Events Add an event Speakers Talks Collections
 
RSAC 365 Virtual Summit
January 27, 2021, Online
RSAC 365 Virtual Summit
Request Q&A
RSAC 365 Virtual Summit
From the conference
RSAC 365 Virtual Summit
Request Q&A
Video
Lessons Learned Developing an AI Cyber Immune System: The Good, Bad & Ugly
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
146
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

This session will include findings and insight about both the opportunity as well as the challenges to transform organizations from cybersecurity to cyber resilience, including: How to establish and lead their own diverse teams down a roadmap toward resilience via a holistic people, process, technology solution.

About speaker

Michael Mylrea
Senior Director of Cybersecurity R&D (ICS, IoT, IIoT) at GE Global Research

Michael Mylrea is the Senior Director of Cybersecurity at GE Global Research, where he leads various cybersecurity R&D efforts, focusing on applied AI/ML and securing and optimizing IoT/ICS. Mylrea has 18+ years of experience developing innovative solutions for complex cybersecurity and technology challenges for industry and government. He holds cyber-patents, 50+ publications, launched a successful cyber-consultancy and held senior positions with US Departments of Energy and Defense, Cyber Innovation Development (CISO), Deloitte, US Cyber Consequences Unit, Lakeside Oil and Harvard Berkman Center. He completed his doctorate as a CyberCorps Fellow at GWU, focused on cyber-resilience and completed various degrees, courses and certifications at Tufts, UW-Madison, WGU, SANS, Harvard and MIT Sloan.

View the profile
Share

I am Michael Millwright director cybersecurity rmd from GE Global Research on the next 30. 35 minutes. I'm going to take you all on a journey in developing one of the first day. I driven industrial immune systems. Really highlighting the opportunities and challenges that good bad and ugly, really Lessons Learned From developing a paper defense. The secure, the world's most critical systems from Aviation of power grids weapon platforms to healthcare, the concept of how we got here.

Starting a major evolution of AI for cybersecurity. I'm really has come from advances in storage compute in mathematics. And while there's a lot of noise and kind of misconceptions Fai, cybersecurity can replace the need for human cyber. Defenders. Take humans out of the loop. You know, we've heard the robots are coming. They're coming for your job. That couldn't be further from the truth. So I'll just be mystified. Hey, I talked about what, it's really good at for this face, as well as some of the gas in it and some of the considerations you

should have when you go back to your organization. To the critical areas are all AI cybersecurity solutions that are effective really require. Our number-one humans. They required deep domain expertise and I like to think of this like you need Architects and Builders. So the architects in this case are the Deep domain cybersecurity Assamese, the subject matter experts the builders are your data scientist and your math magician Wizards. So to say without either one of those AI cybersecurity Solutions are not very effective. Other critical ingredient

in. This mixture here is data, you know, the old adage garbage in garbage out could not be more true for these Solutions. So you really need access to that training day. Make sure that that data is prepped and prepared to be ingested by your machine learning algorithm. And so what this Evolution we seen advancements from kind of a first and second wave of AI. Moving towards a third wave of, they are really improving architectural reasoning, our accuracy and with this, and seen major, major developments in in improving cyber threat intelligence or able

to really tackle some of the challenges with the scale of the problem, a lot lot better. But with all that, it's still not a Panacea. It's knocking, replace your side of the fenders and in many cases as well discuss often times don't need more humans in the loop to really do this right now. So let's start out with the Bad and the Ugly at least, let's talk about the problem. We're trying to solve here before we get into some of the good and great. If you will, I like to use energy infrastructure as an example because all modern organizations are energy organizations

in the sense that there's not a single modern organization that could function without access to electricity and the last 20 years. We've Eyes and networking automated our entire energy value. So we can look at energy infrastructure or Transportation or health or defense your seeing this trend of the digital station networking, which has woven together or information technology and our operational technology, or cyber-physical systems include the incredible tapestry that has significantly increased the speed, the size, and the volume of data,

requirements, being collected aggregated and exchanged and from a cyber perspective. That's reason number of challenges. But the drivers behind this is pure and value creation of Mater City smart. It's helped us in the gridspace. Balance distributed energy resources to Casa close at Maids. Are it makes our transportation system more efficient at makes us better at our health Diagnostics, but with all this week. Big, big data problem. How do we make sense and find patterns in these data sets? Again, not to take humans out a loop

at least with a lot of their traditional roles that can make us better at what we do a grid. Cyber Defender a good one. If we do this right away, I can make a good one break, but it's not again. If we're going to take that human out of the loop again. So that's it. That's a big opportunity to challenge and all. This is the number of the iot or industrial iot, especially the Legacy systems were never designed to be connected to. These are systems that prioritize functionality and ease-of-use encryption authentication, communicating plane attacks.

