Applied parallel coordinates for logs and network traffic attack analysis

J Comput Virol (2010) 6:1–29
DOI 10.1007/s11416-009-0127-3
ORIGINAL PAPER
Applied parallel coordinates for logs and network traffic
attack analysis
Sebastien Tricaud · Philippe Saadé
Received: 20 December 2008 / Accepted: 17 July 2009 / Published online: 27 August 2009
© Springer-Verlag France 2009
Abstract By looking on how computer security issues are
handled today, dealing with numerous and unknown events is
not easy. Events need to be normalized, abnormal behaviors
must be described and known attacks are usually signatures.
Parallel coordinates plot offers a new way to deal with such
a vast amount of events and event types: instead of working
with an alert system, an image is generated so that issues
can be visualized. By simply looking at this image, one can
see line patterns with particular color, thickness, frequency,
or convergence behavior that gives evidence of subtle data
correlation. This paper first starts with the mathematical the-
ory needed to understand the power of such a system and
later introduces the Picviz software which implements part
of it. Picviz dissects acquired data into a graph description
language to make a parallel coordinate picture of it. Its archi-
tecture and features are covered with examples of how it can
be used to discover security related issues.
Keywords Visualization · Parallel coordinates ·
Data-mining · Logs · Computer security
A picture a day keeps the doctor away.
S. Tricaud
Honeynet Project French Chapter, 69 rue Rochechouart,
75009 Paris, France
e-mail: sebastien@honeynet.org
P. Saadé (B)
Lycée la Martinière Monplaisir, Laboratoire de Mathématiques,
41, rue Antoine Lumière, 69372 Lyon Cedex 08, France
e-mail: psaade@gmail.com
1 Introduction
This paper covers how visualization techniques based on
parallel coordinate plots (abbreviated as //-coords) can
enhance the computer security area.
It is common to have thousands lines of logs a day on
a single machine. With private networks of hundreds of com-
puters over complex topologies, this really represents a huge
load of information. How can one separate the important part
of the information from the unimportant one?
To deal with that issue, administrators, most of the time,
use tools such as Prelude LML,1 OSSEC2 or similar software
that are often based on signatures. Besides signatures based
tools, they also use anomaly based tools, that are classifying
the information after a learning phase. One example is spa-
massassin,3 which does a great job at removing spam out of
our mailboxes. Over the years, these tools have proven an
indisputable efficiency.
However, something missing today is dealing with data
exactly as it is. There is often more to see than just the part
of the data having a matching threshold of signature. That’s
why computer visualization is a good choice!
Computer visualization is a neat way to see the picture of
what is really happening and can, in some cases, handle a lot
of information. As //-coords can handle multiple dimensions
and an infinity of events, it became a natural choice to write
a software being able to automate those graphs creation. This
software is called Picviz.
In the first part of this paper, we will introduce the very
basic facts about //-coords. We will explain in the most sim-
ple terms the fundamentals of //-coords as a mathematical
1 http://www.prelude-ids.org.
2 http://www.ossec.net.
3 http://spamassassin.apache.org.
123

2 S. Tricaud, P. Saadé
theory. The first important results will be given, without
assuming from the reader a heavy mathematical background.
In the second part, we will present Picviz in its overall
architecture and then in greater detail.
The last part of this article will be devoted to real-life
examples.
We will first discuss the case of giant log files of Cray sys-
tems, whose syntax was unknown to the authors and gave a
challenging playground for finding important events without
using any pre-existing signatures or tools of any kind.
Then, we will consider the particular case of Botnet
attacks. It came out that //-coords and Picviz can help in
detecting and characterizing these malicious threats. And we
will explain how in the last section of this paper.
Examples including Apache access.log file and wireless
LAN traffic monitoring will also be given.
2 A short mathematical introduction to parallel
coordinates
2.1 Cartesian and parallel point of view
Imagine one has to collect elementary events of a given type
(temperatures of all capitals of Asia, network traffic on a net-
work adapter, etc.). Let’s suppose that each given elementary
event carries N kinds of information and that N is not small
(greater than 4). Since it is not easy to plot vectors belong-
ing to a space of more than 3 dimensions in a 3 dimensional
physical space (not counting the time), it becomes necessary
to adapt the representation technique.
In an N-dimensional vector space E, one needs a basis
of N vectors. Then each vector u ∈ E corresponds to an
N-tuple of the form (x1, x2, . . . , xn). In the usual euclidean
space of dimension N, denoted RN , the canonical basis is
orthogonal, which means that axes are considered pairwise
perpendicular (Fig. 1).
This is the usual cartesian representation of R3 and it
corresponds to our everyday life experience of our ambient
space! But since it is impossible to draw more than 3 per-
pendicular axes in a 3 dimensional physical space, the idea
behind //-coords is to draw the axes side by side, all parallel
Fig. 1 Orthogonal basis in R3
Fig. 2 Four axes
Fig. 3 Four axes and a vector
to a given direction. It is then possible to draw all these axes
in a 2d plane (Fig. 2):
For example, the vector u = (0.6, 1.6, −0.8, 1.2) ∈ R4
should show up as (Fig. 3).
That point of R4 has become a polygonal line in //-coords!
At first sight, it might seem that we have lost simplicity.
Of course, on one side, it is obvious that many points will
lead to many polygonal lines, overlapping each other in a
very cumbersome manner. But on the other side, it is a fact
that certain relationships between coordinates of the point
correspond to interesting patterns in //-coords.
It is the aim of this short mathematical introduction to
present some elementary patterns that can be observed and
to prepare the reader for the definition of geometrical invari-
ants (in //-coords) of simple subspaces of RN such as lines,
planes and p-subspaces (sometimes called p-flats).
2.2 Trivial hyperplanes
The usual definition of an affine hyperplane in RN is that
it is the set of vectors −→u = (x1, x2, . . . , xN ) satisfying an
affine relationship of the form
α1x1 + α2x2 + · · · + αN xN = β
where all coefficients α1, . . . , αN and β are real numbers and
α1, . . . , αN are not zero all together.
That hyperplane is a vector hyperplane or simply
a hyperplane if β = 0 which geometrically means that it
passes through to origin O = (0, 0, . . . , 0).
For example, in the usual plane R2, the relationship
α1x1 + α2x2 = β, where (α1, α2) = (0, 0)
123

Parallel coordinates, logs and network traffic attack analysis 3
Fig. 4 A line in R2
Fig. 5 A coordinate line and a trivial line in R2
characterizes a line orthogonal to vector −→n = (α1, α2)
(Fig. 4).
Naturally, in R3, the cartesian equation
α1x1 + α2x2 + α3x3 = β, where (α1, α2, α3) = (0, 0, 0)
defines a plane having −→n = (α1, α2, α3) as normal vector.
Hyperplanes generalize these geometric objects in higher
dimension and are precisely the (N − 1)-affine subspaces of
RN .
Trivial hyperplanes are those which can be defined by
an equation of the form
xi = b
for a fixed given index i.
In such a case, that hyperplane is simply parallel to the ith
coordinate hyperplane having equation xi = 0 (Fig. 5).
In //-coordinates, such trivial hyperplanes are very easy
to recognize since one coordinate is constant (Fig. 6):
This is the first simple pattern to show up in //-coordi-
nates. For example, if one studies a set of TCP/IP packets,
one can assign to the first axis the source IP address and to
the second axis the destination port. It is then obvious that in
an attack such as DOS on port 80 (www) of a server coming
from many different machines, such a structure is going to
appear. And even with much more than two axes involved,
if one of them represents destination port, the above pattern
will still be there, and easy to notice.
Fig. 6 A trivial hyperplane in R4
Now that we have seen that trivial hyperplanes are easy
todetectin//-coordinates,wemustconsider nontrivial ones.
Are they so easy to discover? For example, let’s say you have
10000 points scattered on a fixed affine hyperplane H in R5.
Will the //-coordinates plot of these 10000 points present a
trivial pattern that can be seen at first glance? The definitive
answer is No! or, more precisely, Not yet!.
So, do not despair! In the following section, we will try to
explain in the most simple terms what can be seen and how
it can be seen using //-coordinates.
3 A detailed overview of some basic facts of parallel
coordinates
3.1 Notations and conventions
As said earlier, all //-coordinates plots are drawn on a
2-dimensional plane. To keep things simple, all the axes will
be drawn vertically. The plane on which they are drawn is
equipped with a usual cartesian coordinate system, denoted
R// = (O, −→ı , −→j )
where (−→ı , −→j ) is an orthonormal basis (Fig. 7).
A point M in that plane will have its coordinate relative
to R// given by
M = x//, y// R//
We will denote by εi the abscissa of the vertical line carrying
axis xi in R//
(xi ) : x// = εi
Definition RN
// (ε1, ε2, . . . , εN ) corresponds to R//
equipped with N vertical axes having abscissas ε1 < ε2 <
· · · < εN .
To shorten the notation, we will often simply write RN
// .
123

