ITBN Tune-in: Self Learning Security Algorithms - David Aviv

Okay, hi everyone. Once again, my name is David Aviv, Radware CTO. And just as a continuation of the session that we have today, the place of AI, we try to start with what we call self-learning technology, fascinating software on the way to AI or part of AI. So I would like really to shed light on what we mean and the way and what are the borderlines between different AI technologies, different algorithms.

So not surprisingly, as we started to say, it's not only AI that we are starting to see. We see automation in the sense of automation. You can look at it in two ways: the bad guys and the good guys. The bad guys use automated tools to create new attacks, like in IoT. Think of the bots, new generation of bots. We'll talk a little bit of a new generation of an AI automated bot called swarm bots, using swarms to attack the network. So AI can be used by the defense guys to automate the defense, to automate the security policies and others. So I'd like to combine them together, and definitely those are two trends that shape the industry, the security, the cyber industry of today in all aspects.

Start with an example of the IoT, the pre-IoT examples. Usually, and I'm talking now on networks, could be any network, enterprise, service provider, usually we wouldn't be talking on high volume attacks, very high volume attacks, you know, DNS, SMTP, etcetera, targeted reflection attacks. To those who understand a little bit in the DNS, all the common practice on that. We used to have also high complexity attacks but low volume. So on one hand you had a volume, low complexity, and then you have a low volume with very high complexity attacks. About two years ago, December 2016, the new wave of Mirai came in and we started to face a new generation, both high volume and complex IoT. Welcome to the world of IoT. We were talking on different things. So think of that, that you have now hundreds of millions, billions of devices on different things, and someone can create a swarm that can change the activity, the behavior of the devices. And a swarm, unlike even other intelligent attacks, acts in a different way. How do you protect it?

First of all, of the trade-offs here, I want to touch between the different trade-offs of analytical classifiers with self-learning and machine learning or artificial intelligence, deep learning, machine learning, the different nuances. The issue starts with what problem are you going to solve. Is the problem complex or not? Sometimes it's not effective to apply an AI tool. More than that, AI can provide more harm than effectiveness. So let's look on that. Look at the network on traffic time series analysis. All of you know that network protocols are RFC-based protocols. Everyone has an RFC explaining the state machine of the protocol. In that case, definitely the way is to build a state machine or anomaly detection that is based on the RFC, the understanding. This is very effective for today what we do in time series analysis. On the other end, when you start having or pack information, not time series specifically, let's say logs, activity logs, user machine behavior, now can you profile a user behavior? It's not an RFC, TCP, UDP, ICMP, HTTP that we all know the state machine. User, a machine spin-off, I'm running in Amazon a machine, an EC2 machine, then comes the DNN family, deep neural networks, clearly.

So when we look at it, when you have really a way to run and characterize traffic, it means that you can build an analytic classifier. Why? Because you can write code that describes the RFC. So we have a family of analytic classifiers here. What we do in Radware, all of our data plane devices utilize self-learning algorithms that I will show you next in a second. What we do in the data plane, while when we move to the other side of the equation, when you have a lot of activities, users, machines, no way to build an RFC to describe the way they behave, you need to use data. If you refer to what we have talked like five minutes ago, the quality that you have here depends on the quality of the data, because based on the data you are going to train the model. So the data, you must be very careful to use data that actually reflects the problem that you're going to solve. And that's the entire magic. If you have good data, you need to train the model and maybe it will provide good. But if you do that, it's kind of not necessarily you need AI for that. For what we are doing, actually, obviously we have big data, you upload it to the cloud, and what we call, we use cloud factory. The cloud becomes a factory accumulating all the data, and what we are building on top of it is a layer which we call AI defense. So self-learning on the data plane and AI defenses at the cloud, which are solving different problems. So each part is solving a different problem. And on top of it, remember the issue of security is not a detection, the really tough part is the mitigation. Automatization, can you do also mitigation automatic? Remember we are talking on AI and automation. So the idea, are those technologies providing you enough information that you can build an automation and an automatic gate? So at the data plane you need to do it in zero minute, why? Because it's real-time. While you do deal with more complex patterns, you move to a trend what we call a zero day, because it depends on much more analytics and analyzers. So self-learning, analytic classifier, and cloud classifier, DNN classifier, completely two types of family. And on top of it you need to build a policy automation, a control plane that can actually organize all the activities, because there will be certain issues you need to do at the data plane, some of them in the cloud.