And you see a lot of human machine. Interface has that connects these critical systems that keep trains on the rails, defense platforms accurate and reliable. You see them. Default passwords and said, this is the challenge. It's a, if there's a scale Challenge and how do we, how do we make sense of and how do we detect and how do we monitor, these large data sets and say, I was really good at identifying those patterns? And through a lot of things like spam, a threat intelligence, identifying those those those patterns in these large lights, datasets is critical.

The challenge being is that our adversaries are also rapidly evolved. We have, we are at we are dealing with cyber adversaries. What it will? Look at the solar winds attack recent solar winds are the grid. Cyber incidents are Ukraine. In Ukraine, you can get our trait and triceps, we could go on and I will, or seeing if you have an adversary that's in flax nonlinear and rapidly evolving. And so we need to evolve our defenses. AI machine learning, use them interchangeably. They are very different understanding that. We need to evolve those defenses to

keep up with that threaten. Keep up with the tactics techniques procedures. And so, what we do from for what we developed this, industrial uses than we'd studied very closely. The tactics techno. The procedures used by the adversary to carry out these attacks in a couple of interesting Revelations. That really high like the limitation of our current cyber-defense it. I'm number one. We see that and it doesn't always need malware or some type of zero-day exploit. Once they get privileged access to your human machine interface often times organization have no

visibility and spells lower layers of industrial control system. Those critical sensors and actuators those controllers again, that keep trains on the rails. That keep transportation systems are reliable and safe, and efficient. Power is balanced. Those critical systems. Often times when I've visibility you saw, once they get access to that team in machine interface. No longer do. They need malware? It's legitimate command. Using the native protocol of the system in the wrong succession, or at the wrong time, of both can actually cause a

physical impact. And so, our intrusion detection systems that work on a signature heuristic. Are we seeing these ones and zeros before? An AI is really good at recognizing patterns are often times, very very limited. So, what the industrial system we've developed is focus, much more on the physics. So we're looking at changes in voltage and frequency in ambient temperatures, as well as the networking there, since it will get into that a little bit later, but you see that the scale of the problem and yai application is the space are

so important 350,000. New pieces of malware registered each day, very very good at recognizing, those signatures and her Fixing building up our robust libraries, identified an hour, but when we start asking questions around, contextual reasoning things that humans are really good at and what was the motivation? Your behavior of our adversary? What were they doing? In that German steel mill. When they were lateraling around different parts of the networking and something went boom. Did they interrogate one of those systems in the wrong way? That

caused it to to go boom, or was that part of their goal? And motivation is really, really bad at that. It's really bad at understanding contacts. And that's really what the third wave is trying to figure out, not touch the attacks, you know about a fifth of the world's shipping capacity. Shut down. Other critical systems. Plus, 10 billion dollars. Worth of damage was asked to be part of the expert. Testimony forensic team is made up of a Ferry County group of individuals. Much more talented than myself, a machine-learning. A way. I

could have potentially identified in the earlier stage. I'm but what keeps me up at night here on this side of all these big numbers. It's really the two million cyber security position sensor on sale again for AI Cyber Solutions platform to work. We need a lot more. He needs deep domain, cyber Security, Experts to fill those roles. We need more diverse. We need more women. We need more people from different diverse backgrounds. Joining the workforce. Because if not, we're going to have the same myopia. A lot of our intrusion detection systems and firewalls have

is that we're all thinking through the same set of prisons and filters. Looking at the challenge when the adversaries evolving and very, very diverse. And then we also have to involve to the current environment, you know, when the world flipped upside down and started the unprecedented. Find during the pandemic. We saw a lot more most organizations now operated remotely. So this remote connectivity, if you're AI solutions, they wait, why, why is the substation? Why is this train? Why is this defense platform now? Have these remote VPN connection? That's an anomaly. That

must be an attack. You're going to get from Hibbett of false positives. Destroy your AI solution out the door cuz it recognizes patterns based on a summation of all its perceptions of previous experience that have been training the features and that's as it's not going to work. Well, we saw with the pandemic, this huge increase in our attack surface and open publicly Bond herbal systems, which is absolutely frightened. You can look for a nickel systems. Are you see where AI is really good at identifying those that small changes and malware and she's every day, but not as good as