Fig. 7 A //-coordinate system for R4
Fig. 8 Notations
And sometimes, we will even forget about the conditions
ε1 < ε2 < · · · < εN , just for fun!
Definition If A = (a1, a2, . . . , aN ) ∈ RN , let
A
i
denote
the point
A
i R//
= (εi , ai )R//
This is simply the point attached to axis xi and by which
passes the polygonal line representing A (Fig. 8).
Definition The line joining
A
i
and
A
i + 1
will be denoted
by
A
i, i + 1 R//
=
A
i
A
i + 1
and, when needed, the segment joining these two points will
be denoted by
A
i, i + 1
Fig. 9 (L) in R2
Remark Most of the time, we will not make any difference
between the line
A
i, i + 1
and the segment
A
i, i + 1
. For
sake of simplicity, we will often draw segments and build
intersection points of such segments as if they were lines.
Remark Sometimes, we will consider lines joining points
on axes that are not consecutive. For example,
A
i, j
corre-
sponds to the line joining
A
i
and
A
j
, even if |i − j| = 1.
The general cartesian equation of such a line is
A
i, j R//
: (y// − ai )(εj − εi ) = (x// − εi )(aj − ai )
or
A
i, j R//
: y// = ai + (x// − εi )
(aj − ai )
(εj − εi )
3.2 A first glance at the case of a line L in R2
In this section, we are going to focus on a line (L) of R2
(Fig. 9).
The way such a line looks like in //-coordinates mainly
depends on the slope of L.
To get an intuitive feeling of what is going on, have a look
at the following plots of L in //-coords for different slopes
(Fig. 10):
It seems obvious that in all cases, when M(x1, x2)
describes L, the line
M
1, 2
passes by a fixed point who’s
horizontal position solely depends on the slope of L.
123

Fig. 10 Different slopes for (L) in R2
3.3 A closer look at the case of a line L in R2
Let M(x1, x2) ∈ L. Then the corresponding line in R2
// has
the following equation
M
1, 2 R//
: (y// − x1)(ε2 − ε1) = (x// − ε1)(x2 − x1)
Fig. 10 continued
• In the particular case where L is parallel to the ﬁrst diag-
onal : x2 = x1, that is, when L : x2 = x1 + β, one has
(Fig. 11)
M
1, 2
: y// = x1 + (x// − ε1)
β
(ε2 − ε1)
123

Fig. 11 (L) : x2 = x1 + β
Fig. 12 Random points
on a line L of R2
meaning that when M describes all of L, the line
M
1, 2
remains parallel to vector
1
β
ε2−ε1 R//
or
ε2 − ε1
β R//
.
• In the more general case where L : α1x1 + α2x2 = β is
not parallel to : x2 = x1, that is, when α1 + α2 = 0, it
can be easily shown that all lines
M
1, 2
have a common
point (Fig. 12).
We will denote that point by
L
1, 2
, and a straightforward
computation gives
L
1, 2 R//
=
α1ε1 + α2ε2
α1 + α2
,
β
α1 + α2 R//
Remark People used to working with projective spaces will
immediately notice that both cases can be described by the
point
L
1, 2 R//
= [α1ε1 + α2ε2 : β : α1 + α2]
in RP2
.
From now on, we will consider that any line L ⊂ R2 cor-
responds to a point
L
1, 2
in R2
// and we will not mention
anymore that in some cases that point has to be defined in
projective plane.
Remark Reciprocally, given the point
L
1, 2 R//
= (x0//,
y0//)R// ,
one recovers a cartesian equation of L
L : px1 + (1 − p)x2 = y0//, where p =
x0// − ε2
ε1 − ε2
The correspondence between L and
L
1, 2
is called by Insel-
berg the line-point duality:
L : α1x1 + α2x2 = β
↓
L
1, 2 R//
= [α1ε1 + α2ε2 : β : α1 + α2]
L
1, 2 R//
= x0//, y0//, z0//
↓
L : (x0// − ε2z0//)x1 + (ε1z0// − x0//)x2 = y0//(ε1 − ε2)
3.4 Vocabulary
The preceding point
L
1, 2 R//
will be called the level 2
invariant point of line L, or simply the invariant point
of line L, in parallel coordinates.
4 First nontrivial results in //-coordinates
4.1 A first glance at the case of a plane P in R3
As we did for the case of a line L in R2, we will start by some
simple geometrical configuration of an affine plane P in R3.
First of all, we already know that if P is a trivial plane
xi = β then random points on P show up in R3
// in the
following manner (Fig. 13).
In such a case, it is completely obvious that P is entirely
characterized by the point of convergence of the segments.
That point has coordinates (ε1, 1)R// in our example or, more
generally, (εi , β)R// in the case of P : xi = β.
We will call that point the invariant point of plane P in
parallel coordinates and denote it
P
1, 2, 3
= [εi : β : 1]
Suppose now that we are in the particular case of P : α1x1 +
α2x2 + α3x3 = β where α1 = 0.
We observe that random points on P plot in the following
way (Fig. 14).
In our example of (P) : 0x1 + 2x2 + 3x3 = 7, what hap-
pens on the first axis (x1) is uncontrolled since the cartesian
equation of P does not involve variable x1. On the other hand,
the shortened equation 2x2 + 3x3 = 7 and our initial study
of a line in R2 explain why we observe that all lines
M
2, 3
123

Fig. 13 Trivial plane
(P) : x1 = 1 in R3
Fig. 14
(P) : 0x1 + 2x2 + 3x3 = 7
(for M ∈ P) have a common point. We can even compute
the (homogeneous) coordinates of that point.
P
1, 2, 3
= [2ε2 + 3ε3 : β : 2 + 3]
In the little more general case of P : α2x2 + α3x3 = β the
invariant point would be
P
1, 2, 3
= [α2ε2 + α3ε3 : β : α2 + α3]
Now we are going to spend some time with plane P : 2x1 +
2x2 + 3x3 = 13 which offers none of the previous partic-
ularities. In a certain sense, we can consider it as a random
affine plane of R3.
If we randomly select some points on it and plot these in
parallel coordinates, we are bound to admit that, this time,
no obvious pattern structures the plot (Fig. 15).
But if we intersect P with plane x1 = λ, for different
values of λ, we notice some well known patterns! (Fig. 16).
The explanation is easy to give: if x1 is a constant, then x2
and x3 satisfy the equation 2x2 +3x3 = (13−2x1) and that’s
the equation of a line in these two variables. The invariant
point associated with this line has (homogeneous) coordi-
nates
[2ε2 + 3ε3 : 13 − 2x1 : 2 + 3]
whose abscissae is independent of the constant value of x1.
As x1 is fixed to different constant values, the level 2 invari-
ant point describes a vertical line (i.e. itself parallel to the
axes of the plot).
This simple observation is going to lead us to the defini-
tion of the level 3 invariant point associated to the plane
P.
4.2 A closer look at the case of a plane P in R3
Going one step further, we notice that the ordinate
β − α1x1
α2 + α3
of the level 2 invariant point of P∩{x1} satisfies, with variable
x1, the affine equation
α1x1 + (α2 + α3)
(β − α1x1)
(α2 + α3)
= β
and in these two variables it is simply the equation of a line.
So it should come equipped with an invariant point.
123

