1. User Profiling For Host Based Anomaly
Intrusion Detection In Windows NT
Debapriyay Mukhopadhyay
Satyajit Banerjee
2. Definition of IDS: Intrusion is defined as the set of
unauthorized activities that violate the security policy
of the system and intrusion detection is the act of
tracing those unauthorized users or activities on the
system.
• Two kinds of IDS:
1) Misuse Detection:- Previous attacks are captured
in attack signatures and this approach looks for
any of these known signatures in the data under
test.
2) Anomaly Detection:- Data that strongly deviates
from the normal behavioral profile are considered
as intrusive. So, mechanism involves learning the
normal behavioral profile of an user/system.
3. Motivation:
1) Prior work on IDS have mainly targeted UNIX machines.
But, majority of world’s computer while is running
WINDOWS OS.
2) A major fraction of intrusive activities is actually
launched from the inside host machines.
Problem Definition:
1) In this paper, we have tried to address the problem of host
based anomaly intrusion detection running Windows OS.
2) Problem can be seen as of learning the “normal behavior”
of an user and then scoring new activities against this model
to identify malicious insiders.
4. Issues
How to model “normal behavior” of an user is a
highly non-trivial problem.
How to ensure a significant coverage of the space of
user’s “normal behavior” – as otherwise there will be
increase in false alarms.
How to utilize the model characterizing “normal
behavior” of an user to detect anomaly intrusions
from an inside host.
5. What we have achieved?
• We have identified and categorized data that are
truly reflective of user’s normal behavior.
• We have taken a User Profiling based approach to
learn and model the “normal behavior” of an user.
• Bayesian Network has been used to profile an user
and also to detect host based anomaly intrusions.
6. Source Data and Feature Selection
• System Processes : - set of processes or services that
starts running when system starts up. These system
processes provide us with a top level profile of an user.
• Application Processes :- launched by the user shell
explorer.exe. One application (user ) process is
launched by another application (user) process.
Exploiting this dependency a DAG can be learnt.
• Window Title Bars :- capture a huge amount of
information related to user’s behavior. Per process
visible window titles can be text mined to gain
valuable information.
e.g. – iexplorer.exe can be related to one’s browsing
profile.
7. Source Data and Feature Selection
• Application Usage Profile: capturing how a user browses
through the different features of an application. For each
application, we need to track both user key strokes and
mouse click events. A nearly related concept is Program
profiling.
• For each user and for each session, the following features
can also be collected.
i) max. number of instances of each application in each user
session;
ii) average time spent on each instance of this application
(normalized by session length);
iii) percentage of the session length being spent on this
application;
iv) average waiting time for an instance of an application
being active (normalized by session length).
8. User Profiling
• Bayesian Network – used to capture the mutual
influence of different domain variables on target
attributes. Its an effective tool to be applied for
reasoning in uncertain situations.
• Categories 1 and 2 data both have a kind of causal
relationship between themselves in a sense that one
process has generated the other.
• Each process is considered as a domain variable and
“normal behavior” as target attribute.
• Detection of intrusion is done by evaluating
Prob(Normal | Evidences), by evidence we mean the
set of domain variables that are true at the time of
evaluation.
9. Learning the Bayesian Network
• Each process exe corresponds to a node in the DAG
and also as a random variable of the underlying
probability model.
• Exploit the parent-child relationship to construct the
DAG.
• For each random variable N, and for each distinct
state S of values of its parents, count the frequency of
N happening in association with S.
• Calculate Prob (N | S) – entries of the Conditional
Probability Table.
For root nodes, these conditional probabilities are
simply the a priori probabilities.
12. Inferencing
• Polytree algorithm is not applicable – we can have
more than one path between two nodes.
• We apply Junction Tree algorithm for inferencing
and calculate the following.
• P1 = Prob (Normal| Evidence of category 1).
• P2 = Prob (Normal| Evidence of category 2).
• If P1 < T1 and P2 < T2, then the data can be a
case of intrusion.
• T1 and T2 are pre-determined thresholds for
Category 1 and 2 data respectively.
13. Conclusions
• This is a work in progress.
• We have identified five categories of data, but only
have provided means of how to use the first two
categories of data.
• Different types of data can be used hierarchically
or parallelly to help in detecting an anomaly
intrusion.
• We have planned to use Probabilistic Temporal
Network to unify temporal information of (5) with
the atemporal information of (1 or 2).