Specifically, I work in the Information Systems Technology Program, which is a, it's a very applied approach to engineering, and specifically information technology systems.
We're currently at the University of Central Florida developing a undergraduate security program that I'm also, I'm the one that's mainly involved in developing that, and that's actually what my second talk a little bit later is on today.
But this is the more, maybe more interesting, more technical one.
My past research includes the correlation of distributed, distributed network sensors.
If you have a, if you have a large scale network, you have intrusion detection systems, you have firewalls, you have other things that are recording information saying what's going on on the network.
Trying to correlate that information together, to bring it all together and trying to see the big picture of what's going on, rather than getting these little kind of half views of what's really occurring out on the network.
And also I'm working on some projects that involve combined data sharing networks for law enforcement in the state of Florida.
You know, this little blurb ad about the University of Central Florida, it's down in Orlando, Florida. It's about, we've got about 42,000 students there, about 5,500 of which are in engineering.
We're located about 13 miles east of Orlando, Florida.
As I said, the Information Systems Technology Program, this is a really, it's an applied engineering degree.
It's very hands-on, very real world, very current tools, not so much the traditional theoretical approach to engineering.
A heavy concentration of hands-on learning in the program.
A little bit about the background of what led me to this current idea.
There was a previous research that I worked on, I gave a couple of presentations back at DEF CON 9 and 10.
And this was on using Bayesian multiple hypothesis to try and analyze reports from these distributed intrusion detection systems and other types of network sensors.
With the idea of trying to determine if what you were seeing were correlated events, if this was an initial attack and a follow-up,
or if the two events you were seeing in two different places might be two systems communicating who shouldn't be communicating together.
And to try and use multiple sensors to distribute over large distributed networks to try and do this.
The basic idea behind what I was doing with that was taking target tracking methods that came out of radar target tracking systems,
trying to track real objects in the sky with various radar systems,
and taking that same approach and treating all of our sensors that we have on our network, whether it's an IDS, a firewall, a router,
even a human sat there making observations about what's going on and trying to automate the correlation of that together.
And this wasn't intended to totally throw the human analyst out of the loop,
but just to kind of act as a sort of a first level filtering system to throw away a lot of the information that you don't need
in order to come to a conclusion about what's going on.
With the idea that we were going to kind of move this more towards a point of view,
instead of looking at this one section of the network or this one part of what was going on that's visible to this one sensor,
like, you know, you've got your firewall sat at the, say you've got it sat at the border of your network.
It sees everything that's coming in and out of the border.
But it doesn't see stuff that's entirely internal to the network.
You might have an intrusion detection system sat on one segment of the network that sees what's going on there,
but it doesn't see what's going on on another segment of the network.
And to take all these little pictures, pieces of information,
and to put them together to get a more complete view of what someone was doing while they were attacking the network,
and did this through using radar tracking algorithms,
specifically Bayesian multiple hypothesis tracking,
and making an analogy to the world of computer security,
where our firewalls, our IDS, our humans looking at what's going on on the network,
anything that records information is a sensor.
That's like a radar system.
The actions that the person out there trying to break into the network are doing,
that's the target, and that's the trace they leave behind.
They're trying to put these traces together.
And we're working in a situation where we've got to do this in real time.
We've got incomplete data.
We've got inaccurate data, potentially, about what's going on,
depending on how your sensors are operating.
Just a little bit more about the system.
How this works, this idea of this multiple hypothesis tracking,
is as events occur on the network,
as an event is when a sensor recognizes that something's occurring.
When something actually pops up on the sensor, when it gives a report,
when you've got an IDS that tells you, hey, I think I see something.
When you've got a firewall that decides it's not going to allow this packet
or it sees something unusual.
When you've got your sysadmin on your network saying, hey, you know,
something weird's going on here.
As these events arrive, they're analyzed.
And say, for example, we're sat there,
and one of our systems sees a port scan happening out there.
We build a hypothesis about what we think's going on.
In this case, if we've got a single scan coming in from the outside world,
hey, you know, that's all the information we have.
That's the only hypothesis we can come to, is that that's what's going on.
We're seeing a port scan right now.
Really, I guess, actually, technically, you can make two hypotheses at this point.
