Securing Critical Infrastructure, Navigating Tech. Risks - Paul Smith

yes all right great cool thank you very much uh so I actually don't own University Welcome to around facilities I won't say yes no yeah so um uh yeah I'm Paul Smith I'm a Prof here in the school of Compu Communications and I uh do research on um critical infrastructure security and resilience uh with a bit of a focus on uh Energy Systems systems and that sort of thing and um what I'm going to try and talk about today is sort of um some of the risks and benefits associated with using various sorts of digital Technologies Computing Technologies in this context um so I have another version of this talk or another slide is like basically

how you kind of maintain keep the lights on whilst we're using sort of Technologies like AI ml all this kind of um so I think there's a sort of tendency in the security Community to sort of think about these Technologies in particular uh in theis infrastructure sector think about Technology Innovation isn't necessarily negative thing right introduces sort of more abilities it's true uh but there's also sort of opportunities uh with the use with these Technologies to support sence security uh I'm going to try and talk about a couple of I think uh new cases working on people students in this space all right that make sense yes good uh if if if you want to

interrupt me and tell me I'm speaking nonsense or too fast go for it yeah uh I won't be upset too much all right so um as we all know soes I'm going to move a lot um yeah so providing all sorts of important services that Mak list elect energy and distribution system what Supply saw talk on Transportation sessions really interesting uh these are sort of being increasingly digitalized so basically using computer based systems systems so in the past know a lot of these entirely analog uh system right nonu various reasons to use more and more Computing technology um and sort of a part of that kind of digitalization is a sort of conversion of what we think about information

techology so of things that we sort of use on a day-to-day basis commercial off the shelf equipment laptops uh Microsoft operating system stuff with sort of operational technology which is typically quite different in flavor yeah [Music] um uh talk a little bit more about what form that it and OT converion looks like um done already [Music] associated with this digitalization they also RAC sorry RAC all right so uh so this who's heard of this term I convergence before speaking so basically the kind of traditionally sort of operation technology operating spere of Engineers control engine and that sort of thing and um of organizations want to sort of optimize the performance operations andu that try to

increasingly net and integrate business size oranization with operations get data out to those system um toble to do things like main for example so a turbine you know all time if you can kind of predict face before c sa a lot money um so that's kind of why we're trying to do this and sort of this my attempt to be sort of psh is that you know this is sort of seems Mark kind integration and there's two ways to think about this this this convergence one is from an architectural point of view when I say architectur l the Network systems architecture yeah so in the past a systems the stuff that sort of you can

find in electricity substations nucle facilities water Delivery Systems in this connected separated by so so there this funny thing like unicons and aaps

common fi it's still a very useful way of securing systems right so basically are not connected to systems not connected to if you want do updates use things like media USB pend on this is changing in several ways so Enterprise systems so the kind of things that you see in the business side being connected to them uh via DM you know to get uh basically data out of s specialized databas called dataor the IDE you can run models over the top of that to see when fail mediates and all this sort of Stu this is a sort of long running thing vendors and suppliers of the equipment environments would like to have remote access yeah so that they can do

maintenance so from the vpm deep into the infrastructure that operating risks uh there's a lot of talk about supply chain security so what your suppli is doing ensuring that your suppliers uh secure as they sort of access your physical assets increasingly field devices are different sorts so the things that you find out in the field operating substations and so on want to try to connect out to Services yeah so anecdotally my old place through working on digital substation and put in this device and it would work until to Cloud right so this is a sort of a trend that's going on and then the kind of devices that you're using to maintain configure them of

Internet connected or commercial shelf so you about this ter industry 4.0 there's a lot of new technology in this space that t where you can monitor process control set points how you can figure it so on so anyway what we end up with is lots and lots of ways in and out of the system that typically previously not connected to each other which know makes our attx surface a lot larger and introduces a greater risk the other kind of um side of this is that there sort of Technology changes um that sort of the typical thought of is dominantly from the it domain moving into the operation technology were historically a very limited technology so theal sort of bit

of technology that use is a progam control so this measurements from from from The Real World runs a control algorithm and actuates to change things like breaker change Val Mak Sur plaster whatever yeah so these are sort of industrialized pieces of equipment should that um just to show you but they sort of very very of limited in their capability so running real time operating system and that's about it right um not a lot else and the other thing historically very very from the security point right that was not really a thing 10 15 years ago in field for 20 30 years so people like you like to have things got a lot of fun with

them this is changing so they getting increased competitional comput capabilities so have more compute more memory more storage than they ever did before uh and I'll keep show you an example of what type of device that's in that direction at the moment using virtualization platforms so basically being able to separate us visualization to like real time operating system part system and then other kind of more general purpose operating systems to services on them if you were into sort of I think it was a competition for uh if you say certain Buzz words you a drink right digital twins with one yeah so digital twins there's a lot of um talk about using digital twins in this space basically

