DESIGN AND IMPLEMENTATION OF A HOST-BASED AND EVENT-BASED
A THESIS PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF SCIENCE
UNIVERSITY OF FLORIDA
Dedicated to my Mom and Dad
I would like to express my gratitude to Dr. Richard Newman for his constant
encouragement and the opportunity to work on this research project. His innovative ideas
and encouragement have made this work interesting and challenging.
I would also like to thank Dr. Randy Chow and Dr. Doug Dankel for agreeing to
serve on my committee.
Many thanks go also to Amit Gandre and Matt Wilson for their support and great
friendship. My very special thanks go to Yuan Li for being a constant source of
encouragement and support throughout.
I thank my parents for their inspiration and support throughout my academic
I sincerely thank all the people who have helped and supported me directly and
indirectly in the course of this thesis work.
TABLE OF CONTENTS
A C K N O W L E D G M E N T S .............................................................................................. iii
LIST O F TA B LE S .......................................................................................... .... ....... vi
LIST O F FIG U RE S .......................................................................................... ....... vii
ABSTRACT ............................... .............. viii
1 INTRODUCTION........................ .........1.....
1.1 D enial of Service A attacks ......................................................................... ............. 1
1.2 Intrusion Detection System s ............................................. ....... .... .... ..... 4
1.3 Knowledge-based versus Behavior-based Intrusion Detection..............................7
1.4 Host-based versus Network-based Intrusion Detection .........................................8
1.5 O organization of the Thesis ............................. ........... ................... .............. 10
2 BACKGROUND AND PREVIOUS WORK ......................... ................. 11
2.1 AAFID (Autonomous Agent for Intrusion Detection) ...................................... 11
2.2 EMERALD (Event Monitoring Enabling Responses to Anomalous Live
D disturbances) .................................................................................................. 14
2.3 GrIDS (Graph-Based Intrusion Detection System) .............................................. 17
2.4 GIDEM (Generic Intrusion Detection Model).............................. 18
3 DESIGN ................... ................... .........21
3.1 Detector Components: Data Collector ........................................ ..............22
3.2 Detector Components: Data Buffer.................................................................... 23
3.3 Detector Components: Data Operator.........................................................25
3.4 Detector Components: Data Container ........................................ ..............26
3.5 Detection Strategy .................................... .......................... ........... 27
3 .6 S u m m a ry .............. ..... ............. .............................................................. 2 9
4 IM PLE M E N TA TIO N ......... ........................................................ .........................32
4.1 Initialization of the D etector ........................................... ........................... 32
4.2 Data Collecting in Near Real Time .......................... ................................... 33
4.3 Implementation of Data Assembly and Compression ............ .....................34
4.4 Implementation of Creating the Data Buffer ................................................. 36
4.5 Implementation of Successful Login Detection.................................. ..............38
4.6 Implementation of Multi-thread System Manipulation and Concurrency Control .41
4.7 Sum m ary ................................... ................................... ......... 44
5 TE STIN G .................................. .............................. ........ ........ 45
6 CONCLUSIONS AND FUTURE WORK ....................................... ..............50
6.1 C conclusions ............................................................... ... .... ........ 50
6 .2 F utu re W ork ...................................... ........................................... . 50
B IO G R A PH ICA L SK ETCH ........................................................... .............. 55
LIST OF TABLES
3.1 Syslogd facility and priority........................................................................ 22
4.1 Sam ple actions of syslog.conf ................................ ............................. ............ 33
4.2 The form at of data entry .................................................. ................................ .. 34
4.3 The description for flag bit................................................................ .............. 35
4.4 Type bits for different login ............................................................ .............. 35
LIST OF FIGURES
1.1 T he IE T F m odel ............................................................... ............................... . . 6
2.1 Physical layout of an AAFID system ............................................ 12
2.2 Logical representation of the same AAFID system ........... ....................... 12
2.3 EMERALD monitor architecture .............. ............... ........................ 15
2.4 The beginning of a worm graph, and a more extensive view of the same worm........... 17
2.5 A architecture of G ID EM .. ....... ....................................................... .............. 19
3.1 The detector com ponents .......... .................. ........ .................... .............. 21
3.2 Architecture of the data buffer .............. .................................. ............................ 24
3.3 Architecture of the data operator ............. .. ......... ...................... 26
3.4 Structure of H ost-based D etector...................... ............... ................. ............. 30
4.1 H washing tree of data buffer................................................................ .............. 37
4.2 Data structure of naming and timing lists...... .......... .......................... .............. 39
5.1 Performance of the deletion of the data buffer .................................................. 46
5.2 Performance of the searching on data buffer........................................ ............. 46
5.3 Performance of the insertion on data buffer............................... ......... .............. 47
5.4 M emory Utilization ............................................... .............. .......... 48
Abstract of Thesis Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Master of Science
DESIGN AND IMPLEMENTATION OF A HOST-BASED AND EVENT-BASED
Chairman: Dr. Richard E. Newman
Major Department: Computer and Information Sciences and Engineering
In the recent decade, network system and e-commerce models can go to every
"corner" of our world. While people enjoy such unprecedented convenience and
happiness, hackers unceasingly "crack" the security of the network. In particular, Denial
of Service (DoS) attacks have become a very concerning issue of network security.
Research on network intrusion detection systems has been widely undertaken, including
the development of several feasible systems such as AAFID, EMERALD and GrIDS.
This thesis seeks to design and implement a host-based and event-based detector,
which is a part of the intrusion detection system in CONS lab. The designed detector can
provide an upper level controller, such as a coordinator, with the capacity to start, stop or
reconfigure the system. The detector conducts a fast detection and response based on its
tree-like data buffer. An aging policy is applied to increase the durable detection
capacity and usage time of the fast detection data buffer. Besides detecting the failure of
login events by the host, the analysis and detection of successful login events draws the
designed detector's interest. A combination of two detection results can enhance the
intrusion detection possibilities and accuracy. A compressed log file created and
maintained by the detector makes it possible to recover a detection system and launch a
post-event polling analysis, if needed. An object-oriented language makes the detector
more generic and scalable for further development.
While the world stepped into the new millennium without Y2K complaints,
people started enjoying a fast growth of network and E-business glory. However, in early
February, 2000, a handful of internet sites operated by such respected Internet electronic
commerce companies as Yahoo, CNN.com, eBay, E-Trade, Buy.com, Amazon.com and
ZDNET.com were shut down one by one for several hours. The attacks, which are so-
called denial-of-service (DoS) attacks, may have directly cost these organizations tens of
millions of dollars in lost sales. The damage to reputations and credibility may be more
difficult to gauge . Consequentially, a big question mark always keeps returning to
people: "Is the web secure?"