Fig. 15 Random points on
(P) : 2x1 + 2x2 + 3x3 = 13
Fig. 16 Different intersections
of (P) : 2x1 + 2x2 + 3x3 = 13
with x1 = constant
This is verified on the plot if we draw the lines going from
x1 on the first axis to the corresponding level 2 invariant point.
The point we have constructed using different level 2
invariant points will be called the level 3 invariant point
associated with P (Fig. 17).
As far as it is itself a level 2 invariant point, we can use
what we know about such points to compute it’s homoge-
neous coordinates. We get, after a quick computation
P
1, 2, 3
= [α1ε1 + α2ε2 + α3ε3 : β : α1 + α2 + α3]
Now we have to tell whether this point has interesting prop-
erties or not and if it is useful to help us decide whether
a given point is in P or not.
Remark The previous construction of
P
1, 2, 3
is not pos-
sible if α2 + α3 = 0 because of division by zero. Later in
this paper, we will give a different construction of the level
3 invariant point, helping us to avoid such accidents.
4.3 A different construction of the level 3 invariant point
of a plane
We study a plane P ⊂ R3 given by
P : α1x1 + α2x2 + α3x3 = β
Let A, B, C be three noncollinear points of P. The plane P
is completely determined by these three points.
As we already know, plotting A, B, C in //-coordinates does
not yield any clear geometric pattern of the kind we had in
the case of a line in R2 (Fig. 18):
What actually happens is that there still is a geometric pat-
tern or geometric organisation in R3
// when plotting points
all belonging to a fixed plane P, but it is not as obvious as
seen before.
To show this, we will use geometric constructions of
points, starting from A, B, C and not the cartesian equation.
Definition Let
AB
1, 2
be the unique point of intersection of
the lines
A
1, 2
and
B
1, 2
. This point always exists in the
123

Fig. 17 How level 3 invariant
point appears for (P) : α1x1 +
α2x2 + α3x3 = β
Fig. 18 Three points A, B, C in R3
Fig. 19 Construction of the first level 2 point
projective plane holding R3
// even if the two lines are parallel
(Fig. 19).
For such an intersection point, we use the notation
AB
1, 2
=
A
1, 2
·
B
1, 2
Proposition 1 With all preceding notations we have
Fig. 20 Construction of four level 2 points
AB
1, 2 R//
= [(b2 − a2)ε1 − (b1 − a1)ε2 : b2a1
− a2b1 : (b2 − b1) − (a2 − a1)]
Of course, in R3
//, the same construction can be done with
axes 2 and 3, and with the points B and C. For example,
BC
2, 3
=
B
2, 3
·
C
2, 3
.
If we add these points to the original plot, we obtain (Fig. 20):
For the final step of our geometric construction, we need to
draw the line passing by
AB
1, 2
and
AB
2, 3
. We will call this
line
AB
1, 2, 3
(Fig. 21):
AB
1, 2, 3 R//
=
AB
1, 2
AB
2, 3 R//
Now let
ABC
1, 2, 3
(Fig. 22) be defined by
ABC
1, 2, 3 R//
=
AB
1, 2, 3
·
BC
1, 2, 3
123

Fig. 21 Preparing for level 3 point
Fig. 22 Construction of a level 3 point
A simple and remarkable fact is given in the next proposition
Proposition 2 The three lines
AB
1, 2, 3
,
BC
1, 2, 3
and
AC
1, 2, 3
are convergent. For their common point of inter-
section, we use the following notations:
ABC
1, 2, 3 R//
=
AB
1, 2, 3
·
BC
1, 2, 3
=
BC
1, 2, 3
·
CA
1, 2, 3
=
CA
1, 2, 3
·
AB
1, 2, 3
This fact can be observed in Fig. 23. The Open Source soft-
ware Geogebra4 was used to create an animated construction
and this image.
From the above definition, it is easy to deduce the follow-
ing corollary
4 http://www.geogebra.org/.
Corollary 1 Let σ ∈ S({A, B, C}) be any permutation on
the set {A, B, C}. Then,
ABC
1, 2, 3 R//
=
σ(A)σ(B)σ(C)
1, 2, 3 R//
There are many ways to prove the first proposition. For those
not familiar with projective geometry and Desargues’ the-
orem, a direct analytical computation is possible. But if no
trick is used to simplify the computation, it might seem hard
to do it by hand. In such a case, the open source mathematical
software Sage5 can easily be used to compute the coordinates
of the level three point.
But in the end, it all boils down to a point with quite
simple coordinates: the ones we obtained in our previous
construction of the level 3 invariant point of P! And that’s a
fundamental coincidence.
Proposition 3 (Fundamental result)
ABC
1, 2, 3 R//
= [α1ε2 + α2ε2 + α3ε3 : β : α1 + α2 + α3]
in the case where (ABC) is a well defined affine plane with
cartesian equation
α1x1 + α2x2 + α3x3 = β
Remark Again, it is visible from the above expression that
ABC
1, 2, 3 R//
is invariant under permutation of the letters
A, B, C.
Remark The coordinates of
ABC
1, 2, 3 R//
only depend on the
coefficients of the cartesian equation of P and of the abscis-
sas ε1, ε2, ε3 used when building R3
//.
This motivates the following definition.
Definition Let P : α1x1 + α2x2 + α3x3 = β be a plane of
R3.
We define the level 3 point invariant of P in R3
// as the point
P
1, 2, 3 R//
= [α1ε2 + α2ε2 + α3ε3 : β : α1 + α2 + α3]
Remark It is very important to understand that P completely
determines
P
1, 2, 3 R//
but the reciprocal is false.
Obviously, the knowledge of
P
1, 2, 3 R//
is not enough to
recover the original cartesian equation of P. At least, you
need the data of a point A in P.
It can be shown that it is then enough to recover P.
5 http://www.sagemath.org.
123