understanding the behavior and motivation behind the anomaly. I think in all this and you know, it's a whole separate presentation, but the current Paradigm that were using respond to these two birds. I see in which environs are very much limited. Actually Focus too much on technology and not enough on the people in process. That makes the solution holistic. And so we need to marry human in machines from an AI for the very much. So to fill some of these gaps. I'm so that's that is the problem that some of the Bad and the Ugly. I think that was the good is also good. Is

recognizing gas and some of the major gaps in this space. Yeah. Absolutely. We want to make her a i l l goes lighter and more efficient and closer to the edge and in better context the reasoning. But to do that, we have to understand trust. We have to understand how we can establish human operator trust, so that human machine human. Great cyber. Defenders are human cyber Defenders. They trust inclusion. And that's part of the Big Challenge and part of the whole area of research run explainable to develop

find Quality Inn truth or going through 2, human understandable decisions with our algorithms. A great example of that that played out and really sad. When, is that a couple years ago, a suicidal German pilot took off the AI autopilot with the goal to crash, a commercial airline. And so that autopilot would have said, corrected, that nose-diving suicidal pilot, but the pilot took that off and, you know, hundreds of people lost their lies and so explainable II and that trustworthiness, it actually goes both ways. It's crazy to think. But how does that

algrim actually trust a human operator? Especially in an area and cybersecurity and cyberwar world where humans, often times our The weakest link as well as that the last line of defense. So there's there's a huge area of research there and part of that research has humbly I and improving awareness of think about imagine a fleet of autonomous drones that are about to carry out a mission in cooperation with ground troops. They locked in Target and I'll suddenly experienced a denial-of-service attack. Humbly. I is part of this third way for this autonomous

weaponized drones, recognize our limitations. Go back to the human operators. Say, I have a lower degree of confidence in my threshold of decision-making, that I can accurately hit this target without any casualties on the ground. And so, that's where we're going with humbly. I like I'm some lighter versions of that could be with diagnostic or an algorithm is going to Diagnostic and helping a doctor. For example, that's good helping them become. And say, hey, I'm not quite sure what this you need to check. Just to help validate verify my findings

or a wind turbine generator, right? Where we think there's a fault in this Fleet with one and three but not too, I need more data. Here's my current evidence and part of getting their part of realizing those goals of third wave is really. Now it's Einstein's reminded us as half and answers in a really good question because you go back to your organization's and you assess your own platforms. You need to ask questions. Like how do you establish contact interdependently among human machine teams that are in power. Bi Nai house, trust affected when humans and machines depend on each

other and how should human machine teams playing with each other really to improve the state-of-the-art an answer. These questions. We need to establish a bi-directional language of explanation to really improve context and trust between human machine. So getting all this the robots are coming, the robots are two years away is here, but it's going to probably, at least in the short-term require a lot more human domain experts. Then a lot of people give credit. So what are some of the bad and ugly? We see are

also rapidly evolving their own to include AI machine learning. So, one thing that the victim was happening right now. Is there a reverse engineering, a lot of AI systems so they can better understand what their decision manifolds are. They can understand that black box nature. This is how the algorithm detected and localized. This specific attack. Once they understand that that can pave the way for food poisoning or enabling them to replicate the system in the manipulate the system with different adversarial. AI attacks were also seen in. This is a whole separate

presentation by the increasing use of deep state. Social engineering is spearfishing. We're seeing the use of AI for Behavioral analysis to find that a tweaker Link in the chain. If you will, we talked about reverse engineering and then we're also seeing some, you know, malware 2.0 as they smell polymorphic a. I type of tax that brute force and just keep attacking till they get around those signature sticks in your firewall. That is something that has been happening for a while now and we've all seen this, you know, the results and impact of some of these issues and

malicious attacks, a huge challenge with a black box nature of AI or an algorithm arrives. At its conclusion and see that this is no longer a panda. This is a given just with a small picture, bation, a small change in in the image, we have Project going on with facial, recognition AI with DARPA. What's the only thing that is funny is pandas. And that's, of course, Will Ferrell. And if you see what this small changes, sometimes, I'm in the image set in the data. We can actually manipulate the results of the algorithm we can cause

and that's, and that's really scary. I mean, you're talking about major major ethics issues, major bias issues, major issues, especially when you're talking about AI controlling these critical critical systems that I'm talking about in aviation, in the power grid two of the big areas, tax. We talked a little bit about Shirley. I really think about poisoning, how the algorithms trained in development, or in this cross-validation stage, or even in the testing on Southern. Seasoning. And there's