1.1 Denial of Service Attacks
The Internet is rapidly evolving as the "Information Super Highway" as academic,
business and government organizations as well as individual users deploy a proliferation
of local area networks (LANs) and wide area networks (WANs). However, this
increasing popularity of the Internet comes with inherent security risks . Ideally, every
computer system should be completely secure. Unfortunately, it is not possible in
today's environment. The 1998 Computer Crime and Security Survey conducted by the
Computer Security Institute (CSI) and Federal Bureau of Investigation (FBI) reports a
continuing growth in U.S. computer incidents . Intrusion activities seek to
compromise integrity, security, confidentiality and availability of a computer system,
including illegally accessing internal information, breaking into computers externally,
inserting erroneous information into files and flooding the network thereby reducing its
effective channel capacity.
A major well-known type of attack is the denial-of-service (DoS) attack, which is
characterized by an explicit attempt to prevent legitimate users of a service from using
that service. In a typical connection, the user sends a message asking the server to
authenticate it. The server returns the authentication approval to the user. The user
acknowledges this approval and then is allowed to use the services of the server. In a
denial of service attack, the attacker sends several authentication requests to the server,
filling it up. All requests have false return addresses, so the server cannot find the user
when it tries to send the authentication approval. The server waits, sometimes more than
one minute, before closing the connection. When it does close the connection, the
attacker sends a new batch of forged requests, and the process begins again--tying up the
service indefinitely . Therefore, the server cannot provide the service to the one who
truly needs it.
Denial of service attacks come in a variety of forms and aim at a variety of
services. There are three basic types of attack: consumption of scarce, limited or non-
renewable resources; destruction or alteration of configuration information; physical
destruction or alteration of network components .
Distributed DoS (DDoS) attacks are a new variation on the theme of denial of
service, and they pose a serious threat to any Internet-based enterprise, regardless of
infrastructure redundancy or robustness. The first DDoS attack in the wild was reported
in the middle of 1999. During the week of February 7th through 11th, 2000, it emerged
as a new major category of attack on the Internet. Their danger stems from their
simplicity: the tools that are necessary to set up and initiate the attack are easy to obtain
and implement, and the victim can experience hours or even days of service interruptions.
Therefore, any organization that uses the Internet must take immediate actions. Unlike
DoS removing a given resource from service, such as a server or network, without
permission, DDoS attacks--in which multiple systems generate the attack on a single
target--are the next logical step. Distributing the workload across hundreds or even
thousands of systems is one of the best ways to accomplish an intensive DDoS attack. It
not only increases the impact, but also makes stopping the attack--much less identifying
the attacker's true source--much more difficult .
The study of security in computer networks is a rapidly growing area of interest
since late 1980s. Controlling access to a computing system is the first choice to reduce
the risk of penetration threats. Several intrusion prevention techniques can be applied to
enforce network security. Authentication can be used to verify the identity of a user.
Different access control policies limit different classes of users' rights to access to data
and resources. Security policies are intended to preserve the privacy and integrity of
every user. Firewalls and cryptography can be used to block and avoid illegitimate
accesses. Good software specification, coding and testing policies can reduce
introduction of intrusion holes during development . With all the above techniques, it
is still not possible to prevent completely security threats in open data networks.
Monitoring and recording activities of users and networks will be necessary.
1.2 Intrusion Detection Systems
Intrusion detection is defined as "the problem of identifying individuals who are
using a computer system without authorization and those who have legitimate access to
the system but are abusing their privileges. (i.e. the 'insider threat')" . An intrusion
detection system is a computer system (possibly a combination of software and hardware)
that performs intrusion detection. Most intrusion detection systems try to perform their
task in real time .
Desirable characteristics of an intrusion detection system include the following
1. Continuous running: An Intrusion detection system should run under
minimal human supervision.
2. Fault tolerance: When the system crashes, the intrusion detection system
must have the capability to resume its operation and with no need to
rebuild its knowledge base.
3. Resist subversion: An intrusion detection system must have the capability
to monitor itself and prevent itself from being attacked by intruders.
4. Minimal overhead: A well-designed intrusion detection system should
ensure a minimal system load.
5. Configurable: An intrusion detection system should be configurable.
Therefore, it can be deployed on different hosts.
6. Adaptable: An intrusion detection system should be able to adapt to
changes in both global and local activities.
As the number of systems to be monitored increases and the chances of attacks
increase we also consider the following characteristics as desirable :
1. Scalable: An intrusion detection system should have the ability to increase
the number of hosts it monitors by adding new components without any
2. Graceful degradation of service: If some components of the intrusion
detection system stop working for any reason, the rest of them should be
affected as little as possible.
3. Dynamic reconfiguration: Changes in intrusion detection configuration
should be done independently and invisibly to other components. They
should not require restarting the whole intrusion detection system.
Intrusion detection systems usually include three components: sensors, analyzers
and a user interface. Sensors are responsible for collecting data. The input of a sensor
may be network packets, log files and system call traces. Sensors collect and forward
this information to an analyzer. The analyzer receives input from sensors or its peer
analyzers. It determines if an intrusion has occurred and generates an alarm when an
intrusion occurs. The analyzer may provide guidance about what actions to take as a
result of the intrusion. The user interface allows a user to view the output from the system
or to control the behavior of the system. A generic requirement is summarized in the
IETF (as shown in Figure 1.1) .
Intrusion detection systems can be classified by a variety of characteristics.
Based on the way their components distributed, the intrusion detection can be
distinguished as centralized or distributed intrusion detection systems. When the data
analysis is performed in a fixed number of independent locations and the location of the
collection components are not considered, the system is recognized as centralized
intrusion detection systems, e.g. IDES, IDIOT, NADIR and NSM. When the data
analysis is performed in a number of locations proportional to the number of hosts that
are being monitored and only the locations and number of the data analysis components
need to be considered, the system is recognized as distributed intrusion detection systems,
e.g. DIDS, GrIDS, EMERALD and AAFID .
Administrator Security Policy
Figure 1.1 The IETF model 
The different characteristics of analyzers divide intrusion detection systems into
two categories: knowledge- and behavior-based intrusion detection systems. These are
discussed further below. The different response models of the intrusion detection to
attacks categorize intrusion detection systems into two models: passive and active
models. The passive system merely generates alarms while the active one takes more
energetic actions in response to the attacks. Depending on the audit source location,
intrusion detection systems can be classified as host- and network-based intrusion
detection systems. Usage frequency distinguishes among intrusion detection systems
based on whether they have real-time continuous monitoring capabilities or they run
periodically . The next two sections introduces knowledge-based versus behavior-
based and host-based versus network-based intrusion detection systems in detail.