You can say, hey, there's a port scan going on,
or alternatively, my IDS isn't working properly,
and this is a false positive.
Then, as other events occur, say you see a buffer overflow attack
against a machine following that.
Well, what could be going on now?
We could make the assumption that the port scan is being followed up
by a buffer overflow.
Someone scanned a machine for a vulnerability,
then they attacked that particular vulnerability on it.
Or, potentially, these could be two unrelated events,
that the port scan could be part of one attack,
the buffer overflow could be part of a different attack.
On the system.
And what we were trying to do with this system
was trying to automate the process of figuring these out.
Now, this is a simple example.
This is a pretty trivial case.
In this particular case, if the port scan is a scan against,
you know, I don't know, against port 80,
and the buffer overflow is a buffer overflow against an IIS server,
chances are those two things are related.
If they're not, if the port scan doesn't match up
with what the buffer overflow was, chances are they're not related.
But, you can extrapolate from this idea
into much more complicated systems,
and much more complicated series of events,
of trying to piece together the pieces
and figure out whether or not all of these events
and all of these things that we see occurring across the network
are related together, are they not related together,
are they likely to be something real,
are they likely to be a false positive
caused by a imperfect sensor,
or imperfect IDS, or firewall, or human in the loop.
Actually, when I was working on this,
one of the most interesting sources of data that I had to work with
came from the Capture the Flag games here at DEF CON.
Luckily, the SHMOO group captures all of that data
that goes on during the Capture the Flag game.
They capture all of that network traffic,
and that provided a great playground to work in,
to try and see if that information could be correlated together.
And this little graph I got up here,
I mean, this is kind of just pulled right out of context.
These are some slides I grabbed just to give a bit of background
about the previous research in this.
Basically, what this was showing is,
as you let the number of hypotheses that you consider increase,
in that previous example I said,
excuse me, I said that we had the two hypotheses
either these two events are unrelated,
or these two events are part of a chain of a single attack.
Well, we let those, we say these are our two hypotheses,
and then we look at the evidence that's going on,
and we evaluate them, and we find out which hypothesis we think
is currently the best one we have available.
Then, we wait for more information to come in.
We'll have a certain confidence level,
and how confident do we believe in this hypothesis?
As more information comes in,
we'll build more hypotheses,
and we re-evaluate the likelihood of our existing hypotheses.
Eventually, at least in theory,
this will eventually converge to a point
where one hypothesis is clearly the winning hypothesis.
Now, the problem with, I said that's in theory,
the problem with that is,
you have to consider every possible hypothesis in order to do that,
and the number of hypotheses grows at a super exponential rate.
So, it's practically impossible to do that.
There's a lot of work on the pruning of them,
deciding which ones you want to keep,
which ones you want to throw away.
And what this graph is just showing here was,
along the bottom is a number of hypotheses
that were allowed to remain in the system,
going from, well, at the bottom one,
all the way up to a thousand,
and the accuracy of how well it correlated
the different attacks together.
And it never quite got up to 100%,
but it got up into about 93, 94% accuracy
on the capture-the-flag data,
being able to piece together what were related attacks.
And now, of course, this was actually correlated
against the baseline was done by hand,
taking me, I and a few others that helped me,
took the data over the capture-the-flag game
over about a, I think it was about an eight-hour period,
and by hand figured out what was related information,
what were related attacks,
and then built a system that could do,
to automate that process.
So this thing, it worked.
This was a, it was a pretty decent idea.
It worked.
It didn't necessarily work as well as we had hoped initially.
If you've got everything on a same network,
or related network,
where you're likely to see lots of related events,
it works well.
If you've got this distributed over large networks,
where you're likely to see many more unrelated events
than related events, it goes, the accuracy goes down a bit.
Or at least the amount of work you have to do
to get good accuracy goes way up.
It's more accurate.
It's more accurate to say.
So that's what it had done in the past.
And this got me thinking about,
so this was a system that had developed
to try and detect related information on a network.
And then, of course, you know, once you can do that,
you always think about the next step.
And once the next step, once you can detect something,
you figure out how to hide it even better.
Right?
You know, it's always kind of that one-upmanship game
that we're playing.
So it got me thinking, how do we defeat a system
that works like that?