also a Bayesian attacks, which are carefully. Your carefully adjusting in. Misclassifying the results of the algorithms are allowing that attack action in support of things. Some of the things we can do. I'm to get around and defend ourselves. From some of those adversarial attacks are really, really real firms. The importance of understanding their near boundary condition. So understanding where that algorithm failed to detect and what you can do it during the learning stage as you continue to an attack

that that algorithm to really really have that really good understanding of those near boundary conditions on which is really an area that is going to focus on. Having the completeness, a robust set of current boundary conditions. Is that absolutely critical. We talked a little bit about adversarial training defensive distillation. Strategy where we train them, all the output probabilities of different classes rather than hard decisions about which class output. And so then those probabilities are supplied to the model and then their training, the class and was really hard

label. This makes it different more difficult to exploit. So with all of this, with all these advances, some additional bad and ugly is, is again, he's not a Panacea. Let's go back to garbage in garbage out to need those Rich data sets. There is a limitation of Cyber attack data in cyber threat intelligence. There's a limitation and understanding the motivation behind the attacks. Why they're being carried out. So establishing these classifications and feature sets explainable the eyes, very difficult. And there's also the issue that is if you drop an AI Solution on a network or

system, it's already compromised. Think if you introduced in the ice, I was curious Lucian last eight months when you're Network systems potentially were infiltrated to the solar winds attack and then that Solutions train. Hey, this is what normal looks like. This is normal. I would, in fact, you're compromised, you're getting a false sense of security. I think the other issue is actress he's so if you're getting a lot of false positives like think about that, as the pandemic hit me offered to work for remotely, work from home. If your algorithms,

and solution was an agile and adaptable to that. It could be tuned in train. I'm you're going to get for Hibbett of numbers of false positives and just turn the turn, the thing off, and it's not going to be effective at all. And then in the otics face or you have a lot of stochastic, loads you have trans you, you have computational environmental, human factors that are changing what normal looks like establishing that manifold is very very difficult. A great example of that is A jet engine, you know, how do you develop an industrial mean system for a jet engine?

That's flying at different times to different fireman's. There's a lot of pollutants in certain environments that will cause a faster degradation curve of that system. And so, how does your AI solution adapt and recognize those natural degradation curves from the environment are fun? How it's being operated, not a trivial task to do, especially considering those requirements to get it right? Is that you can't have those false positives or it's not effective. And so I know what the solution we develop part of our journey as it's called digital

ghost. I we wanted to take a very different approach based on some of those very sophisticated attacks, sophisticated attacks and incidents we discussed earlier. We wanted to look at the physics of the system because while you can Flip a better one or zero to get around a signature horrific is polymorphic attacks. A guy attacks that are going to be around your firewalls. The physics are more difficult to spoof and we have deep deep domain expertise in the physics of our assets. We take a defense-in-depth approach to develop

impressive multidisciplinary team from different backgrounds, looking at everything from how white set white cells of a cat in your, your body's immune system attack on different, viruses in different seasons. When your attack will. You look at this from a very, very diverse perspective. And what we started out with his again, you those two core components, any effect over solution. You need the domain expertise in space. And the data scientist, of course, is part of that. And then you also need these really

rich, big data sets to train. And then once you have those, what we've developed, we start out with a digital twin starts out with a digital twin living breathing at a set of the system, include a high-fidelity view of the control systems, the controllers, right? Cuz the controllers can dictate the physics of the overall system. And so these are highly complex tightly coupled systems of system with hundreds of sensors and so changes in the environment

and operations and degradation curse can really change what Mama looks like. So it's not easy to get right. This is where this is the area. We are operating in because we, again, back to some of the earlier incident realize that once an adversary gets in, They no longer need now. We're to cause damage, you need to be able to look at the house of the system. Not only from a fault. What are the natural system to operate outside of its manifold? And we'll work with what we have. The ability in

doing this and taking this approach to use our forty machine learning algorithm detective localized, where we've been attacked. So here once we have that digital twin up and we're training, the learning were attacking it constantly using very sophisticated. Attach this tablet or manifold is what's normal. What's in a cat? What's a fault? Establish more features and this is exactly the type of solution your lawyers. We've seen with cybertech stuxnet, solarwinds