Fig. 23 Constructions of level two and level three points for (ABC)
5 Geometrical proof of Proposition 2
We will begin with a short review of projective geometry.
5.1 A quick introduction to projective geometry
Projective geometry only considers points and lines and does
not use notions as distance or angle. It is based (classically)
on 4 axioms:
• By two points, there always is a line passing, which is
unique if the two points are different
• Two lines always meet
• It is possible to find 4 points such that any three of them
are never aligned
• Pappus’ theorem
Pappus’ theorem has many different expressions but an easy
one can be read from the following figure
Let (ABC) and (A B C ) be two triangles such that C ∈
(A B ) and C ∈ (AB).
Let i = (AC) ∩ (A C ), j = (BC) ∩ (B C ) and k =
(AB ) ∩ (A B).
Then i, j, k are aligned.
OneofthefirstconsequenceofPappus’theoremisthefamous
Desargues’ theorem.
Consider the following figure (Figs. 24, 25)
Desargues theorem states that if (ABC) and (A B C ) are
two triangles such that (AA ), (BB ) and (CC ) meet in a
common point O, then the points of intersection of sides of
both triangles having same names are aligned.
More precisely, if i = (AB)∩(A B ), j = (AC)∩(A C ),
k = (BC) ∩ (B C ), then i, j, k are on a same line.
By a very important duality principle in Projective
Geometry, Desargues’ theorem admits a reciprocal. This
reciprocal states that, with the preceding notations, if i, j, k
are aligned then (AA ), (BB ) and (CC ) meet in a common
point (Fig. 26).
Using this result, it is possible to prove the first proposition
concerning the level 3 invariant point.
5.2 A geometrical proof of Proposition 2
Recall that our geometric construction of
ABC
1, 2, 3
is based
on the following process.
We first compute the six points obtained by intersecting lines
in-between the same axes (Fig. 27).
Then, we pretend that the three lines
AB
1, 2, 3
,
BC
1, 2, 3
and
AC
1, 2, 3
are convergent (Fig. 28).
To prove this, we will simplify the drawing and only keep
the central axis (Fig. 29)
123

Fig. 24 Pappus’ theorem
Fig. 25 Desargues’ theorem
We notice two triangles satisfying the hypothesis of the recip-
rocal of Desargues’ theorem and therefore can conclude that
our three lines meet at a same point (Fig. 30).
6 Hyperplanes in RN
It is possible to generalize the preceding constructions and
define points associated to general hyperplanes H of RN .
Definition 6.0.1 Let H : α1x1 + α2x2 + · · · + αN xN = β
an affine hyperplane of RN .
Let RN
// (ε1, ε2, . . . , εN ) be a fixed //-coordinate system.
The//-coordinate level N pointassociatedwith H in RN
// is:
H
1, 2, . . . , N RN
//
= [α1ε1 + · · · + αN εN : β : α1 + · · · + αN ]
The next result shows that level N points can be build
recursively by a simple geometric process from four level
N − 1 points.
123

Fig. 26 Reciprocal of Desargues’ theorem
Proposition 4 For any family of N points (A1, A2, . . . , AN )
defining hyperplane H, one has:
H
1, . . . , N R//
=
A1A2 . . . AN
1, 2, . . . , N R//
where the last point is defined recursively by:
A1A2 . . . AN
1, 2, . . . , N
=
A1 . . . AN−1
1, . . . , N − 1
A1 . . . AN−1
2, . . . , N
·
A2 . . . AN
1, . . . , N − 1
A2 . . . AN
2, . . . , N
The whole process starts with the N2 points
Ai
j
.
Fig. 27 Construction of the level 3 invariant point
Remark It is very important to note that it is not an easy
computation to prove that both definitions of the level N
invariant point coincide.
To be more precise, the fact that the recursive process
produces, in the end, a point with coordinates that can be eas-
ily expressed in terms of the coefficients of some cartesian
equation of the hyperplane looks like a beautiful algebraic
results to both autors.
Fig. 28 Construction of the
level 3 invariant point
123

Fig. 29 Simplifying the figure
7 Detection of more general geometric structures
with //-coords
As we want to keep this mathematical introduction quite
elementary, our journey in the theoretical //-coords universe
will stop shortly. We hope that the very basic aspects of //-
coords are now familiar to you and that you understand better
how classical affine geometric objects can be detected.
Of course, we know that it is not enough to detect linear
structures and that in real life, the mathematical relationship
between variables can be more complex. It is therefore essen-
tial to know wether //-coordinates are able or not to deal with
such datasets.
Furthermore, it is one thing to examine a pure
mathematical object defined by a set of algebraic equations in
RN (as in real algebraic geometry), and it is another to man-
age experimental data that are supposed to obey a certain
class of mathematical laws. It is even a third story to investi-
gate computer security logs given that it is very doubtful that
any analytical relationship exists between the variables.
In a forthcoming article, we hope to report about a classi-
fier based on //-coords and pioneer work of Inselberg [6–9]
and some new ideas. This classifier, hopefully, will show how
//-coords can go much beyond our actual exposition.
8 Automating //-coords creation with Picviz
Now that the theory behind //-coords is known, and since
Picviz goal is to automate the creation of //-coords images
out of logs, this section introduces its architecture and how
it can be used to discover security issues. Picviz is not lim-
ited to computer security, however since it is a good goal to
demonstrate how powerful can be such a tool, this is what
the paper sticks to.
Because digging visually for security issues is aimed to
be very practical, Picviz presentation starts with an authen-
tication log investigation. After this short presentation, the
architecture is detailed. Once the reader knows features pro-
vided by the software, two examples are given.
Fig. 30 How we apply the
reciprocal of Desargues’
theorem
123

8.1 Picviz
Picviz6 is a software transforming acquired data into a
parallel coordinates plot image to visualize data and discover
interesting results quickly. Picviz is composed in three parts,
each facilitating the creation, tuning and interaction of //-
coords graphs (Fig. 31):
• Data acquisition: log parsers and enricher
• Console program: transforms PGDL into a svg or png
image. Unlike the graphical frontend, it does not require
graphical canvas to display the lines, it is fast and able to
generate millions of lines.
• Graphical frontend: transforms PGDL into a graphical
canvas for user interaction.
It was written because of a lack of visualization tools able
to work with a large set of data. Graphviz is very close to
how Picviz works, except that is has limitations regarding
the number of dimensions that can be handled by a directed
graph, such as when dealing with time.
8.2 Data acquisition
The data acquisition process aims to transform captured logs
into the Picviz Graph Description Language (PGDL) file
beforePicviztreatment.Inthispaper,logisusedinterchange-
ably with data to express something that is captured from one
or several machines. That being one in:
• Syslog: System and application log files. Containing at
least four variables: time, machine, application and the
logged event.
• Network: Sniffed data.
• Database: Structured information storage.
• Specific: Log file for applications not using standard log
functions.
• Other: Any other way to record events.
CSV being a common format to read and write such data,
Picviz can take it as input and will translate it into PGDL.
8.3 Console program
The Command Line Interface (CLI) allows Picviz to easily
generate //-coords graphs, including frequency analysis, fil-
tering and powerful filters. As the CLI allows to deal with
millions of events, this may be the first step prior using the
graphical frontend. The CLI is the pcv program that is being
used in this paper to show how graphs were generated.
6 This paper describes Picviz version 0.6.x.
8.4 Graphical frontend
While the Graphical frontend does not allow to deal with as
much objects as the CLI does, it still gives valuable insight
for the analyst. Live interaction is something the frontend
can provide:
• moving the mouse over the lines will actually show the
lines values being displayed,
• changing the axis order on the fly by selecting the appro-
priate value on the axis description box,
• zooming to allow line-to-line relationship finding
(Fig. 32),
• slider to move in the event time scale: moving the cursor
to the left-most will show the first event,
• brushing to find the relationships in multiple dimensions
of a given event (Fig. 33).
8.5 Understanding 10000 lines of log
Visualization is an answer to analyze a massive amount of
lines of logs. //-coords helps to organize them and see cor-
relations and issues by looking at the generated graph [3].
Each axis strictly defines a variable: logs, even those that
are unorganized, are always composed by a set of variables.
At least they are: time when the event occurred, machine
where the log comes from, service providing the log infor-
mation, facility to get the type of program that logged, and
the log itself.
The log variable is a string that varies widely based on the
application writing it and what it is trying to convey. This
variability of the string is what makes the logs disorganized.
From this string, other variables can be extracted: username,
IP address, failure or success, etc.
Log sample: PAM session
Aug 11 13:05:46 quinificated su[789]:
pam_unix(su:session): session opened
for user root by toady(uid=0)
Looking at the PAM session log, we know how the user
authenticates with the common pam_unix module. We know
that the command su was used by the user toady to authenti-
cate as root on the machine quinificated on August 11th at
1:05 p.m.. This is useful information to care about when
dealing with computer security. In this graph we clearly
identify:
• If a user sometime fails to give the correct password
• If a user logged in using a noncommon pam module or
service
• Time when users log in
123