How could you avoid detection on a network
that was well-observed, well-monitored,
and where the information was well-correlated,
where you couldn't rely on the guy sat over on this part
of the network not knowing what the guy sat over
on this other part of the network was doing,
or seeing, or knew about what was going on?
If you assume that there was complete coverage of the network,
how could you hide what was going on on there?
And I'm not necessarily, I'm not talking about so much the detection
of a specific attack, like someone does a buffer overflow
against a web server.
They do, you know, whatever other type of initial attack
to gain access to the system.
But assuming they're already on there.
They've got onto the system, and they want
to be communicating back and forth.
They've got their remote machine they're sat at.
They've got access to a machine on the network.
They want to communicate back and forth,
but they don't want to be detected while doing it.
How could you go about that and avoid detection
on a well-monitored network?
So you have to look at this from the point of view
of where do we have the choices about what we do?
When you're talking about the initial attack,
you don't have a whole lot of choices.
If you've got a, you know, I guess, I mean, I guess,
unless you get lucky and you see a system out there
that's got a ton of vulnerabilities on it,
then you've got some choices about what to do.
But assume that you have limited choices
of how a system is broken into.
There's limited ways into the system to begin with.
You don't have many choices there.
So any tool that's written to initially gain access
to a system usually has a fairly limited lifespan.
At least it's limited lifespan in the range
of people who actually care about security.
You always find some bozo out there who never updates anything,
never puts any security on their system.
You can always get into them.
But talking about the people who are serious about it,
you have a limited lifespan with any sort of tool
for initially getting in there.
However, once you have access to a machine
and all you're worried about doing is communicating
with that machine, you've got a lot of choices.
You know, we're not stuck with one type of communication protocol.
We're not stuck with any traditional communication protocol
that's out there whatsoever.
The trick to this is figuring out which ones work
and which ones won't be seen.
So here's the question.
How do you communicate on a network
without letting anyone know you're doing it?
First things first.
How do they detect what's going on?
How does someone see what you're doing on a network?
You've got IDS.
You've got firewalls.
You've got observers.
What's going on on the network?
In general, I pick the IDS and I say software or human
because whatever's looking at your network,
whether it's automated or whether it's a person,
they're trying to determine
if someone's intruding on your network.
You get two basic forms of this.
And this applies...
I think the IDS system, in at least one sense,
is the most general case.
And you can kind of find derivative cases
that cover the same concept for firewalls and routers
and everything else out there.
There's basically two forms of detection out there.
Anomaly and signature detection.
Signature detection works on known attacks, known events.
You know, if you know exactly what you're looking for,
you know when you see this packet go by, it's something bad,
you want to throw up an alert about it.
Hey, that's great, you can catch it every time.
It'll never get by.
You know, you've got that one nailed.
The only way to avoid that type of detection is
to use either an unknown method
or something perhaps that changes constantly,
and never use the same method twice.
The anomaly detection, it doesn't detect misuse of the system.
It doesn't detect something malicious.
It just detects something unusual.
So it's not necessarily really great at trying to pick
up something bad happening on your network,
but it'll pick up something unusual.
And it's really, if the signature detection doesn't work,
that's the only chance you have left,
is looking for something unusual.
If you don't know what to look for,
look for things that stand out.
So we can't really avoid the signature detection.
If you've got a way of communicating,
or a way of doing anything on the network,
and someone can write a signature to detect that, that's it.
You can't do anything with it.
It's got a limited lifespan.
Eventually it's going to be seen.
So the only alternative is to use something that you can't write
a signature for, something that changes constantly,
quote-unquote random, or at least appears random
to anything that's observing the network.
And we have to be able to do this without looking unusual.
And that's the key.
Not only do we have to do something,
it's kind of a catch-22 in a way,
because we've got to do something that's very unusual,
you know, this kind of random changing communication protocol,
but we've got to do it without looking unusual.
You can't stand out on there.
So the first stage of this is to detect the baseline of the network.
What does the traffic on the network look like to begin with?
What is occurring on the network?
What happens out there all the time?
And this can be done passively.
You can sit back and discover what type of traffic is allowed
on the network, what does normal traffic look like,
and kind of build yourself a map of both what is going on
on the network, what the structure of the network is,
and what's allowed in and out of this network.