these are virtual representations of the real world system that you can do analytics on again things like maintenance so it's like a model thatle data from data from The Real World feed it to model then use that to support decisions about how to operate sometimes ously sometimes with in the loop and these sort of there's kind of an implied often you talk about twins implied use of the cloud to sort of locate the models of course our friend AI in ml which we heard a lot about uh before the break um and internet protocols uh for monitoring and control or Internet like protocols so historically these protocols that were used to um control industrial control

systems were of and closed yeah so not open like you would get from the internet uh a little bit obscure right so it's a bit of security by security going but uh sort of some examples of kind of industrial control systems decid to use more internet like protocols so in here was involed in a project looking at electri charging descriptions yeah so the sort of things car side of road they control using a protocol called ocp open charge Point protocol that's entirely web sockets Json base right and the reason to do that is because it's cheap and easy to develop implementations that compared to uh what it was in the past you get to Market that kind of drive to use more of

these kind internet and like yeah so these program controllers that have hats running on them yeah which is really cool and interesting but these are of apps running in the core of infrastructure yeah and you know what happens when he goes wrong you go to Schneider Electric or seens uh sheets contact or whatever you choose your there and what anecdotally we found is what they say well that's not our problem need to go to the app so again this isort of interesting issue around response so I've sort of tried to lay out um some of the risks sorry so they kind of basically where the sort of industry is going as supp in terms of

this digitalization how are we feeling about that confused con concerned yeah you should be um which is why you need to fund research [Music] whoever all good plug you have to get it in there right all right so what I'm going to do is talk about some of the kind of risks and benefits of this digitalization that uh I've been working on with with some some students and and stuff over the over the last year uh so the first thing I'm going to talk about is our friend machine learning everybody's an expert on machine learning right in the room you're an expert yeah expert is always a relative term right relative to the other individuals in the room

right um so basically the way the way machine Learning Works is basically you have a training phase we have a bunch of training data do some feature extraction normally so if using network data for example nwork traffic captures or something like this could do some uh feature extractions oring at entropy of ports and IP addresses and and blah that and then you train test and validate some some model right this is for predicted machine learning not the kind of gen gen stuff that for the break a little bit different I'm not person yes and then no problem uh and then the sort of um then once you sort of train your model you s of then use it

operationally so you get some sort of sample and image and network flow summary or something like this you querium model and it hopefully gives you an influence of prediction yeah so is the image person or an animal or a certain type of animal is a FL representative Serv attack or or port scan or it benign or malicious depending on what kind of model you're using and be use for and all through these kind of like uh if you like life cycle or training phas operation different types of attacks that could manifest yeah so you could poison the training data steal the model clear the model to find out the training data all sorts of different

sorts of things that you can do and lots of research is how you how you kind of what are the different types of atts and how you protect ma model I worked with colleages a few years ago on a particular type of attack to machine learning models called aerial examples aside from the 439 students who's we have some experts in the room basically what they are are carefully crafted Imports to the machine in the learning model that are intentionally cored to subvert the prediction that the model is making yeah so you make manipulations for the query in this little example here this is a sort of very classic example by in fell where you uh make modifications to pixels of

an image so that we can't see uh any change but the machine learn basically a bit of noise or some noise and then the model predicts yeah so the machine learning model given so you take a training model train model manipulate query it gives you an incorrect classification and there are different classes of this type of attacks targeted attacks where you say I want this input to be always benign or always to be a port scan or always to be given whatever it is that you want to do now this sounds like a lot of fun right yeah it's really cool until you start using machine learning models in critical infrastructures that are support in

control for example so there's a lot of our sort of model control manipulating Control Systems set points and all this St and you can use uh machine learning for cyber security right so there's a lot I'm sure you've heard um of kind of work looking products that ought to use Amal to support things like virus detection and network traffic classification [Music] yeah [Music] so uh exactly yeah dark trace for example is a good example so one of the things that some colleagues and I looked at is to see if we could generate up serial examples um uh for inputs to a model that's been used to do Network intrusion so in this case we were

looking at sort of multi classification model is doing classification basically you feed Network form Network summary summary of network exchange and then the model predicts what it thinks it is is it is it ATT is it P or something else okay what's interesting about this or challenging about this is that unlike a sort of panda where the constraint is only that it needs to look like a panda to you and I the end you can't arbitrarily change the features of a network flow there are constraints on how you can manipulate the flow because it very quickly doesn't make any sense so if you change one feature it has to be consistent you do that and if you're

an adversary out of the internet and you want to essentially generate Flor records using your book by changing your behavior a little bit there's only so much there's only certain features that you can affect because for example you don't control features Rel to reverse reverse direction that making some sense yeah good finally I this was 2020 so it's 2024 yeah uh um I've been able to explain this in way so the the long story short is is that um we were able to generate seral examples by manipulating so basically what we did is we had an algorithm that was uh manipulating features uh and giving it constraints we said you can't touch these features touch these features Go Go Wild on these