1.3 Knowledge-based versus Behavior-based Intrusion Detection
Intrusion detection systems have different strategies for the interaction with the
input information. There are two trends in intrusion detection: knowledge-based and
behavior-based. Knowledge-based intrusion detection is based on accumulation of
signatures from specific attacks and system vulnerabilities and is referred to as misuse
detection or detection by appearance. Once it detects any attempt to exploit a known
vulnerability, an alarm is triggered. Activities not explicitly recognized as attacks are
considered acceptable. Knowledge-based techniques have good detection accuracy
because of their low false alarm rates. However, their detection completeness is affected
by the difficulty of collecting the necessary information on the known attacks and
updating them with new vulnerabilities and environments. Expert systems, signature
analysis, Petri-nets and state-transition analysis can be used to implement knowledge-
based intrusion detection [11, 12].
Behavior-based intrusion detection is implemented by detecting the behavior that
deviates from normal or expected behavior of the system or the user. It is also known as
anomaly detection or detection by behavior . The current activities are compared
with a reference mode. If any significant deviation is observed, an alarm is generated.
Therefore, its good detection completeness can contribute to automatic discovery of new
attacks. However, it may cause a relatively high false alarm rate, which reduces its
detection accuracy. Implementation of behavior-based intrusion detection includes
expert systems, neural networks, user intention identification and computer immunology
1.4 Host-based versus network-based intrusion detection
There are two kinds of intrusion detection models, network-based and host-based,
which recognize attacks. In either case, these systems check for attack signatures and
specific patterns that usually indicate malicious or suspicious intention.
The primary difference between host-based and network-based intrusion detection
is the data source from which the system derives the information used to conclude the
occurrence of an attack. Host-based intrusion detection is the first area that was explored
in intrusion detection, and it involves loading a piece or pieces of software on each server
that needs to be monitored. The loaded software uses log files, process accounting
information, user behaviors or output data from application-based intrusion detection
systems operating on the host. Network-based intrusion detection monitors the raw
network packets on its network segment. It does not require software to be loaded and
managed on a variety of hosts .
Since host-based intrusion detection involves looking at the logs containing
events that have actually occurred, it can measure whether an attack is successful or not
with more detail and accuracy than a network-based intrusion detection system. It can
monitor all user log activities as well as what each user does during the time that he
connects to the network. Host-based intrusion detection may also check the integrity of
the system files and watch for suspicious processes, such as file accesses and attempts to
install executables or access-privilege service. Host-based intrusion detection systems
reside on each host to monitor the activities that occur at the user level, much of which
can not be detected by a network-based system. For example, attacks from the keyboard
or rlogin within a LAN do not cross the network and, therefore, cannot be detected by
network-based intrusion detection. In switched environments, it is difficult to achieve
sufficient network coverage by network-based intrusion detection systems. Host-based
intrusion detection provides better performance due to residing on as many critical hosts
as need. Furthermore, encrypted network traffic can be hardly handled by a network-
based system. When the incoming traffic comes to an operating system, it has been de-
encrypted. A host-based system is able to monitor such events easily. Finally, a host-
based intrusion detection system does not need any additional hardware and is cost-
effective. These are the reasons why host-based intrusion detection system has been
deployed first [14, 15].
Network-based intrusion detection systems operate at the network level without
regard to the type of operating system. They monitor all packet headers and can
investigate payload contents that may slip by a host-based system. In addition, a network-
based system examines live packets in real-time so that any attempt to remove or hide
evidence of an intrusion is impossible. By monitoring network traffic, network-based
intrusion detection systems can detect malicious attacks. Real-time detection and
response is one advantage of network-based detection. In host-based intrusion detection,
events have already happened before a host-based intrusion detection system can observe
them in log files. In addition, network-based detection can also detect unsuccessful
attacks. If placed outside a firewall, a network-based system can identify an activity
targeted at resources behind the firewall, even if the firewall rejects the attempts. This
unsuccessful attack information can be used to evaluate and refine the security policy.
Accordingly, it re-enforces the system to ensure that newer attacks will not pass through
[14, 16, 17].
Both network- and host-based intrusion detection system solutions have unique
strengths and benefits that complement each other. Integrating network- and host-based
approaches is the choice of most advanced intrusion detection systems. Network-based
detectors monitor network activity, while host-based detectors sit on various hosts, and
both report back to a single management center/administrator. A center may have a view
of the network at all levels so that the enforcement of security policy is ensured and a
great flexibility in deployment options is provided.
1.5 Organization of the Thesis
This thesis begins with the introduction of problems in network security and
intrusion detection systems. In Chapter 2, four distributed intrusion detection system
architectures, AAFID, EMERALD, GrIDS and GIDEM, are introduced. Chapter 3
presents a host-based and event-based detector architecture and the detector's
implementation is described in Chapter 4. Chapter 5 will state the detector's performance
results and Chapter 6 gives a conclusion and possible future directions.
BACKGROUND AND PREVIOUS WORK
There are hundreds of distributed systems available that perform intrusion
detection, intrusion prevention and system security checking. Among those, AAFID,
EMERALD and GrIDS are representative. At the end of this Chapter, the intrusion
detection model (GIEDM) in the CONS lab is introduced.
2.1 AAFID (Autonomous Agent for Intrusion Detection)
AAFID is a distributed monitoring system with an orientation towards intrusion
detection developed at the CERIAS, Purdue University. It is an agent-based intrusion
detection system that uses many autonomous agents  working independently and
collectively to perform distributed intrusion detection.
AAFID architecture has four components which are referred to as entities: agents,
filters, transceivers and monitors. The AAFID system is designed as a hierarchy of
components with agents at the lowest level performing the most basic functions, and
monitors and transceivers at the top-level overseeing the operation of agents and filters
on a per-host and per-host set basis. Figure 2.1  shows the physical layout of the
components in an AAFID system. Figure 2.2  shows the corresponding logical
organization of the same system .
An AAFID system can be distributed over arbitrary number of hosts, each of
which may contain any number of independent agents working simultaneously to monitor
certain aspects of a system and reporting anomalous behaviors or occurrences of specific
events . Some agents may perform network monitoring functions, while others may
perform host monitoring functions. The results produced by the agents are collected on a
per-machine level. In this way, the system may find correlated events that are reported by
different agents, such as ones caused by the same attack. Furthermore, detection of
attacks involving several different hosts is feasible since the reports produced by each
H O ST__ _ **********........................... H O ST
OA ........ .. ..