Fig. 31 Picviz simpliﬁed architecture
Fig. 32 Graphical frontend zoom feature
Fig. 33 Graphical frontend brushing feature
Figure 34 shows the representation of the auth syslog
facility:
Analysis
What one can easily see in Fig. 34 is how many users
logged in as root on the machine: red lines means root des-
tination user. Also, the leftmost axis (time) is interesting: it
has a blank area and using the frontend we discover that no
one opened a session between 2:29 a.m. and 5:50 a.m.:
The second axis is the machine where the logs origi-
nated. Since this example is a single machine analysis, lines
converge to a single point.
Fig. 34 Picviz frontend showing pam sessions opening
Fig. 35 Zoom on time axis
Fig. 36 PAM module convergence
This third axis is the service or application that wrote the
log. We can quickly see four services (one red, three blacks,
the line at the bottom is also a connection between two axes):
moving the mouse above the red line at the service on top of
Fig. 35 shows that only the ’su’ service is used to log a user
as root. Hopefully no one logs in using gdm, kdm or login
as root.
The fourth axis is the pam module that was used to
perform the login authentication: again, as only local authen-
tication was done using the pam_unix module, lines are con-
verging. If we would have had a remote authentication, or
other modules opening the session, we would see them on
this axis (Fig. 36).
The ﬁfth and sixth axis are the user source and destination
of the logs. We have as much source logins as destination lo-
gins. On this machines, logins are both su and ssh.
As experts might know, //-coords are already used in
computer security [1] but face a problem of not being easy to
automate or with various data formats. This paper focuses on
how relevant security information can be extracted from logs,
whatever format they have, how anyone can discover attacks
123

or software malfunctions and how the analyst can then filter
and dig into data to discover high level issues. The next part
covers how Picviz was designed, its internals. After we will
see how malicious attacks can be extracted, and how it can
help you to write correlation rules.
9 Picviz architecture in detail
9.1 Picviz graph description language
It has been designed to be easy to generate and as close as
possible to the Graphviz [2] language (mostly for proper-
ties names). It is a description language for //-coords which
allows to specify all kinds of properties for a given line
(data), set each axis variable axes and give instructions to
the engine in the engine section. Also, a graph title can be
set in the header section.
Below is an example of a PGDL source which represents
a single line:
header {
title = "foobar";
}
axes {
integer axis1 [label="First"];
string axis2 [label="Second"];
}
data {
x1="123",x2="foobar" [color="red"];
}
Despite CSV being a standard for log normalization in
order to create a graph, and even though Piviz can convert
CSV into PGDL, CSV is not recommended. It is impossible
to set properties on an individual line when a specific item
is encountered. This would require an external configuration
file to be parsed at every new added line. This would greatly
decrease performances. Also, //-coords have the weakness of
hiding interesting values according to the axis order. Chang-
ing the axis order in PGDL is as simple as moving a line in
the axes section.
9.1.1 The axes section
It defines possible types you can set to each axis, as well
as setting axis properties. Labels can be set to axes with the
label property. Axes types must be one of them:
Type Range Description
timeline "00:00" - "23:59" 24h time
years "1970-01-0100:00"
-
Several years
"2023-12-3123:59"
integer 0 - 65535 Integer number
string "" - "No specific A string
limit"
short 0 - 32767 Short number
ipv4 0.0.0.0 - IP address
255.255.255.255
gold 0 - 1433 Small value
char 0 - 255 Tiny value
enum anything Enumeration
ln 0 - 65535 Log(n) function
port 0 - 65535 Special way to
display port
numbers
It is indeed possible to specify the maximum value an
axis can get as a numeric value instead of the axis type. This
would allow value 1234 to be the maximum of axis1:
axes {
1234 axis1;
...
}
Other available properties are:
• print: when set to false, removes values printing on lines.
It is usually used when an axis had too big values which
are overlapping the next axis.
• relative: when set to true place values on the axis rela-
tively to each other. Which decreases performances but
can improve the axis reading.
9.1.2 The data section
Data are written line by line, each value coma separated. Four
data entries with their relatives axes can be written like this:
data {
t="11:30",src="192.168.0.42", dport="80"
[color="red"];
t="11:33",src="10.0.0.1", dport="445";
t="11:35",src="127.0.0.1", dport="22";
t="23:12",src="213.186.33.19", dport="31337";
}
Thekey=valuepairallowstoidentifywhichaxishaswhich
value. Since axis variable type was defined in the previous
axis section, the order does not matter. For the previous data,
if one want to respect the order, the axis section would be:
axes {
timeline t;
ipv4 src;
port dport;
}
123

Changing the axis order and repeating the axes is explained
in the Grand tour section.
Every line can receive two properties: color and pen-
width, which allow to set the line color and width, respec-
tively.
Data lines are generated by scripts from various sources,
ranging from logs to network data or anything a script can
capture and transform into PGDL language data section.
This paper focuses on logs, and Perl was chosen for its con-
venience with Perl compatible regular expressions (PCRE)
built-in with the language. The next part explains how such
a script can be written to generated the PGDL.
9.2 Generating the language
PicvizdeliversasetoftoolstoautomatethePGDLgeneration
from various sources, such as apache logs, iptables logs,
tcpdump, syslog, SSH connections, …
Perl being suited language for this kind of task, it was chosen
as the default generator language. Of course, nothing prevent
other people to write generators for their favorite language.
The PGDL is generated with the Perl print function, along
with Perl pattern matching capabilities to write the data sec-
tion. The syslog to PCV takes 25 lines of code, including
lines colorization where the word ’segfault’ shows up in the
log file. Then, to use the generator, type:
perl syslog2pcv /var/log/syslog > syslog.pgdl
To help finding evilness, a Picviz::Dshield class was writ-
ten. Calling it will check if the port or IP address match with
dshield.org database:
use Picviz::Dshield;
dshield = Picviz::Dshield->new;
ret = dshield->ip_check("10.0.0.1")
It can be used to set the line color, to help seeing an event
correlated with dshield information database.
9.3 Understand the graph
9.3.1 Graphical frontend
Aside from having a good looking graph, it is good to dig into
it, and see what was generated. An interactive frontend was
written for this purpose. It is even a good example on how
Picviz library can be embeded in a Python application. The
application picviz-gui was written in Python and Trolltech
QT4 library.
The frontend provides a skillful interaction to find
relationship among variables, allows to apply filters, drag
the mouse over the lines to see the information displayed
and to see the time progression of plotted events. Real-time
capabilities are also possible, since the frontend listen to a
socket waiting for lines to be written.
The frontend has limitations: on a regular machine, more
than 10000 events makes the interface sluggish. As Picviz
was designed to deal with million of events, a console pro-
gram was written.
9.3.2 Command line interface
The pcv program is the CLI binary that takes PGDL as input,
uses the picviz library output plugins and generate the graph
using called plugin. To generate a SVG, the program can be
called like this:
pcv -Tsvg syslog.pgdl > syslog.svg
As SVG is a XML and vectorial format, it will perform
well when a few thousand of line are drawn: the operator will
be able to do actions on items, select them, grep for a certain
value, move the lines, etc.
However, with a big set of data, SVG frontend will fail
doing the rendering.
This is why a PNG capable plugin was written. Using the
cairo7 library. The plugin is named ‘pngcairo’ and can be
used like this:
pcv -Tpngcairo syslog.pgdl > syslog.png
Usually is it better to use the PNG plugin, filter data and if
needed,usethenchoosebetweentheSVGplugintouseallthe
features a vectorial image can provide or the Picviz frontend,
that is designed to deal with //-coords issues. Section 2.3.4
explains how filters can be used with Picviz.
9.3.3 Grand tour
Because choosing the right order for the right axis is one of
//-coords disadvantages, Picviz provides via the pngcairo
plugin a Grand tour capability. The Grand tour generates
as much images as pairs permutation of axes possible, the
idea is to show every possible relation among every available
axes. Plugin arguments are provided with the -A command.
So to generate a grand tour on graphs, pcv should be called
like this (Fig. 37):
pcv -Tpngcairo syslog.pcv -Agrandtour
...
File Time-Machine.png written
File Time-Service.png written
...
File Log-Machine.png written
File Log-Service.png written
7 http://www.cairographics.org.
123