other features yeah because these are things that we can control arbitrarily and generate an SLE C classification of of a network FL and what we were able to do was of get similar this is Det this get similar uh sorts of results in terms of our ability to craft samples and uh respecting the constraints that are in place uh with respect to network F that make sense yeah so these problems exist lot of the plac physical systems or industry Control Systems right so if we have uh machine learning models that are supporting um I don't know the behavior of control system that's uh in the electricity distribution domain you can't arbitrarily change things like phase voltage power there there's L of

physics in place right so there's kind of interesting sort of parallels in physical systems domain where these models are being used and some people have done some work on on generating apperal examples for those sorts of systems so that's a big risk and you know um there are lots of kind of uh ways that you can make these sort of models more robust to these types of attacks so you train them in an adversarial way yeah but it's never going to be perfect yeah so the the you know the attackers are going to get smarter or whatever or we'll miss it or whatever so we now need to account for that uh right I need to pick up the pace

uh just to give you a bit of a flavor of some technology that we've been playing with so as I mentioned earlier there this kind of iic of industrial devices so the the the the sort industrial device sits almost every infrastructure does some sort of process control AIC controller yeah it's a sort of industrialized device that's me to sort of sit in hostile environments and it normally runs like I said earlier a real time operating system and you basically provide called user program which is a control program takes some measurements make some sort decision and then actuates so something like a p control control systems engineer um so seens have this product the S7 1518 mfp multifunction platform

LC all about industry 4.0 sort of improving efficiency and all this kind of stuff and they have this uh functionality on there called a CP C++ run time and it's basically sort of modified version of Linux that you can sit on this little PLT sixs for example on your power manufacturing uh plan and you can kind of run here and get information about process via protocolization sounds like a horrendous night from the C security point of view right this general purpose oper system into into this environment um this we saw an opportunity for cyber right so what we did is we came came so came up with this this little to Goose I left my student

alone for a little bit too long and he came with his name um Goose makes noise and the W is um um and basically what the the state of the PLC so like has it been restarted cycle times changing and also um monitor whether it's behav controlling the system correctly is it trying to apply on safe control options and if it is it generates an alarm and this is a sort of way of implementing some U good practices with P some security coding oops so there's opportunity for this right we can put it right on the PLC so a smaller attack surface it's hard for Pro to to manipulate digital twins this is thing that's coming people are investing fast

amounts of money in twins um largely for operational reasons do things like predictive maintenance and so on as I said earlier uh there's a lot of security issues around um use of digital twins so the twin2 represents a lot of intellectual property and if you're in certain sectors represents sensitive information operation facility and you also have a control Loop right you measurements coming out controls going in right so you can manipulate this manipulate this manipulate this and you can make bad decisions right if this Lo doesn't represent reality um so to protect it like with any control system way what we we've been looking at is can we use uh a digital T to support s

security incident response uh fores where usually there's a bit of a disconnect between sort of it in resp like an 19 Operation Center and maybe there's an operational technology SC Operation Center but he use a digital twin to get situational awareness use it to do all sorts of different types of queries about okay if I change the system what the behavior that's going to occur because we have high ability requirements so making changes different pressures is difficult all right so this is my research peacock uh I said peacock um sorry to be certain to that um so so so there's all this digital technology coming right and we could just very our head in sand and say this is a terrible

idea stop doing it and then life will carry on and Commercial Enterprises will keep having these capabilities the system they'll find out way we can't do that we've got to somehow Embrace to some extent in a risk anded way this sort ination and some of this Innovation some of these things are necessary in order to operate some of these critical infrastructures in a way that we need them to to do in the future right to be cost effective like you think about rable Energy System Energy System energy sources we to Opera in a smarter way than you do now right so what we need is sort of approaches to realize the use of these Technologies in a safe because safety is

a huge topet here in Secure way right so we need so technologies that are reped to to attack right so all of the stuff that we do around hardening systems and all that kind of stuff when we think about a ml we need to train our models right keep training them to pen testing against ouri all that cor of the Technologies secur as we can get then we need to think about how these Technologies assist in a wider system architecture right app principles things you know like defens and death yeah which is sort of well established in these field uh Gils in place so if the AI model tells you to do something stupid like limitation layer that says

no don't do that that's that's not safe so St doing it people and organizations a lot of the use of this technology requires modifications to the way that we operate nor yeah more digital more convergence to this Human Performance issues that we need to come up with this process issues that need to address regulation of course so we can all come up with really really great ideas all this Innovation and reg says sorry and then we spend billions or Millions on technology for context not Mee approval so we need to think about how do we uh give the tools to Regulators to support analysis of these Technologies in place yeah how do we make a case as people that want to use

these potentially to a regulator like the use of these is safe and secure one of the tool set one of the arguments new talk about arguments evidence cases argum weop and of course we need good guidance and right good practices can follow um some of those are coming right but that's probably still a little bit uh there's a summary I guess my kind take away is when you when you look at this there's a bit of tendency for us to S think about these things in a very negative fashion right so let's let see around say okay what opportunities do this realize for Cy security and then how do we go about introducing some of these Technologies

in the set be away uh I want to say thanks to all these very very clever people who helped me or actually did all the work on just the mouthpiece thank you