E Monit ,
S""'" E A
SControl flow UI
................ Data flow HOST
Figure 2.1 Physical layout of an AAFID system 
Figure 2.2 Logical representation of the same AAFID system 
machine are aggregated at a per-network level. Agents in an AAFID system can be
added, started, stopped or removed without altering other components. Additionally,
agents may employ mechanisms of reconfiguration and update at run-time without the
need to restart the whole system .
Agents may use the data collected by filters that are started by transceivers. There
is one filter for each data source that provides a subscription-based service to multiple
agents. When an agent subscribes to a filter, it specifies the data it needs, and the filter
only sends the agent the data that it wants. This is the data selection function that the
filter performs. The other function that performed by filters is data abstraction. Filters
implement all the architecture- and system-dependent mechanisms for obtaining the data
that agents request. Therefore, the same agent can run under different architectures
simply by connecting to the appropriate filter .
AAFID agents report their findings to one or more transceivers instead of
communicating directly with each other. Transceivers are per-host entities exerting
control, such as start and stop, over the locally running agents, and they perform analysis
and reduction processing on the data received from the agents. Transceivers on different
hosts may transmit data to each other. By combining the reports from different agents,
transceivers oversee the status of their host. If any suspicious activity takes place, it is
reported up the hierarchy to one or more monitors .
Monitors are the highest-level entities in the AAFID architecture. They analyze
the data received from transceivers to detect intrusions in the network and can do higher-
level correlation to detect an event that involves multiple machines. Furthermore,
monitors have the ability to detect intrusions that may escape the transceivers. A monitor
may report information to a higher-level monitor. They can control transceivers and other
monitors. Monitors can access network-wide data, therefore they have a global picture of
the status of the network they are monitoring. In the AAFID architecture, the user
interface is separated from the data collection and processing entities. Monitors provide
the access point for the user interface to obtain information and send commands .
The desirable feature in AAFID is flexibility and scalability. However, it has
several limitations. For example, it may be useful to allow more communication between
entities in the AAFID architecture, that is, to allow one agent to communicate directly
with another agent.
2.2 EMERALD (Event Monitoring Enabling Responses to Anomalous Live
EMERALD is a distributed scalable intrusion detection project undertaken in SRI
International's System Design Laboratory. It addresses detection of malicious activity
through and across large and complex networks.
The EMERALD architecture is composed of a collection of independently tunable
service monitors deployed in hosts and performing monitoring functions. EMERALD
monitors can be used to prevent DOS (denials of service) and loss of availability. They
work independently with application logs and network services to detect and respond to
malicious activity at the operating system and network layers, and can interoperate to
form an analysis hierarchy. Hierarchical data processing helps scalability to large
The EMERALD monitor architecture is illustrated in Figure 2.3 . The
EMERALD architecture employs three analysis units: a signature engine, a statistical
profile engine and a resolver within the monitor. These three types of units surround the
target-specific resource objects.
Figure 2.3 EMERALD monitor architecture 
The signature-based and profile-based anomaly detection subsystems will analyze
data at various logical layers of abstraction, on appropriate platforms to minimize the
amount of data that must be analyzed at each logical layer. Signature-based analysis
engine uses historical records to build normal detection models to verify a new instance
as a known intrusion activity. The profile-based component is used to monitor network
traffic and detect unanticipated anomaly modes .
A resolver is a countermeasure decision engine interpreting analysis results
generated from signature and statistical engines. It correlates alarms from anomaly
detectors throughout an enterprise network. The EMERALD resolver provides further
analysis of security violations and intrusion events and implements responses to
intrusions, which involve reconfiguration of the intrusion detection system. Signature
analysis and the integration of an expert system also improve the correlation capabilities
EMERALD monitors demonstrate a streamlined, distributed intrusion detection
system providing localized real-time analysis of multiple network services (e.g., FTP,
SMTP, HTTP) and infrastructure (e.g., routers or gateways). An EMERALD monitor may
interact with its environment either passively or actively. Monitors provide a well-defined
application programming interface (API) to interoperate with other third-party intrusion-
detection tool suites, such as deriving and receiving event data and analytical results
among other third-party security services .
Each monitor also contains a resource object that provides a reconfigurable
library of information to customize the monitor to the target analysis. The monitor code
base is analysis target independent. Therefore, the EMERALD monitor is scalable with
no modification to anything except the content of the resource object .
EMERALD presently includes several eXpert components that use P-BEST rule
based inference. The EMERALD eXpert is a signature-analysis engine based on the
expert system shell P-BEST. P-BEST is an expert system shell with a set of rules that
enable it to identify or suspect an anomalous activity in the system. P-BEST-based
eXperts may be independently distributed to analyze the activity of multiple network
services or network elements [22,23].
EMERALD builds multiple local monitoring capabilities coordinating the
distribution of analyses to detect both local attacks and coordinated attacks such as
distributed DoS. It can interoperate with other analysis platforms within its own scope.
Providing global rather than local analysis is an important issue that further research
2.3 GrIDS (Graph-Based Intrusion Detection System)
GrIDS is intended to detect large-scale automated attacks on networked systems
and was designed at UC Davis. It is a hierarchical system, which puts together results of
localized host-based and network-based intrusion detection systems into graphs and is
able to aggregate those graphs into simpler forms at higher levels of the hierarchy. The
whole GrIDS hierarchy requires an Organizational Hierarchy Server (OHS) to oversee
the operations and to provide global view of the hierarchy .
GrIDS employs data source modules running in each host to report incidents and
network traffic to graph engines that build a causal structure of activity in the network
called an activity graph. The activity graph allows automated attacks to be detected in
near real-time. For scalability to large networks, these graphs are constructed locally and
passed upward in the hierarchy where a parent engine maintains a graph of coarser
resolution. Tracking of a worm is an example of building such an activity graph which is
shown in Figure 2.4 . The nodes of an activity graph correspond to hosts in a system,
while the edges in the graph correspond to network activity between those hosts. In
addition to forming graphs from network communications, GrIDS allows the nodes and
edges to be annotated with attributes that supply additional information about the events.
Figure 2.4 The beginning of a worm graph, and a more extensive view of the same worm
User-customizable rule sets describe to GrIDS how to build these graphs to find
certain activity patterns and when to send alerts. These graphs are then compared against
those patterns of intrusive or malicious activities, and if they look similar an alarm and
corresponding reactions are generated. The rule set has rules to build the graph and also
assessment rules to evaluate the graphs and to take appropriate actions. Not all types of
network behavior are related to each other. GrIDS implements several independent
graphs representing different types of activities, as specified by the rules. It may be much
easier to form and to analyze the structures independently than a large single graph
containing unrelated information .
