Welcome to NVIDIA's AI podcast, where we explore the expanding world of artificial intelligence with the people who are developing and applying AI
On this segment of the podcast, we're going to talk about one of the trickier problems we face, cybersecurity.
It used to be that cybersecurity was something that people paid attention to in the shall we say less frequented recesses of our workplaces and homes.
In other words, we didn't do much about it, but not anymore.
We all know about Russians and hacking and the election and the Democratic National Committee.
There's been DDoS attacks recently that we all experienced when our favorite online services were taken offline for a weekend.
So there's an increasing understanding that we all have a stake in finding better cybersecurity as our lives and our data get more and more intertwined.
Our guest, Dr. Eli David, is an expert in computational intelligence and the CTO of Deep Instinct.
And he believes AI has a huge role to play in fending off the bad guys.
Eli, welcome. You are Skyping in from Tel Aviv, Israel, correct?
That's correct. Eli, how did you end up in this space?
You're an academic. How did cybersecurity and the application of AI to it become...
Part of your life and your life's work. That's correct.
I'm an academic most of my work The work has been done in the academic world, supervised many students, over 30 publications.
And when we thought about founding Deep Instinct three years ago, We knew that we would like to apply deep learning, but to a kind of problem that on one hand the impact would be huge, And on the other hand, there is no direct competitor doing that.
There are companies applying deep learning to computer vision, to speech recognition, to text understanding, Google, Facebook, Microsoft doing all great jobs.
But looking at cybersecurity, we see on one hand, huge problem.
And on the other hand, nobody's really obtaining good results.
And we have deep learning that can potentially revolutionize the field.
So that was the premises on which we founded the company.
Bringing deep learning to a new domain that has not been applied previously and hopefully obtaining the same huge improvements that have been obtained by deep learning in vision, speech and text.
As a computer science, how did you get into deep learning and machine learning to begin with?
Like you said, some time ago, It wasn't working very well and it must not have been a whole lot of fun to be banging your head against a deep learning wall or a machine learning wall when it wasn't working.
You have no idea how correct you are. My passion ever since I was in elementary school was to create a brain of my own, just to copy a brain.
So when I started the computer science, naturally I was attracted to artificial neural networks, trying to understand how our brain works and create an artificial copy of that.
So 10 to 15 years ago, neural networks was considered a complete waste of time. that only a few academic groups in the world were wasting their time and money doing that and nothing good would ever come out of it.
We could train neural networks back then, but only shallow neural nets with a few layers and a few thousand synapses, connectors. cute and good toy results but nothing really exciting.
Since three years ago, we see this revolution of deep neural networks that led into all these great results that we see today.
And this revolution is due to two facts.
The first is improvement in the algorithmic side. we understand neural networks much better, we have new methods that allow us to train deep neural nets, but even more importantly, the advent of GPUs or more accurately, the use of GPUs not only for playing Counter Strike for our children, but for running neural networks. and we see that we get up to a hundred times speed up.
If I give you just an example of the kind of things that today we can do and we couldn't have done otherwise, What a deep instinct we train our deep learning brain on hundreds of millions of legitimate and malicious files, we run it on clusters of GPU machines.
It takes us about 24 hours one day to train the brain.
Had we not run it on GPUs, it would have taken us more than three months.
100 times slower. So these are the kind of things that we couldn't have done several years ago and today we can do. at a high level, how do you think about cybersecurity?
You know, in my head, at least, it's sort of good guys and bad guys or good women and bad women battling it out.
But it's... That's just a sort of cartoonish way to think about it.
To use sort of Sherlock Holmes' words, what is the game and how is it played between predator and prey?
Actually, I like the cartoonish analogy and I will say that it is exactly like Tom and Jerry. the cat and the mouse.
With the difference that in this case, Jerry, the mouse always has the upper hand.
We have many cybersecurity solutions But despite the existence of so many cybersecurity solutions, there are about 1 million new malware created every single day.
