending six years developing expertise to make that work. So it's helpful to understand that if you're an integrator or a government end user, that companies like PacStar and our partners already pre-certify our hardware and our solutions. So from PacStar and from Ultra what you'll have are components that are already validated for the CSfC solutions and they're on the NSA components list. The other thing that you can do to make your life easy is engage a trusted integrator and these are integrators who are qualified by NSA to build CSfC solutions and many of them have significant experience shepherding government organizations through, you know, the stigging and the RMF process and the NSA registration process for CSfC solutions. The last thing I'd say is we've done a bunch of solutions now that do have registrations and some of our customers are actually willing to help other customers, either delivering their complete set of IA artifacts or even making solutions available on behalf of other customers so if someone's interested in that we may be able to really accelerate the process by delivering stuff that's already been fielded elsewhere.
Bob Ackerman
Okay, thank you. Andrew, your favorite topic is in the next question. What are the biggest barriers to adoption of AI in an operational tactical environment?
Andrew Puryear
Thank you. Well you're exactly right, I do have a passion on this. So yeah, so the Defense Innovation Board has already done quite a bit of thinking around this. They've already articulated five principles for the use of AI in the defense department applications on responsible, equitable, traceable, reliable and governable. That a good list on the certainly entrance criteria but I'd also add three additional - resilience, explainability and predictability. So resilience. I mean, this is a slightly different use of the phrase resilience. AI is currently somewhat fragile. So I mentioned some of the adversarial attacks that people are able to launch on these algorithms. Resilience means that hey, we're able to train the algorithms on a certain data set and if the algorithms see something out there in an operational environment that doesn't fit within the box that was covered by that data set, it's going to react appropriately, react how you would want it to. so that's resilience. explainability is key - we need to be able to understand why these agents are making the decisions that they're making and then predictability, ensuring that, you know, in some operational scenario that this, that these agents will operate within some predictable box. So traditionally the way that things are certified, algorithms are certified in DoD applications is you're able to characterize that the input and output, for all inputs you know exactly what the output would be. The point of artificial intelligence machine learning is you can't do that. You shouldn't be able to do that. If you could do it you shouldn't use artificial intelligence and machine learning. But techniques to enable you to predict that the output will maximize whatever your objective function is, mission focused quality of service in this case, I think is also going to be critical.
Bob Ackerman
Thank you. Okay. Charlie, this next question is wrapped in quotes - iI don't know if they're quoting you but I'm going to read it anyway. 'This looks rugged.' But what do you mean when you say that?
Charlie Kawasaki
Oh, thanks Bob. Yeah that's a fairly broad term actually and even more so now that PacStar is part of Curtiss-Wright Defense Solutions. So our products have been designed initially for things like soldier carry, for things like command post applications and they are designed with the expectation that they have a certain amount of shelter. We do do things like MIL-STD-810 and 461 for uh, you know, blowing sand and dust and shock and vibe and temperature and EMI but when you want to start looking at things like you've got a wash down requirement where you're spraying it with a garden hose, that's not PacStar equipment but we have a partner business unit, Parvus where that's what they build and they have a similar suite of technologies to PacStar that support a similar set of capabilities. You can wash those things down with a hose and you know, but of course they're going to be heavier. You've got a lot more sheet metal around them, big connectors. Then on another part of the spectrum that we now serve as well is the embedded very high lifespan rugged equipment such as stuff that meets VITA standards and is designed for mounting in things like combat vehicles and that's supplied by our C5ISR BU. So as you can see, you know, I've got RJ45s on these devices so engineers will understand, yeah I probably don't want to wash that down with the hose but we have this mounted on all kinds of things now, aircraft such as the MB22 Osprey and C-130s, we've got it on ground vehicles like JLTV, amphibious vehicles but just don't mount it outside.
Bob Ackerman
Okay, understood. Well, that concludes our SIGNAL Media webinar for today. I want to thank our experts for their presentation and thanks to all of you for joining us. Our presenters will try to respond directly by email to any unanswered questions in the queue. You can link to the archived version of this webinar along with previous SIGNAL webinars on the SIGNAL magazine website at www.afcea.org/signal/webinar. Thank you again and have a good day.