GrIDS detects large-scale correlated or automated attacks by checking to see if
the constructed activity graphs at any level have violated some policy as captured in a
rule set. Policies are compiled into rule sets .
GrIDS provides a scalable design for intrusion detection in large networks. It
allows specifying finer grain access control policies than firewalls. It also has some
limitations. It may not be possible to capture all intrusions as violations of an explicit
policy. Further, since the system depends on a centralized OHS for hierarchy
management, it may not be useful in a large dynamic network.
2.4 GIDEM (Generic Intrusion Detection Model)
A real-time generic intrusion detection model (GIDEM) is presented in
Seetharaman Balasubramanian's thesis . The system consists of coordinators,
detectors, tranceivers and response agents. The structure of the design is shown as Figure
The coordinator is the highest level of the system. The design of the coordinator
applies an IETF model , which is a part of the analyzer and the manager. It provides
the event analysis and detection system control. When detectors report events to the
coordinator, the coordinator parses them. Using a rule-base, the coordinator decides
whether to trigger an action and issue a specified response to the corresponding response
agent. The coordinator also can start, stop and reconfigure the detectors as needed.
Monitor profile Coordinator
Response profile Coi
Transceiver Transceiver Response
Figure 2.5 Architecture of GIDEM 
The detectors distribute in the entities of the GIDEM. Their primary duties are to
collect data and conduct possible data analysis. They have the capability to communicate
with the transceiver and the coordinator. The idea of the transceiver is similar to that of
the AAFID architecture. The transceivers provide the external communication interface
for each host. Basically, the transceivers have the same functions as the coordinator does.
The only difference between the coordinator and the transceiver is that transceiver only
controls one or limited number of detectors while the coordinator globally controls all the
detectors in GIDEM. The responder is responsible to execute commands from the
coordinator. In addition, the user interface is modeled as a response agent.
This thesis presents a host-based and event-based detector for GIDEM. The goal
of this thesis is to design a near real-time host-based detector, which can analyze raw data
quickly and efficiently, eliminate the redundancy in the log data and be capable of
reconfiguration and recovery. More details are explained in Chapter 3.
The purpose of this thesis is to design a host-based detector that can efficiently
monitor and analyze routine log activities and report an anomalous intrusion to the
coordinator. Its duties are to bring and transfer the log data from the log file in near real-
time, to parse the data and rebuild it to the standard format. Then it analyzes the routine
login data and reports any anomalies. Finally, it compresses the raw data to save space in
the log file. In this chapter, the detector design is described in two parts: detector
components and detector strategy.
Detector components include four functional units: data collector, data buffer,
data operator and data container. After the detector is initialized by loading its
configuration file and user profile, all of detector components are started based on the
configuration information. The detailed design for each component is presented in the
Buffer e Operator
Figure 3.1 The detector components
3.1 Detector Components: Data Collector
Raw messages are generated by Syslogd, which is a collection mechanism for
logging various messages generated by the kernel and applications running on the UNIX
operating system. Every message collected by Syslogd is associated with a level of
importance, which consist of the facility the system component from which the
message originated. Combinations of priority and facility specify actions related to the
message. The values of both parameters (priority, facility) are described in detail on the
syslog(3) man page as shown in Table 3.1  (priorities are listed from lowest to
highest below). The priority levels between debug and warning is the major concern in
Table 3.1 Syslogd facility and priority 
Facility Description Priority Description
Auth Used by authorization Debug Normally used for
systems loginn) (lowest) debugging
Used for the cron and at
Cron Info Informational messages
Conditions that may
Daemon System/network daemons Notice C
Produced by kernel
Kern Produced by kernel Warning Any warnings
Lpr Printing system Err Any errors
Critical conditions like
Mail Mail system Crit
Mak Internally used for time Any condition that demand
stamps immediate attention
News Reserved for the news Emerg
News (highest) Any emergency condition
User Default facility, used for
Reserved for the uucp
Local0...7 Reserved for local use
The function of the data collector is to collect data from the Syslogd daemon,
multiplex the raw log to the corresponding data buffer and finally pass the formatted log
messages to data container. It has three parts: data reader, data parser and data
compressor. The responsibility of data reader is to obtain the raw data in real time or
near real time and pass it to data parser. The data parser is to parse the raw data string
that is read from the Syslogd and to multiplex it to the corresponding type of data buffer.
Another responsibility of the data parser is to assemble the data into the compressed file
format, which is delivered to the data compressor. When the data compressor receives
the log data (which is in standard format), it converts it into a binary format and passes it
to the data container.
3.2 Detector Components: Data Buffer
When data is converted to the standard format by the data collector, the data
buffer holds the data in memory for fast analysis. However, holding all this information
wastes memory and is unnecessary. The data buffer consists of two basic units, the data
indexing tree and temporally ordered data indexing list, which has a sliding window. The
sliding window plays a major role in dynamically maintaining all log data read from the
Generally, the IP address is composed of four numbers between 0 and 255. A tree
structure is selected by the data indexer to maintain an index of all log data. The depth of
a detector tree is four. The main advantage of the tree structure is to provide for faster
searching performance than a circular list buffer does. If the frequency of log data
redundancy is high, saving more space will be another benefit as well.
Leaf node of
I W in d ow... / .' ..................................................
Data Indexing Tree window list
Oldest Compressed file ewes
Figure 3.2 Architecture of the data buffer
However, if a detector tree keeps growing, it will finally reach its threshold and
use up all the pre-allocated system resources. To monitor the capacity of the detector tree,
a data list is created and is extended as the detector tree is modified. Every leaf node of
the buffer tree (representing an IP address) has a link with one node in the sliding
window list (representing timestamped audit event). When time passes a certain interval,
an aging policy is applied to the data buffer. That is, the sliding window will move
backward on the list with a certain size, even if an anomalous intrusion is detected. All
out-of-date log data to the left of the window will be removed from the buffer tree and
the sliding window list. As stale sliding window list nodes are deleted, the information
on the corresponding data indexing tree nodes (both leaf and non-leaf) are necessarily
updated. Therefore, the capacity of the buffer tree is changing. Selecting a suitable
sliding window moving rate will result in increasing the efficiency of the buffer tree.
3.3 Detector Components: Data Operator
Monitoring and updating the data in the data buffer is the major function of the
data operator. It contains a series of engines that perform searching, inserting and
deleting data on the buffer tree that function upon request. To keep every engine running
without blocking any other engine, a multi-threaded detection system is selected. Each
thread does not have to wait for any other thread to finish. All the threads proceed
simultaneously with fair concurrency control. Properly formed threads can often improve
the performance of an application and they do not incur significant implementation
overhead. They effectively give good performance for the effort expended.