Fig. 37 Syslog grand tour
An other way to change the axes order is simply changing
their order in the axes section.
axes {
integer axis1;
string axis2;
ipv4 axis3;
}
Shows in the order: axis1, axis2 and axis3, while:
axes {
ipv4 axis3;
integer axis1;
string axis2;
ipv4 axis3;
}
will show the axes in the order axis3, axis1, axis2 and then
axis3 again. Thus changing the axes–pair relationship. This
section also allows commenting to hide a given axis.
It is recommended to have the data separated from the
image and axes properties to allow easy changing, this can
be done using the @include keyword, that will include data
from a third-party file. In the case of numerous data, this is
very convenient to find relationships and allow easy changing
of axis type and order:
axes {
ipv4 axis3;
integer axis1;
string axis2;
}
data {
@include "mydata.pgdl"
}
9.3.4 Real-time
Picviz can be set up to perform real-time line drawing by
listening to a socket. Then programs can send their lines to
this socket.
When Picviz listen to a socket, it should have a template
associatedwithit.Thisisagraphwritteninatemplatederived
fromPGDLnamedPGDTforPicvizGraphDescriptionTem-
plate. The difference with a PGDL file is that a template may
not have any data, so that the program will know which vari-
ables are associated with the axes. Of course PGDT is more
convenient to express a file that is created for real-time, but
it can interchangeably be used with PGDL.
To start PCV in real-time mode, one can start on one side:
pcv -s local.sock -t samples/test1.pgdl
-Tpngcairo -o /tmp/graph.png
And on the other side, the client will simply need to write
onto this socket:
echo "t="12:00", i="100",
s="I write some stuff";"
The OSSEC HIDS8 can output its alerts to Picviz, using
the template ossec.pgdt provided with sources.
9.3.5 Filtering
To select lines one want to be displayed, Picviz provides
filters. They can be used on the real value to match a given
regular expression, line frequency, line color or position as
mapped on the axis. It is a multi-criterion filter. It is set with
the CLI or Frontend parameters.
With the CLI, they can be called like this:
pcv -Tpngcairo syslog.pcv ’your filter here’
With the frontend, filter can be called like this:
picviz-gui syslog.pcv ’your filter here’
Filter syntax is:
display type relation value on axis
number && ...
Where:
• display: show or hide, select if we hide or display the
selected value
• type: value, plot, color or freq, choose what is filtered
• relation: <, >, <=, >=, ! =, =, relation with selected
value
8 http://www.ossec.net.
123

Fig. 38 SSH scan
• value: selected value to compare data with
• on axis: text to express the axis selection
• number: axis number to filter values on
For example, to display all lines plotted under a hundred
on the second axis, one can replace your filter here by plot
< 100 on axis 2. Specific data can also be removed, such as:
pcv -Tpngcairo syslog.pcv ’value = "sudo"
on axis 2’
A percentage can be applied to avoid knowing the value
that can be filtered: ‘plot >(42% on axis 3 and plot < 20%
on axis 1’.
9.3.6 String positioning
One of the basic string algorithm displaying is to simply add
the ascii code to create a number. Among pros of using this
very naive algorithm, is to be able to display scans (strings
very close to each other coming from one source) very easily.
As for the cons there is the collision risk, but in practice this
low risk of having such events. As Picviz is very flexible, it
still offer other string alignment algorithms, using Levenstein
[10] or Hamming distance [4] from a reference string. This
still makes collision possible, but differently.
The basic algorithm highlights scans evidences, and then
one can quickly spot an issue. This way, without having any
knowledge of how the log must be read, little changes will
appear close enough to each other to grab the reader attention.
The following lines are logs taken from ssh authentication,
and appear like this:
time="05:08", source="192.168.0.42",
log="Failed keyboard-interactive/pam
for invalid user lindsey";
time="05:08", source="192.168.0.42",
Fig. 39 Same shared value
Fig. 40 Syslog heatline with virus mode
for invalid user ashlyn";
time="05:08", source="192.168.0.42",
for invalid user carly";
Figure 38 shows a generated graph from twenty lines of a
ssh scan.
On the third axis, one can clearly see the lines sweep,
showing the scan.
9.3.7 Correlations
With //-coords, several correlations are possible, as shown in
Fig. 39, where it is known for sure all events share a common
variable.
One other way to correlate is applying a line colorization
for their frequency of apparition between two axes and col-
orize the whole line, according to the highest value. Picviz
can generate graphs in this mode with the heatline rendering
plugin and its virus mode (Fig. 40).
123

Fig. 41 Picviz curves
As of today, only the svg and pngcairo handle this feature.
Picviz CLI should be called like this:
pcv -Tpngcairo syslog.pgdl -Rheatline
-Avirus > syslog.png
9.3.8 Curves
Curves is the idea of drawing arc circles instead of lines to
maybe uncover clusters more easily. While this technique
will go against all the mathematical theory described in this
paper, some people consider this. Nothing useful has been
found using this technique. It is shown to get an idea of what
one can do with //-coords. See Fig. 41.
9.3.9 Section summary
Picviz has been designed to be very flexible and let anyone
capable to generate the language, filter data and visualize
them. This can be done statically on a plain generated file, and
with the graphical interface, this is even possible in real-time.
Ofcoursetheknowledgeoflogslinesisbettertosetmoreaxis
and have more information to understand the graph. How-
ever, the naive approach is sometime enough to see scanning
activities.
Now that the Picviz architecture and features have been
explained, we will show how we can efficiently use it to
dig into logs and extract relevant information.
The next section covers how we can efficiently see attacks
from many lines of log and finally write a correlation script
in perl.
10 Understanding the unknown
This section explains how one can handle in a very short
amount of time the analysis of failures a system can have
based on their logs. The log issues coverage will not be
exhaustive but using //-coords can easily spot some.
During the latest Usenix Workshop on the Analysis of
System Logs 2008, Cray systems gave logs to analysts to see
what they could extract from them.
Cray logs are available for download in the Usenix web-
site. 9 This analysis is done on the file 0809181018.tar.gz.
10.1 Files overview
Once uncompressed, there are sixty-seven files and four main
directories, which are:
• eventlogs
• log
• SEDC_FILES
• xt
The log directory does not contain enough information: a
single 18 lines file. As Picviz is good to manage big files, the
eventlogs directory was chosen.
The eventlogs directory has two files: eventlogs and
filtered-eventlog. As the last file is smaller and from its name
seems to be filtered, it is best to run Picviz on eventlogs.
10.2 Dissecting the eventlogs file
The eventlog file looks like:
2008-09-18 04:04:54|2008-09-18
04:04:54|ec_console_log| src:::
c0-0c0s3n0|svc:::c0-0c0s3n0|
2008-09-18 04:04:54|2008-09-18 04:04:54|
ec_console_log|
src:::c0-0c0s3n0|svc:::c0-0c0s3n0|7
[?25l[1A[80C[10D[1;32mdone[m8[?25h
2008-09-18 04:04:54|2008-09-18
04:04:54|ec_console_log|
9 http://cfdr.usenix.org/data.html#cray.
123