So let's look a little bit on the analytic classifiers. Since we have very limited time, the analytic classifiers provide today two important things, two components: they self-learn. The reason for self-learning is to be able to create smart and self-learning baselines in order to do the detection, and then you need to use other algorithms to do the auto-mitigation. Very effective for network and application attacks, namely there are three to layer seven attacks, both at the edge, network, data centers, etc. To look a little bit on top of it, the idea behind them, I'm sorry because we have very limited time, nine minutes left, so let's try to play with it. You take a vector of observation, those are the time series X1, X2, X3, can reflect different things. X1 can be packet size, TCP, you need distribution of flags of TCP, different parameters that represent the RFC or the problem that you are going to look. And what we are doing, we are building an analytic classifier. Then the classifier basically is sampling all the time and creating those baselines. Those are the self-learning baselines. It's like magic, create a multi-dimensional self-learning baseline. Baseline reflects what we call the NSA: the normal, suspicious, and attack surfaces. In every second basically you are looking to create a classifier that can point you what we call what is the current degree of attack that you observe. This is not every second, it's done by self-learning algorithms. Those are built across, and the idea, if you do it the right way, you can reduce the false positives, because always the issue of detection and characterization lies, what is the quality, how many false positives do we have. So the way we do it is completely building a self-learning automatic agent that looks at the traffic. You build the policy, you monitor, you train the parameters that you are looking, and what you are doing, you are building, running the traffic in, and all the time you are measuring based on that self-learning technologies what is the degree of attack. When the degree of the attack starts to move to the red lines, which means I see now an abnormal activity is detected, something which is abnormal, doesn't matter, there is something there, then you start the profile, the characterization. And this is the magic to build in real-time. Actually the algorithms can do it today in seconds, 18 seconds you understand and you create a signature or a fingerprint or footprint that explains the anomaly. Now in order to validate it, you will need to challenge a little bit in order to reduce the false positives. But once the positive and the challenge was validated, now you can block it while speed the traffic in. Today the silicon we are using are about up to the largest devices today in the industry are about 400 gig per box. So the state of the art of self-learning in the industry today is multiple hundreds of gigabits per second that we can actually do that kind of close loop of detection, characterization, and automated mitigation. By the end of the day, then you can add a regex extension to look very, be very specified. And guess what, you can now use the cloud with AI intelligence to feed the regex extension. So there's now a complete complex which combines both self-learning and AI technologies in that sense.

Just to give you, so you can ask me, yeah, but what are the really percentages that you can provide today? So these are real numbers that we have taken from one of our Tier 1 customers in the US, running across 40 data centers across the globe, from Hong Kong, Australia, United States, European Earth. You can see the numbers that out of, in one week, this is a weekly attack report the devices produced and built the same way I'm showing you, 130,000 attack events, 99.49 were actually automated by the device successfully. Other 663 was a second state, doesn't matter. The idea is that those are what we call combat-proven technologies, combat-proven self-learning technologies that can really solve a lot of issues, real-time issues when you run attacks against networks, data centers, applications, multiplayers.

Okay. So we were observing that side, analytic classifiers, which are fine up to a certain point because you can as a model that reflects the problem. But what about the other side? So let's look on the other side when we have really other issues or let's say things that you cannot do, you cannot classify by any analytic classifiers. So in that sense we look at the cloud in two aspects. First of all, cloud itself becomes a factory, used to do analytics. So in the cloud we are using, as we explained before, taking a lot of information on the device as other things, and we are providing new tools, malware workload protection. And then as a cloud can be seen also as a hosting place for new customers. So customers are building the private data centers but are moving the workloads to run in the cloud, building what we call native cloud data centers with a native cloud. Probably many of you have heard about microservices, running Kubernetes. This is the area for that we are using a new set of AI technologies to build a new domain of solution called cloud workload protection. And by the idea of it, and this is the thing that we are going to launch actually next week, is the way to provide a snapshot of your native cloud data center in all domains: storage, network, servers, users, database. And here you can see the different colors represent the anomalies in different zones, only on those suspicious. So the tools are used to what we call reduce the signal-to-noise ratio, because when you run a cloud you have a lot of noise, too many activities. This is a very smart filter that reduces the noise level, and based on that, using AI to gain, you can build a metro line tracking attacks. This is a metro line that provides the different stages of infiltration and exfiltration of a data breach.

So those are the things that AI is augmenting to what we were talking before, can really solve probably very well today. So let's tie all the pieces together. So when we tackle the pieces together, we say we have multiple layers, we should have multiple layers. So we have the self-learning data plane, the control and automation that control and automates the policies, obviously management and operation, and on top of it, this is the new tier that everyone looks, AI defense. AI defense augments the self-learning, it's another part of a self-learning. Together you provide data, control, management, and cloud. This is a new domain in AI, in AI plan. So when we look on all of these, basically you can open a traditional unprotected object, you start building multiple layers to protect it. So the first one will be what we have seen at the data plane, you will have a self-learning data plane will address in real-time, in seconds, the threat and mitigate in seconds activities that running against your data, against your application, against your network. And then you have other layers, member updates into regex engine from the cloud, up to you run a cloud factory to run on top of it additional feed, additional layer that can overarch a lot of information, a lot of new enrichment sources to provide a completely new view on what you need. And with that, zero seconds. We will be very happy to answer questions in our booth on mass loss, or feel free to come address us at Radware. Thank you very much.