The searching engine seeks to search the information of a specific or suspected IP
address. Usually, the searching starts from the root of the buffer tree. The insertion
engine is requested when needed. When standard log data comes from the data collector,
it will traverse the buffer tree and update the corresponding node's information. If it is a
non-existing node, the insertion engine creates a new node. The deletion engine starts to
work upon being informed that one specific IP needs to be deleted from the buffer tree.
The deletion request will be needed in either of the following cases: 1) all the sliding
window list nodes related to that IP address appear to the left of the sliding window; 2) an
anomalous intrusion is detected and confirmed.
Besides its updating function, the data operator is responsible to communicate
with the upper level coordinator and to report an anomalous intrusion at different warning
levels. For example, when an IP reaches one warning level's threshold, the detector
needs to report to the coordinator that a warning state is occurring and to consider that IP
as suspect. When a suspected IP later reaches the threshold of higher level warning state,
the detector needs to report that higher state to the coordinator again.
Figure 3.3 Architecture of the data operator
A client and server model is applied to implement the relationship between the
detector and the coordinator. When the detector finishes the data analysis of the raw log
message, the data response engine of the data operator cooperates with the coordinator.
Its responsibilities include notifying the coordinator about its initial launching, sending
any detection result to the coordinator and then waiting for the coordinator's response.
When the coordinator makes a decision on the reported event, it will send the response
back to the detector. Once the response engine obtains the response from the coordinator,
it will update the configuration of the detector or user profile, if necessary. Then the
detector will start to execute detection based on the new configuration. Therefore, it
provides the coordinator a facility to control the detectors operation. In a word, the data
operator functions by keeping the buffer tree running more flexibly and efficiently.
3.4 Detector Components: Data Container
When the data collector delivers the standard format data to the data container, the
data container then writes the data into the compressed log file. Three issues force the
designed detection system to compress the log file. First is the growth of the log file,
second is Unix or Linux log file rotation policy and third is the detection system
recovery. Left unaddressed, the first issue will result in a huge log file occupying too
much storage space. The second and third issues will cause losing some old data which
might be useful for slow attack detection. In the designed detection system, all login data
needed by the detector is necessarily stored in a compressed file so that two advantages,
keeping records and saving space, are realized. Keeping records helps necessary polling
analysis. When a polling analysis request is invoked, all of the records can be grouped
according to the requirement. Saving space is implemented by converting each log data
format to a binary format. The compressed log file's creation depends on help from the
data collector. The compressed data interpreter helps to read data from a compressed file.
More detail on compressing files is given in Chapter 4.
3.5 Detection Strategy
Data analysis is based on two different situations, namely failed login and
successful login. Regarding login failures, keeping a count of how many attacks from a
certain IP address is the most important key for intrusion detection. A login failure may
occur from accidental failure or intentional attack. Occasional failure login occurs when
one user forgets his password or mistypes it. Although login failures from same remote
node may appear for different usernames or passwords, their source IP should be
identical. Therefore, only one leaf node in the data buffer can represent all such login
attempts. That is, space within the data buffer can be saved by eliminating redundant
records with the same IP address. In the process of detection, the occasional login failure
can easily be discovered because it is not repeated over time. Once the user recalls the
correct password or obtains it from the administrator, this type of login failure will not
appear again for an indefinite period. Therefore, its representation in the data buffer will
not be updated. With the motion of the sliding window, such login failures will be
removed from the data buffer when they become out-of-date and appear to the left of the
window. As a result, the capacity of the data buffer is made available to hold other
different IP logins.
Prevention of intended attack is primary function of the detector. Such attacks
have two kinds of format: burst attack and slow attack. Burst attack implies many login
failures happen within a short time. This kind of attack can be from a single source or
multiple sources. As for those from a single source, they can be detected by finding the
counter of the leaf node increasing dramatically. When the login counter reaches
detection threshold, it can be detected immediately. Multi-source attacks have two
formats: non-targetted attack and targeted attack. The solution for non-targetted attack is
the same as that for a single source attack. Targetted attack indicates that multiple
sources are attacking the same user account simultaneously. A polling analysis on that
account will be necessary. Checking each user account's login data within a certain time
can detect targeted attacks. Such a polling analysis can be invoked by the data collector
when it discovers login time has surpassed its reasonable threshold.
A hacker may attack the system more cautiously and patiently. He may not try
each attack continually but keep trying periodically; this is a "slow attack." To detect this
kind of attack, a polling analysis needs to be based on the weekly or monthly log data.
With a more extensive search, the possibility of detecting such slow attacks is increased.
For example, if there may be an IP address of a hacker from which many failure login
actions originate. However, since the interval time between two adjacent actions is
longer than that of the sliding window moving interval time, the hacker's data is always
removed from the data buffer and he escapes detection. Fortunately, this can be
discovered by weekly or monthly statistical analysis.
A lucky hacker may get a chance to log onto the system successfully due to
various reasons, e.g., obtaining a normal user's information in an inappropriate way.
Depending on knowledge-based successful login analysis, the chance to catch an
unauthorized login is enhanced. If the same "user" logs in from two separate sites at the
same time or in a short time, that account can be considered to be "infected." If a user
logs in at unusual times, that account can be suspected of being infected. In order to find
if there is an intrusion, it is necessary first to confirm the authenticity of the user by
coordinator. If coordinator or detector gets negative response from the user, a polling
searching on previous failure logins from that suspected IP is necessary.
A host-based detector is designed to deal with data extraction, data analysis, data
compression and data storage with the Linux Syslogd facility. The architecture of the
detector is shown in Figure 3.4. Real-time data acquisition makes it possible that an
intrusion detection system can be running in real time or near real time. Meanwhile, such
acquisition also reduces the chance that a hacker may remove a record from the log file.
A tree-like data buffer enhances data analysis performance, which is most important for
abrupt attack detection. Furthermore, with the help of sliding window aging, the data
buffer can dynamically change its size thus enabling it to continuously monitor login
actions efficiently. To keep a data buffer of a relatively small size running efficiently
without losing information, a compressed file acts as a complementary part of the data
buffer. Since all out-of-date as well as current data are recorded in the file, a polling
analysis then becomes feasible, which is beneficial for slow attack detection. As a
supplement, knowledge-based detection is added to detect those anomalous intrusions
with successful unauthorized login. As a result, the host-based detector is designed to be
capable of re-configuration, to run concurrently as a multithreaded system, dynamically
to detect both login failures and successful unauthorized logins and to be able to do post-
event polling analysis.