This is not a new concept.
It's just this general idea of doing this.
This is done all the time.
This is how many tools and other fun little things are written
and developed.
But the key thing is that this is what, generally,
this is what people do this process.
Someone looks at a network or you think
about what's going on on the network.
You know, a good example of this is a reverse WWW shell.
That's an old tool.
The idea of this was that you've got a computer inside the network.
You're not allowed to make connections from the outside to it.
You know, the firewall is going to drop any inbound connections to it.
So you put a piece of software on that computer inside the network
that at certain intervals, perhaps at random intervals,
it looks like it's going out to a web browser.
It goes and makes a web request.
Except that what that web request is doing is coming to your computer
on the outside, getting the next command that you've told it to do.
It carries it back as the HTTP response to the target computer,
to the one that you got access to,
and then it carries out that particular command.
It's masqueraded looking like the computer
on the inside is just surfing the web.
That's a great way of hiding it, unless someone knows to look for it.
Once someone knows to look for it, if you stick with it too long,
they're going to look for it, they're going to see it,
it's going to stand out.
Similarly, carrying data commands and ICMP echo response packets,
same sort of idea.
What I'm trying to do with this project is to automate this process
and make it run on the fly and have it adapt itself
to whatever network it's currently working on.
Look at what's going on on the network, decide what to do on the network,
decide what will work there, and then carry it out.
Some aspects of network traffic are going to have a low entropy,
meaning a very low randomness.
They're going to be very kind of, the same thing is always going to happen.
Certain communication protocols are only going to go to certain computers.
DNS requests are only going to go to your DNS servers.
They're not going to be jumping around to every computer on your network.
Incoming web requests are probably only going to come to your web server.
They're not going to come to every computer on your network.
Other things are going to look like this.
They're going to be much more random.
Maybe outgoing web requests, you know, people, different people circling the web.
Outgoing FTP, all sorts of things.
Other things are going to be very random.
There's going to be a lot more entropy in that.
And entropy is the key to hiding information.
The more entropy you have on the system,
the more information you can hide in that.
This is like if you look at, if you look at a system
like a steganography, for example, you know,
you're trying to hide information in something else.
If you've got a bitmap, a bitmap has lots and lots
of redundant information in it versus a JPEG image.
A JPEG has taken a lot of that redundant information out.
That's why a JPEG is smaller than a bitmap.
You can hide a lot more information in a bitmap than you can in a JPEG.
Higher entropy, you can hide more information inside it.
So if we find what has the highest entropy on that network,
what communication channels have the highest degree of randomness,
we can then pick what we want to work in.
And we've got all sorts of, all sorts of things that we can work with here.
We can work with the timing of the packets, we can work with packets,
sending out packets that have bad checksums, malformed checksums,
all sorts of different things that we can play with.
We don't have to use a SSH or Telnet or some other standard type
of communication format.
We can send crafted packets where the information is carried in all sorts
of different little pieces of information inside that packet.
The key is, we just have to make sure it doesn't look unusual
when that packet goes out.
So the overall concept here is we detect what the network baseline
condition is.
On the specific network that we're looking at, what does normal look like?
What doesn't look unusual?
And then select from that the potential communication channels.
Determine which ones have highly random information on the network.
Prune it down.
Initially, we can select from anything.
We can craft a packet that does anything we want.
We're going to look at what happens on the network, figure out what's allowed,
limit it to that, and then limit it to what has a lot of entropy within that.
And then we have to somehow communicate this method to the receiver on the outside.
We have to get the information of what communication scheme we're going to use.
We have to get that to the receiver outside and then maintain a, you know,
I put almost undetectable presence on the network.
You never should think you're undetectable.
You should never think anything's absolute because it never is.
That's how you always get yourself into trouble.
But we can monitor and update the network conditions to stay hidden.
If the network conditions change,
we can change the scheme we're using to communicate.
On the network, if we just sit there and listen, what can we learn
by listening on the network?
There's all sorts of things that we can find out by just sitting there
and listening, not doing anything active, not doing anything to be discovered.
We can figure out what types of packets are allowed inside the network,
what types of packets are allowed outside the network.
We can figure out the architecture of the network.
Where are the switches?
Where are the routers?
Where are other machines on this network?