And a large portion of them go undetected by even the best cybersecurity solutions.
That is, it is extremely easy to mutate an existing malware, turn it into a zero day malware that nobody detects, and render all current solutions irrelevant.
So it's a cat and mouse game, the mouse has the upper hand, and that's why we have such a big problem here.
So in this analogy, there's like you say, there's a million new cats coming online every day and we may or may not even see them.
That's right. We used to be protected by anti-viruses.
Antiviruses are good for remembering currently existing and previously existing malware.
They don't even... provide basic support for detecting those one million new malware that are created or rather mutated every single day.
And that has led to the new genre of cybersecurity solutions, which are called APT protection solutions.
Those solutions that use advanced AI, machine learning, dynamic analysis, sandboxing, and they try to detect as many as possible of these new malware.
So again, it sounds like, and this is in some discussions we've had about AI generally, is that It's a data problem.
And in this expanding world of data, There's no person or no group of people who could ever kind of track these things manually or as a human day to day but that AI systems can do a better job of just kind of absorbing and thinking about or looking at all these changes that occur every day.
When looking at the application of AI in cybersecurity, It is easy to say a decade gap.
What do I mean by decade gap? If you look at the AI in itself, not cybersecurity, more conventional areas of AI, In the 1980s, The best AI solutions were expert systems.
In 1990s, they were heuristic based. In 2000s, they were traditional machine learning based.
I mean, 2010s, the best AI solutions are deep learning based.
Now, if you look at cybersecurity, we see exactly the same trend, but a decade later.
In the 1990s, These cybersecurity solutions were expert system, just signature-based.
In the 2000s, they were heuristic-based, again a decade gap.
And in 2010, as we speak, everyone's talking about machine learning, traditional machine learning.
And there are currently just one company doing deep learning.
And we expect in 2020s, in cybersecurity, everyone will start talking and doing deep learning.
What happens is that the more data you have, the more challenging it becomes for humans, for manual or semi-manual methods to be able to deal with the complexity.
And we have to select more and more automated methods.
That's why we move from heuristics to machine learning, from machine learning to deep learning, et cetera.
Well, I want to get into the Internet of Things because it seems like that that complexity is only getting more complex now.
Pardon me for saying that. Because there's more and more devices connected to networks.
But tell us first, then, how deep learning can be applied to the problem of cybersecurity. security.
I understand that I have to have a strong password and two-factor authentication, etc., etc., but how does deep learning help?
As we said, we have about 1 million new models created every single day, and the vast majority of them are just small mutations over previously existing ones.
And even those Supposedly brand new families of malware, even the nation state attacks, even they are far from being brand new.
They're just 10 to 20% different. than previously existing malware.
An analogy is as if I give you a picture of a cat or a dog and then we just mutate or remove five to ten percent of the pixels and nobody would recognize that there is a cat or a dog in the image.
Sounds ridiculous, but this is what's happening in cybersecurity.
Now, wherever deep learning has been applied, in computer vision, in speech recognition, in text understanding, it has always been applied to the raw data.
The biggest difference between deep learning and traditional machine learning is that wherever traditional machine learning is applied, you always need to do feature extraction.
If the problem is, for example, face recognition, you cannot fit in the raw pixels into your machine learning module.
You need to bring a human expert, image processing expert, who would analyze the problem domain and tell you that the most important features are distance between pupils, distance between nose and mouth, proportions of the face, etc. a few tens or most a few hundred features.
This is very problematic in traditional machine learning. because when you're converting those millions of pixels in the raw data into small vector features, you're inevitably throwing away most of the data.
Deep learning is the first and currently only family of methods within machine learning that's capable of operating directly on the raw data.
When you apply deep learning to computer vision problems, the input is just raw pixels.
You don't do any image processing. would apply deep learning to text understanding, many times the inputs are just the wrong characters, not even words, completely agnostic of the language.
And similarly, wherever deep learning is applied it always operates on the raw data and furthermore you see that the more data you have the better deep learning modules learn.