The implementation of the detector is described in five parts: initialization of the
detector, data collection, data buffer creation, successful login detection and system
controlling. All parts of the detector are written in C/C++ programming language.
4.1 Initialization of the Detector
To be an easy-controlled detector, the designed detection system has implemented
a configuration and re-configuration policy. A previously prepared configuration file
specifies major characters of the detector. The configuration file must be read when a
detector starts. During the detection process, other agents such as coordinators or
response agents can modify the configuration file. Each time this file is changed, the
detector is requested to re-read it and operate on a new configuration.
As a part of configuration and knowledge-based detection, a user profile is loaded
at the beginning of detector initialization. The profile helps the detection system in
making a judgment whether a successful login is suspicious. During the process of
detection, the user profile can be updated to make the detection knowledge system more
After finishing the configuration procedure, the detector needs to register with a
remote coordinator, at a location that is present in the detector's configuration file. The
registry process is completed by using UDP datagram socket communication.
4.2 Data Collecting in Near Real Time
Once the detector is initialized, the next step is to obtain the raw log data in real
time or near real time. The original message data are created by the syslogd daemon.
Actions are associated with the syslogd's combinations of facilities and priorities. Those
combinations specify what can be done with the actual log messages that match a given
combination as shown in Table 4.1 .
Table 4.1 Sample actions of syslog.conf 
Action (example) Description
/dev/console Send messages to devices
/var/adm/messages Write messages to files
@ loghost Forward messages to a loghost
fred, userl Send messages to users
Send messages to all logged-in users
Under Linux, syslogd can write a log message into a named pipe (FIFO). This
makes it possible for the detector to obtain log messages from the named pipe in real
time. In /etc/syslog.conf, a new line is added like:
This means that all log messages are required to write into a named pipe,
"syslog.pipe." When a detector is activated, the named pipe is opened by the detector
with a "READONLY" flag. Then one thread is kept blocked and awaits the passage of
log messages from the opposite side. Once a new log message is written into the pipe
from the other side, the receiving thread can obtain it without any delay. From time to
time, obtaining real-time data is implemented by that thread.
4.3 Implementation of Data Assembly and Compression
The data in the log file are saved in ASCII text format. Although ASCII format is
more user-friendly, it is wasteful of disk storage space. If the data is saved in a binary
encoding, it will dramatically reduce the storage size of the log file. Therefore, it is
necessary to convert raw data from its text format, such as ASCII, into a binary format.
For each data entry, there are five significant attributes: timestamp, user id, source IP
address, login type and flag. One data entry can be formed as shown in Table 4.2.
Table 4.2 The format of data entry
User id 16-bit Type 6-bit Flag 2-bit
IP address 32-bit
Each entry contains two fields: header and content. The header includes flag bit, type bits
and user id bits. The flag bits have four different combinations which are listed in Table
4.3. Those four different combinations result from the comparison between the current
and the previous data entry. If they have the same IP, the content part of the entry only
has a timestamp field so that it reserves storage room. Also, because the first bit of the
flag indicates whether current entry is from another IP, it functions as a delimiter within
the whole compressed log file. If its first bit is equal to 1, the content field will only have
a timestamp field.
Table 4.3 The description for flag bit
Flag bit Description
00 Same user from different IP address
01 Different user from different IP address
10 Different user from same IP address
11 Same user from the same IP address
The IP field contains 32 bits which correspond to four numbers within the 0-255
range. Within the timestamp field, the detection system chooses data type time_t to
represent calendar time. When interpreted as an absolute time value, time t represents
the number of seconds elapsed since 00:00:00 on January 1, 1970, Coordinated Universal
Time (UTC). The login types can be classified in varying formats, namely ssh, ftp,
rlogin, etc. Type bits can represent a maximum of 64 different login types. Table 4.3 lists
several kinds of types.
Table 4.4 Type bits for different login
Type Binary code Type Binary code
Failure ssh 000001 Successful ssh 100001
Failure ftp 000010 Successful ftp 100010
Failure rlogin 000011 Successful rlogin 100011
Unknown failure 000000 Unknown success 111111
Once the raw data is translated into standard format, the data compressor will
write standard data word by word. In case that the system crashed, the detection system
can be recovered with the help of the compressed file.
4.4 Implementation of Creating the Data Buffer
After the raw data is tokenized and converted into the standard format, it is ready
to be added into the data buffer for analysis. In previous work, data redundancy has been
a severe problem in the circular list buffer. Such a buffer is not flexible and efficient for
modifying, updating or deleting data records. In this detection system, a tree-structure is
being applied to solve these conflicts. Such a tree-structure is based on the idea of
hierarchy of IP addresses. It consists of four levels that are related to IP address format.
The root is the highest level. The address mapping between levels is aided by a hashing
function. A hashing chain is used to save memory space and achieve fast searching
performance. The relationship between levels is "pseudo double" linked. The upper
level node (parent node) has a pointer which points to its hashing chain. Every node
(child node) in the hashing chain has a pointer which points to its parent node.
In every node, the information is kept in the following format.
timet least ts;
timet latest ts;
level 0 level 1 level 2 level 3
........ ........ ............-- T o next level
-- To next level
Counter of the node
STo next level
Figure 4.1 Hashing tree of data buffer
The num_child signifies the total number of its children; ptr_parent is a pointer
to its parent; ptr_child is a pointer to its children level; least_ts is its oldest timestamp;
latest_ts is its most recent timestamp. Because every data entry in the compressed file
cannot possess a pointer that refers back to the node on the indexing tree, a linked list is
maintained between the data indexing tree and the compressed file. The leaf nodes on the
indexing tree have a one-to-one relationship with the nodes of the linked list. At each
node of the list, there are only three attributes that need to be stored: ptr_node,
num_byte, node_ts and prev_ts. The ptr_node is a pointer which points to the leaf
node on the indexing tree. The num_byte records the beginning location of each data
entry in the compressed file. The node_ts denotes the timestamp of the pointing leaf node
on the indexing tree. The next_ts is a pointer to point to next list node with the same IP
address, if any. The indexing list node format is shown below.
timet node ts;
Detection tree initialization starts from the root. It proceeds down each non-leaf
level and creates the appropriate values for num_child, ptr_parent, ptrchild, leastts and
latestts. When it reaches the leaf level, it updates its counter, ptr_parent and latestts.