And then look at what types of packets are flying around.
What types of packets does the machine that I'm on send and receive?
What types of packets do other machines out there send and receive?
And from this, get an idea of what is allowed to move around this network,
what is moving around this network, and what we can inject into the data flowing
around without being, without sticking out like a sore thumb
and being immediately identified.
So if we narrow down these choices, we start with the ability to send any type of packet,
you know, send it to any port whatsoever.
We start paring it down based on what's available,
figure out what has the highest entropy.
Look at this, look at the type of packets based
on the protocol the ports are using.
And look at the entropy based on how often do these packets fly?
How different is the data in these packets whenever they go by?
What different machines do these types of packets go to?
And then pick a channel that we know we can get from point A to point B.
We can get from inside the network we've got access to the machine on,
out of that network over to the place where we want to get the information to,
or back and forth, vice versa, either way.
We can figure out what can get into one, out of one network, into another,
and then back the other way.
It doesn't necessarily have to be a symmetric communication channel.
The packets that you send out, the packets that you send going out one way,
don't have to look a heck of a lot like the packets coming back in the other way.
As long as they're getting to the right segment on the network,
as long as they're getting down to the right branch off of your switch
so that it can be picked up by a, by a network card in promiscuous mode,
as long as you can get down that far and you know what to look for,
each guy on each end knows what to look for,
you don't have to have it even be a symmetric communication channel between them.
You can have essentially two different channels going back and forth.
One of the big challenges with this system is starting the conversation.
If you've got a piece of software running on the inside of this network,
and it's going to start communicating with you on the outside,
and you don't know what particular communication scheme it's going to be using,
there's some difficulties in getting, in getting that first packet out of there,
that first piece of information to the outside,
without having it be something that can easily be written as a signature.
Because it's hard to, it's hard to change things or inject randomness
into the first bit of communication.
There's some, there's some ways that we can limit how signature vulnerable that first bit
of communication is, but that's one of, that is, that's one of the points that needs,
that presents some challenge and needs some work on in this.
Having the communication software itself remain undetected
on the remote machine, not talking about what the software is doing,
but actually having the software itself avoid detection.
Of course, that's another challenge, but that's a completely different story.
That's, whether it's something that's using a polymorphic code that is,
that changes from one installation to another,
or even changes itself constantly.
There are ways around that, but it is a different,
but it is another challenge in there.
The third one that's actually, I think, a fairly interesting one,
is the problem on a low entropy network.
What do you do when there's not a whole lot
of randomness going on on the network?
You know, if there's not a lot of randomness in the packets,
how do we stay hidden in it?
We have to ensure that every packet that we send
out doesn't significantly affect the statistical baseline of the network,
so that an anomaly detection system sat
on the network will not recognize this as something unusual.
It will not stand out.
It won't push the statistical analysis of the network beyond the limits,
beyond the thresholds that we currently have,
and it will avoid becoming an anomaly.
This is actually the area right now that I'm doing, that I'm concentrating on,
trying to figure out the best methods for communicating and picking channels
and picking well-hidden channels on a low entropy network.
So, just as a summary in the system, you know, what's real, what's not real.
So far, the allowed protocols discovery,
the network architecture discovery,
the statistical profiling of the network,
those are working prototypes, those modules,
those parts of the system are working prototypes.
Next stages that we're moving towards are the selection
of the communication scheme, implementation of the communication,
and the adaptation to changing network conditions.
I hope that by this time next year, this will actually be a working package
that I can give maybe a bit better presentation and demonstration
of what it's actually doing.
But this is the current idea.
Anybody that has any interest in this or any ideas or feedback
or wants to tell me why I'm crazy for thinking this, feel free.
There's my contact information.
And that's it for about a 25-minute presentation.
There's one question.
That is a great question.
I'll take that just, I'll try and answer one more real quick.
He asked, is there any concept of trying
to actually change the baseline of the network
so that you can do something that used to look unusual,
make it not look unusual anymore?
That is, it's not one of the first level items on there.
That's a pretty advanced way of working it.
But actually changing, injecting stuff in slow enough that the system,
the anomaly detection system adapts to it and thinks that's normal,
that is one of the long-term ideas on there.
That's a great idea to put into this.
Thank you.