And the gap between their accuracies and those of traditional machine learning methods becomes wider and wider. improvement of deep learning over traditional machine learning and other alternatives. in the areas where we're used to seeing half a percent or one percent improvement a year.
Suddenly we see 20, 30 percent improvement.
The deep learning is the greatest, the largest leap in performance in the history of computer science and AI.
So if in all those domains it provides such a big improvement, For the same reasons, it should work in cybersecurity.
That was, by the way, the assumption based on which we founded our company three years ago.
What you describe deep learning applied to image recognition or to language, it can know the characters, it can know... you know, just absorb the pixels and then recognize it as something, right?
As a word, as a phrase, as a something. And are you saying that the same application, it can recognize an AI or a deep learning system trained to, you know, recognize cyber threats can recognize a new threat without ever having seen it before.
Exactly. After all, in order to detect a cat or a dog The deep learning brain doesn't need to see all the cats in the world or all the dogs in the world. it's sufficient for it to see a large representative sample of dogs or cats to understand the concept and detect all of them.
In our case, we train our deep learning brain on hundreds of millions of malicious and legitimate files, different file formats, different operating systems, And we don't do any preprocessing.
The input to our deep learning module is just the raw bytes comprising the files.
In other words, we view malicious and legitimate files just as if they were pictures of cats and dogs.
But the only difference is that instead of a bunch of pixels, they contain a bunch of bytes.
So all the principles of deep learning apply here.
And the results we obtain are again similar to results in computer vision, about 20% to 30% higher detection rate.
There we detect cats, here we detect malware.
And so how does it work in practice? It's sitting on the edges of a network, constantly monitoring.
Does everything kind of funnel through this deep learning system?
You know, if I'm a company, I want to, you know, make sure that I'm secure, where does this interaction take place and kind of how does the alarm get set off?
Another benefit of deep learning in this context is that it is agnostic to file format or even the operating system.
This allows us in both agent-based and agent-less contexts.
In the agent-based context, we deploy our solution on any endpoint or mobile device that you have.
Any laptop, desktop, server or mobile device.
After all, All the knowledge that we have is in that neural network, in that deep learning brain, and we can put it locally on each device that we would like to protect.
We can also run it agent-less, that is connected to the gateway of your network or as a cloud service.
Basically, deep learning provides you with that versatility that you cannot obtain by any other methods.
For example, those companies who use traditional machine learning and they have a solution, let's say, for Windows operating system, they cannot use it in Android because the features will be completely different.
But deep learning can equally work well in Windows and Linux and Android because it's agnostic to the domain.
I see, so it's sort of multilingual. If you build these tools, if sort of the deep learning, the deep neural nets are getting better and better at fending off or recognizing malware.
Isn't cybersecurity always an arms race?
Can't the bad guys and the bad women, for that matter, do the reverse?
Can't they bring these tools to bear on? and kind of probing our own security and our own systems?
Using deep learning, we obtain much higher detection rates in comparison to alternatives.
However, our detection rates are not 100% and I don't expect us or anyone to ever be able to provide full protection.
So the idea here is unfortunately not to provide a bulletproof 100% protection but to make it much more difficult for the attackers and hackers to attack the system.
Now, currently, any attacker can take a well-known malware that is detected by everyone and just mutate it slightly in a few minutes, and nobody will detect their attack.
It's extremely easy to do that. with deep learning we are not making it possible for anyone to create such mutations even after considerable amount of time it will still be detected.
It is possible, of course, of the attackers that they would still manage to evade detection by everyone. but that will be much more difficult.
So our goal, and I think the whole goal of cybersecurity industry, is not to make attacks completely impossible, but to make them more and more difficult.
I see. So it's never, I was going to ask, I was hoping that we could relax finally, but...
It doesn't sound like that's the case. I mean, how do you, whether it's for individuals or for companies, how do you advise them to kind of approach this?
What stance do we need to take now and going forward?