Fig. 42 Cray eventlogs graph
src:::c0-0c0s3n0|svc:::c0-0c0s3n0|cat:
/sys/class/net/rsip333x1/type
2008-09-18 04:04:54|2008-09-18 04:04:54|
ec_console_log|
src:::c0-0c0s3n0|svc:::c0-0c0s3n0|:
No such file or directory
Bylookingthroughthislogfileonemaydiscernfivefields:
• 2008-09-18 04:04:54|2008-09-18 04:04:54: time value,
because of the lack of knowledge of Cray logs for why
there are two identical timestamps, the first is taken. The
variable timeline should be appropriate.
• ec_console_log: unknown value, which seems like an
enumeration of a few possible items. It looks like an event
category. The variable enum should be appropriate.
• src:::c0-0c0s3n0: should be the source node where the
even come from. Again, since there are a limited num-
ber of distinct resources, the variable enum should be
appropriate.
• svc:::c0-0c0s3n0: should be the destination. Just like for
sources, the enum variable should be appropriate.
• : No such file or directory: is the log itself. The string
variable is appropriate and will be put as relative with
the other strings.
BelowtheperlcodeforthegeneratortohandleCrayevents
log file:
print "axes {n";
print " timeline t [label="Time"];n";
print " enum e [label="Event"];n";
print " enum s [label="Src"];n";
print " enum d [label="Svc"];n";
print " string l [label="Log",
relative="true"];n";
print "}n";
10.3 Graphing
From the PGDL source generated and using the axes pair
frequency rendering, a graph is created in Fig. 42.
//-coords reveals the following facts:
• Time range: logs were written in a very short at very
specific time. Without being accurate, one can say all
logs are written between about 5 a.m. and 11 a.m.
• Coverage: one event (the one in the middle) covers all
sources.
• Frequency: the red lines shows that one event occurs
way more than the other, that there is a source-destination
triggering most of the events and a log occurs more than
the others.
• Differentiation: Some low-frequency logs are far away
from all the other logs. Interesting.
10.4 Filtering
As it is usually interesting to start with those events that
appear in low frequency on top of the fifth axis, a filtering is
applied:
pcv -Tpngcairo event.pcv ’show
plot > 90% on axis 5’ -a
Which gives the Fig. 43.
No need for frequency analysis here, since it was done
in the first graph. This graph outlines that a very few logs
were written, that they all come from the same class of event
(second axis) and a few machines among all that were in the
first graph make them.
Logs generating those lines are:
123

Fig. 43 Cray ﬁltered eventlogs graph
10.4.1 Log #1
Bootdata ok (command line is earlyprintk
=rcal0 load_ramdisk=1
ramdisk_size=80000 console=ttyL0
bootnodeip=192.168.0.1 bootproto=ssip
bootpath=/rr/current rootfs=nfs-shared
root=/dev/sda1 pci=lastbus=3
oops=panic elevator=noop xtrel=2.1.33HD)
Interesting keywords: earlyprintk, oops=panic.
10.4.2 Log #2
Bootdata ok (command line is earlyprintk=
rcal0 load_ramdisk=1 CMD LINE
[earlyprintk=rcal0 load_ramdisk=1
ramdisk_size=80000 console=ttyL0
bootnodeip=192.168.0.1 bootproto=ssip
bootpath=/rr/current
rootfs=nfs-shared root=/dev/sda1
pci=lastbus=3 oops=panic elevator=noop
xtrel=2.1.33HD]
Interesting keywords: earlyprintk x2, CMD LINE,
oops=panic.
10.4.3 Log #3
Bootdata ok (command line is earlyprintk=
rcal0 load_ramdisk=1 Kernel
command line: earlyprintk=rcal0
load_ramdisk=1 ramdisk_size=80000
console=ttyL0 bootnodeip=192.168.0.1
bootproto=ssip bootpath=/rr/current
rootfs=nfs-shared root=/dev/sda1
pci=lastbus=3 oops=panic elevator=noop
xtrel=2.1.33HD
Interesting keywords: earlyprintk x2, Kernel command
line, oops=panic.
10.4.4 Log #4
Lustre: garIBlus-OST0005-osc-
ffff8103f3fab800: Connection to service
garIBlus-OST0005 via nid 19@ptl was
lost; in progress operations using this
service will wait for recovery
to complete.
Interesting keywords: Connection to service … was lost,
wait for recovery to complete.
As one can see, Picviz successfully help to trace events
on the most used node based on both frequency analysis and
high values on the log axis. Because printk are the Linux
kernel print function handling anything the kernel wants to
print in usual abnormal situations, it is worth taking a look
at it.
pcv -Tpngcairo -Wpcre event.pcv ’show
value = ".*printk.*" on axis 5 -a
Which gives the Fig. 44, which outlines on the ﬁfth axis
lines previously seen, and others, appearing at about 10%
on the same axis, that were mixed with the other logs. By
looking at those logs value, there is:
10.4.5 Log #1
printk: 64 messages suppressed
10.4.6 Log #2
printk: 336 messages suppressed
123

Fig. 44 Cray eventlogs graph with printk
Which reveals a set of repeated problems.
Of course, knowing exactly what occurred would require
more work and knowledge of Cray systems. However using
//-coords helped to understand easily what was happening
on those machines, the way they log, the number of sources
and what seems to be destinations.
In this section, the way //-coords were used may be looked
as if one would have use the grep tool. Actually, the first
generated picture and its frequency analysis helped to target
data that could be most likely the source of an issue. Then,
because lines were drawn far away from all the other and in
low frequency, it went fairly easy to spot and focus on the
printk.
11 Botnet analysis
Botnets attacks consists of spreading a malware to usually
vulnerable Windows systems to get a maximum of machines.
This big amount of machines allows an attacker to spread
spams and make denial of services (DoS) attacks. Botnets
can be controlled from IRC commands or similar custom
protocol. Also files can be sent through various nodes using
a peer to peer protocol [5].
When facing a Botnet attack, one will see connections
from a big load of different source IP addresses. Making it
hard to block. Most of the time, empirical tricks are being
used to kill an ongoing DoS attack, such as renaming the
target web file, etc.
This section does not intend to explain details of a botnet
attack, but how such an attack looks like and how it can be
understood to help defending against it.
Because Picviz is good with such a number of events, a
Botnet attack was captured and Argus was used to clear out
network flows, it was used to generate a //-coords graph.
In this example, two types of images based on the same
data were generated. Only the color varies:
• Figure 45: UDP traffic is in blue, TCP traffic in yellow
• Figure 46: Frequency analysis per pair-axes wise
Based on those two figures, it is easy to outline:
• The attack happens on two times: the first starting at about
1 p.m. and stopping 30 min after. The second attack is
bigger: starting at about 2 p.m. and stopping 3h latter.
• The TCP flow starts during the second attack and last
during a short period of time.
• The frequency analysis is more interesting: one can eas-
ily see that several IP addresses knocked the same source
port. Also, one single source IP receives most of the traf-
fic. This is because it is the target machine.
• When superposing the two pictures, most of the generated
traffic happens at the same time TCP traffic happens.
As an emergency counter-measure, two solutions can be
considered: blocking any UDP traffic that is not DNS
requests, NTP traffic or any critical service required by the
network. Also, since the source port is shared for most of
IP sources, it seems this value was hardcoded. Blocking this
specific port may help resolving the problem for the time the
administrator can take to understand more about the attack.
In all, even though the botnet understanding requires some
time, using //-coords to see obvious facts can help having
more time to investigate deeper.
12 Apache access.log analysis
Trying to find evil requests within 2years, the first thing one
may do is to use known-patterns against the log. Picviz allow-
123