focused on the ones, and zeros at the networking layer working focusing on. What normal looks like so whether this is an Insider attack, a zero-day exploit operator error. Either way you operate outside of its normal manifold. It's going to be test in real time. Exactly. Which one of those sensors is being manipulated localize, the sensor and then it's going to use those algorithms and there's other sensors that are not compromised to neutralize the impact

of that attack. This is really a new paradigm in cybersecurity. This is a completely different approach. I were saying that in these converge itot environments with International Supply chains, with hundreds of thousands of lines of code, with malicious insiders, going to find a way to get in, or going to find some type of way to get free. This is more of a resilient solution and that once an adversary gets in we can respond with covered in dirt. He's attacking keep these critical assets morning lawn

and I want to conclude with what what exactly does that look like. I better get in here. You see no currently when there's some type of fault with sensor actuator in a combined cycle. Power plant, for example, has seen because you have his Transit pipe tightly coupled system, or one thing goes wrong, is going to have an impact on the system of system. So in addition to detecting localized and neutralizing cyber-attacks other really exciting thing here is where no longer. Cyber security system cost for finding real ways

year to reduce operations and maintenance and improve efficiency and crew, production include and energy management and then again back to making a good operator. Great. And so that's what you know, that's what we're doing here is very, very different. It's not a Panacea, but it's a unique way of the unique approach to fought to apply advances in AI machine learning advances in physics, advances in computer storage mathematics of the only improve the state-of-the-art. So this is what it looks like. I'm this is some of the life testing

the actually performed on on one of the world's largest combined cycle, power plant. Where we set it, put on our hacker hat. We said if we wanted to cause damage, I'm through this critical system. What are the sensors actually has control systems that we would try to manipulate us to cause the most damage? The system is very, very stealthy. When I'm, you seen the truth versus the neutralization algorithm of reconstructed. I'm so, even well on Drina attack, the Centre, using with machine guns algorithm

to reconstruct the attacks sensor. The other way this could have played out as we saw was stuxnet, was stuck. Sent you unfreeze your own exploits. You had the human machine interface. Telling leave the scientist that everything is okay. You know, it's spinning at $30,000 p.m. When it was potentially spinning at $60,000 p.m. And causing degradation and destruction of the app. And a lot of trust between these human machine scene and sell welded detection and localization, algorithms are very, very mature. We

demonstrated a 99% accuracy. So this was a major scientific accomplishment. I'm even with all that and we've demonstrated neutralization, major scientific accomplishment, even with all that, as I mentioned earlier. This is not a Panacea. The gaps, the areas, the research gaps here, are very much related to human machine, interdependence cross, explainable a. I humbly, I it's only with those advances in filling those gaps. Will we get to the Third Way for artificial intelligence? Is it? In fact, intelligent?

You still have a number of barriers to that, as I mentioned? So in conclusion and realizing here, I'm getting low on time. And if these are some of our results when you go back to your organization, when you either assess your current solution or consider investing in one, I'm know that know what they're good at and know where there's a gap know, where the opportunities and challenges are. No, the good bad and ugly. Remember that. This is not going to replace your side working. So if you were Cicero or

in some type of Cyprus, three roll and considering the investment and a i driven solution, you know, you need that really really understand things. Like what's your level of false positives? Are you able to detect and localized attacks in the converge itot environment? What type of OT, protocols? Can you handle? Are you just dealing with networking? Are you just dealing with ccpit remember? Also remember that for any solution to be effective? It's not about technology. It's also About the people in process. The form of the technology has to compliment the function and so

AI. Cyber as we learn is very good to help a scale and deal with the increasing volume of attacks. Not a Panacea know that as we adopt new a, I typed driven Solutions a new TTP, a new framework. Our adversary will continue to adopt similar solution. Six boys is Emma trees which will continue most likely to be in their advantage. I'm so wet that I can clued. Really appreciate the opportunity to join the RS. 18 talented team and an audience of great to be back.

Please, stay in touch. Thank you very much.

Cackle comments for the website

Buy this talk

Access to the talk “Lessons Learned Developing an AI Cyber Immune System: The Good, Bad & Ugly”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “RSAC 365 Virtual Summit”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Similar talks

Jason Rivera
Director, Strategic Threat Advisory Group at CrowdStrike
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Israel Barak
CISO at Cybereason
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Behnam Dayanim
partner, chair of Advertising, Gaming & Promotions and co-chair of Privacy & Cybersecurity practices at Paul Hastings LLP
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “Lessons Learned Developing an AI Cyber Immune System: The Good, Bad & Ugly”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
843 conferences
34172 speakers
12918 hours of content