Until a few years ago, until recently, the large corporations, those with lots of sensitive data, even they looked at cybersecurity and we're looking at the silver bullet.
What kind of solution can I install to be protected?
Now nobody's living in that illusion. Most of our customers right now, they have several solutions.
They have, for example, an antivirus. which is protecting them against currently existing and previously existing threats, and they augment that with another solution such as Deep Instinct, to protect against new threats, new malware.
They have additional layers for network analysis, firewalls, I guess, phishing.
So we see more and more a growing trend towards having many, many layers.
You can view them as many layers of fences and walls that make it more difficult for the attackers to infiltrate, to expand in the network, to exfiltrate data.
That is what I also recommend from individuals up to the largest Fortune 500 companies. to try to, first of all, find the best solutions in each domain and use layers of defense.
So what about our own human behaviors? I mean, a lot of cybersecurity and where we fall down is because As humans, we get a notice and there's a phishing attack or something and we click on it or we give them our information.
And it usually falls outside of our normal patterns of behavior.
Can AI help us with detecting and sort of saving us from ourselves?
There have been some attempts, actually many attempts at creating an anomaly normal detection and behavior analysis solution that look at your normal behavior, try to learn what is your normal patterns, normal behavior, and then find anomalies and alerts regarding them.
Previous attempts were not so successful because of the huge amount of false positive, false alarms that they triggered.
And it is not surprising because previous methods used standard machine learning, which is much less sophisticated than the kind of deep learning and AI that we have nowadays, which is not surprising that the results that we used to obtained until a few years ago were not so successful.
I can tell you that internally we are also working on anomaly detection solution based on deep learning It is still not commercially available, we expect to have it available only later this year, but even there we see that deep learning is providing substantially higher accuracy and much lower false positives in comparison to previous methods.
So for the same reasons that we say that deep learning provides much better detections in the context of malicious file detection and attack detection, for the same reasons it is doing much better in anomaly detection and in the next few years I expect that many solutions would be based on deep learning for that purpose.
Well, let's go back to the beginning and talk about cats and mice and spin this forward. how does AI change our experience in cybersecurity?
And is it something that, you know, we'll always be operating in the background and hopefully it'll just disappear and we'll go about our day?
Or do you think we'll always be aware of those cats out there?
We will have more and more of this kind of attacks.
It was 1 million new malware two years ago.
Today, the number would, I guess, be much more than 2 million.
However, the trend that we see until recently, most of these advanced protection solutions we're based on doing detection only that is they're not preventing the attack they're just providing you with an alert usually after the attack has already started, propagated, many of the damage has been inflicted,
And so the system admin's life was, and still is, quite miserable.
You always have to chase these alerts and try to extinguish the fire here and there.
Today we see a growing trend towards doing prevention and not only detection.
That we expect advanced solutions, deep learning based ones, to be able to take care of the attack by themselves even before it has started doing full prevention. which should make life much easier for us and for the system administrators, even though the number of attacks would keep increasing our lives will become easier.
So that is the optimistic trend that we see.
Do you feel vindicated now or all the deep learning people who people said we're wasting their time and we're just building toys?
Are they having a moment of sort of satisfaction these days?
Of course, of course. We're now the happiest people in artificial intelligence.
For many years we've worked in that. Everyone, all my colleagues in our computer science department said that I'm wasting time, nothing would come out of it.
And now we're doing the most exciting stuff in AI.
We meddle into everything. any domain that we would like without understanding anything about it, we just apply deep learning and we get 20-30% improvement over the results obtained by people who have devoted their entire lives to specific domains.
So we are the happiest people now in artificial intelligence.
Well, the happiest people in artificial intelligence.
Sounds like a good place to be. Dr. Eli David, thank you so much.
This has been great. Thank you very much, Michael.
PostScript here. If you're listening to this podcast on iTunes, Google Play Music, or whatever your favorite podcast player is, do us a favor and leave us a review.
Thanks for listening.