Fig. 45 TCP versus UDP
botnet analysis
Fig. 46 Botnet frequency
analysis
ing one to search for what she/he was not even looking for,
going step-by-step into the attack discovery is what this sec-
tion shows. The first thing to do is to generate a graph from
the logs as they are (Fig. 47).
The frequency analysis mask is applied on every graph,
using an axis-to-axis relationship. In 2years, only two HTTP
protocols were being used: HTTP/1.0 and HTTP/1.1.
To clear things a little, three axes one may find important
are selected: IP address, User Agent and Request. The result
of this selection can be seen in Fig. 48. What is made obvious
in this graph is the red line that goes from a very specific IP
address to a User Agent, and the same User Agent performing
the GET request. This is actually the Google bot.
There are also interesting HTTP requests made, not the
common GET. Looking at the activity made on the other
requests is also interesting.
In Fig. 49, the same axes selection is performed, but the
GET request type is removed. The following line was used
to generate this image:
$ pcv -Tpngcairo -Rheatline access
.log.pgdl -r ’value != "GET" on axis 3’
Putting the request made next to the IP address shows an
interesting pattern (Fig. 50). All lines going from 1 to 2 look
like request being made from the same machine. The way
strings are placed place points on the axis according to the
123

Fig. 47 First full access.log
Fig. 48 Axes selection in
access.log
string length. This shows requests that are bigger and bigger.
Indeed, this is a Google bot fetching any available pages. The
red line shows that this is event occurs frequently, show how
a bot can be active (Fig. 51).
Picviz can also ﬁlter on events frequency, this allows sort
events regarding how frequent they appear (ﬁltering with
‘freq < 0.0005’).
13 WLAN images
Current technology involves Wireless when connecting to
the internet. While this paper does not intend to explain WEP
cracking such as in [11], the two images in Figs. 52 and 53
show some obvious patterns: one looks like normal, and the
other looks like there is some sort of broadcast happening.
While understanding the attack requires how WEP crack-
ing happens, several kind of people can deal with this prob-
lem: administrators can give the picture to analyst, who will
then look for the relationships and tag the source emitting
the broadcast.
What the Fig. 53 shows is the initialization phase of WEP
cracking
Even though it is fairly easy to write a simple script able to
detect the broadcast, the visualization and the multi-dimen-
123

Fig. 49 Axes selection in
access.log
Fig. 50 Pattern in access.log
Fig. 51 Low frequency events
from access.log
123

Fig. 52 WLAN without any attack
Fig. 53 WLAN initialization vector generated on the fly before a WEP key cracking
sion aspect of //-coords allows to spot what looks obvious,
but would not have been without using this technique.
14 Conclusion
This paper explained a disruptive way of responding to com-
puter security related events using //-coords. This was exclu-
sively run from system logs and network traffic. It can of
course be extended to other parts of a machine, such as sys-
tem calls or application behavior.
However, this paper neither covered all theoretical aspects
nor all practical usage of //-coordinates. Much remains to be
done, proved and experienced in both directions.
First of all, on the theoretical background:
• one should decide whether //-coordinates are efficient or
not in analysing randomly generated points of mathe-
matically well defined object.
• one should precisely explain how and why //-coordinates
can be of any interest on datasets not coming from any
mathematical definition (such as one encounters in com-
puter security).
On the practical background of computer security, we
proved how efficient //-coordinates images can be to help
security administrators to find rare and unexpected events,
as much as understand more complex mechanisms hidden in
huge amounts of information.
We hope to investigate soon on the two questions above
and report clear explanations (more or less based on some
already existing literature).
Picviz offers a very efficient tool to get a quick and faith-
ful look into gathered datasets: frequency is taken into
account (axis by axis or on all axes together), offering a
transversal information correlation which proved not to be
obvious.
123

The generated image can be very technical and has its
learning curve to be able to tune and help to get an efficient
image one can easily find things in. However, even when
reacting as a naive person, generating a picture a day can
keep the doctor away and help to discover tendencies before
it is too late. The kind of images Picviz can generate can be
read by several kind of people and thus improves the network
security reaction at several layers.
The future of Picviz will be to automate visually correla-
tions and go further than //-coords to make a better represen-
tation of axes correlations.
References
1. Conti, G., Abdullah, K.: Passive visual fingerprinting of network
attack tools. In: VizSEC/DMSEC ’04: Proceedings of the 2004
ACM Workshop on Visualization and Data Mining for Computer
Security, pp. 45–54. ACM, New York, NY, USA (2004)
2. Gansner, E.R., Koutsofios, E., North, S.C., Phong Vo, K.: A
technique for drawing directed graphs. IEEE Trans. Softw.
Eng. 19, 214–230 (1993)
3. Grinstein, G., Mihalisin, T., Hinterberger, H., Inselberg, A.: Visu-
alizing multidimensional (multivariate) data and relations. In:
VIS ’94: Proceedings of the Conference on Visualization ’94,
pp. 404–409. IEEE Computer Society Press, Los Alamitos, CA,
USA (1994)
4. Hamming, R.: Error detecting and error correcting codes. Bell Syst.
Tech. J. 29, 147–160 (1950)
5. Holz, T., Steiner, M., Dahl, F., Biersack, E., Freiling, F.: Measure-
ments and mitigation of peer-to-peer-based botnets: a case study
on storm worm. In: LEET’08: Proceedings of the 1st Usenix Work-
shop on Large-Scale Exploits and Emergent Threats, pp. 1–9. USE-
NIX Association, Berkeley, CA, USA (2008)
6. Inselberg, A., Avidan, T.: The automated multidimensional detec-
tive. In: INFOVIS ’99: Proceedings of the 1999 IEEE Sympo-
siumonInformationVisualization,p.112.IEEEComputerSociety,
Washington, DC, USA (1999)
7. Inselberg, A., Avidan, T.: Classification and visualization for
high-dimensional data. In: KDD ’00: Proceedings of the Sixth
ACM SIGKDD International Conference on Knowledge Discov-
ery and Data Mining, pp. 370–374. ACM, New York, NY, USA
(2000)
8. Inselberg, A., Dimsdale, B.: Parallel coordinates for visualiz-
ing multi-dimensional geometry. In: CG International ’87 on
Computer Graphics 1987, pp. 25–44. Springer, New York, NY,
USA (1987)
9. Inselberg, A., Dimsdale, B.: Multidimensional lines ii: proximity
and applications. SIAM J. Appl. Math. 54(2), 578–596 (1994)
10. Levenshtein, V.I.: Binary codes capable of correcting deletions,
insertions, and reversals. Tech. Rep. 8 (1966)
11. Stubblefield, A., Ioannidis, J., Rubin, A.D.: A key recovery attack
on the 802.11b wired equivalent privacy protocol (wep). ACM
Trans. Inf. Syst. Secur. 7(2), 319–332 (2004)
123

Applied parallel coordinates for logs and network traffic attack analysis

Recommended

More Related Content

What's hot (20)

Similar to Applied parallel coordinates for logs and network traffic attack analysis (20)

More from UltraUploader (20)

Applied parallel coordinates for logs and network traffic attack analysis