Meanwhile, it updates the ptr node of the corresponding node in the linked list so that a
one-to-one relationship is built up. When the sequential new log data are read, the
detection tree is updated from the root and goes down each non-leaf to the leaf level. If
the expected child does not exist, the tree creates a new tree node with appropriate values
for each of its attributes. Every time a leaf node is formed, the relationship between the
indexing tree and linked list is again updated.
4.5 Implementation of Successful Login Detection
Even if a successful login occurs, it does not mean that it can be totally trusted.
Two possible types of unauthorized logins need to be addressed. One is the same user
logging in from a different IP address within a very short time; another is a user's login at
an unusual time slot.
} Naming List
-| -- - -W--
by uid and
S1 c3 13 c3 timestamp
} Timing List
Figure 4.2 Data structure of naming and timing lists
This detection system needs to maintain a profile for each user. The profile
contains the basic login information for each user, such as normal login time slots and
source IP domain. The user profile is pre-loaded at the beginning of detection to build a
user knowledge table in the system. As the detector running, the information on the
knowledge table can be refined if requested by the coordinator or response agents.
The detection buffer for successful login is implemented in a multi-linked list
format that contains the naming link and timing link. Each entry on the list contains a
user id, source IP address, timestamp, file location and deviation. The user id is the key
for each entry and the naming list is sorted by user id. A sorted, linked list of user ids is
built, with each uid node pointing to a linked list of successful logins for that user id in
timestamp order. All successful logins are also double threaded based on timestamp
order; this is called the timing list. The timing list also has pointers for the head (most
recent) and tail (oldest) entry. Each entry on the list also has a file location field to record
its location in the compressed file and a deviation field to record suspect information.
The value of the deviation field is inherited from the entries with the same user id. The
format of the successful buffer node is given below.
int uid; /* user id */
time_t nodets; /* timestamp */
int ip; /* address ip */
int location; /* the starting location in the compressed
SListNode *ptr_prev_id; /* pointer to previous different id */
SListNode *ptr_next_id; /* pointer to next different id */
SListNode *ptr_next_siblingid;'/* pointer to next same id */
SListNode *ptr_prev_ts; /* pointer to previous timestamp */
SListNode *ptr_next_ts; /* pointer to next timestamp */
When a new entry is added, it first finds its user id in the naming list and then
appends itself to the list of entries for that user id. If there is no matching user id, the
detector will create a new user id node and insert it at the right place within the naming
list. After the naming list is updated, the timing list will be adjusted accordingly using
the head pointer. After finishing the insertion of the node, data analysis is then
performed. In every entry's deviation field, there are four basic counters to monitor four
aspects of basic detection: normal login times, login times at an unusual time slot from a
normal IP domain, login times at the normal time slot but from an unusual IP address and
login times from an unusual IP address and at an unusual time slot. In the designed
system, there are different threshold levels for each category.
When the checking process starts, the new entry's timestamp and source IP
address are first compared with those in the user knowledge table. If an unusual login
time slot or source IP is found, its deviation field will record such data. Then, its
timestamp and IP address are compared with the one preceding it in both sub-lists of the
naming list and timing list. With this kind of checking system, two kinds of anomalous
behaviors easily can be found. The comparisons within the naming list can detect an
anomalous login from two different IP sites within very short time. Comparisons within
the timing list can detect a user login at an unexpected time.
4.6 Implementation of Multi-thread System Manipulation and Concurrency Control
The designed detection system applies a multi-thread mechanism so that the set of
global variables and function calls in the data buffer can be fairly shared and accessed by
every thread. A simple approach would be to create a thread for each request received.
In this designed system, pthread in the GNU thread library is used to create a thread.
There are several situations that will require a thread request. At the stage of the
detector initialization, the detector needs to launch two basic threads: a listening thread
and a deletion thread. The listening thread functions as a response engine in the data
operator being designed. Since the designed detector uses a client/server model in the
communication between the detector and the coordinator, the listening thread's
responsibility is to receive the message from the coordinator. The listening thread
prepares to deal with several messages:
1. When the coordinator requests current detector state, the listening thread
will create a thread to obtain the current state and send it to the
2. When the listening thread receives a reconfiguration request, it will re-
read the configuration file in the assigned location and update the current
detection configuration. Since all configuration variables are shared global
variables, this recent modification will immediately reflect on the current
3. If the listening thread gets updated user information, it will change the
appropriate item of that user in the user knowledge table and reset the
deviation field of the buffer entry for each user.
4. If the listening thread receives many messages that are not sent by the
coordinator, it will report to the coordinator that its port has been attacked.
5. If the coordinator sends a "quit" message, the listening thread will set a
global variable, "TERMINATION," as "true" so that every thread will
terminate its execution and the whole detector quits.
Another thread created at the initial stage is called the deletion thread. It
implements an aging policy in the detection system. As long as the system is running, the
log data will keep increasing. A fixed data buffer can not handle too many data storage
requests. It is necessary to delete some out-of-date data.
A sliding window is selected to continue to update the data on the detection
buffer, which is moving along the timing linked list of data buffer. A global variable is
used to record the moving speed of the sliding window. It controls the moving speed of
the sliding window. As the sliding window moves, some old records on the indexing
timing list are removed from the data buffer. The deletion thread directs the pointer
ptr_node to delete the right node and to update its parent's information. Therefore by
such reduction, the size of the detection buffer can be moderately lessened. The original
moving speed is read from the configuration file. After that, if the thread finds the data
buffer size is reaching a warning limit, it will increase the moving speed. On the other
hand, if the thread finds the current buffer size is far below the warning level, it will
decrease the moving speed. Therefore, the implemented aging policy always keeps the
data buffer of a reasonable size, which effectively utilizes the capacity of the data buffer.
However, at present, the hackers have become more sophisticated and patient.
They can use slow attack strategies to avoid being caught by any fixed length data buffer.
The deletion thread has to deal with those attacks. The key is to conduct a post-event
periodic analysis. The deletion thread launches a temporary thread to scan the monthly
compressed file and builds a temporary buffer for analysis use. If any anomalous
intrusion is found based on the assigned detection threshold, it will report it to the
coordinator. Once the polling analysis finishes, the temporary thread is terminated and all
occupied system resources are released. The detection system can commit the polling
analysis in the very early morning so that it does not increase the system load too much
during the daytime.
In the process of data analysis, two threads are needed. First, when the insertion
engine of the data operator encounters the leaf node of the indexing tree, it finds that the
counter of the leaf node has reached a certain warning level's threshold. It needs to
launch a communication thread to send the message to the coordinator. Secondly, when
the system needs information on a specific IP address, a searching engine thread is
created. It traverses the data indexing tree and writes the requested information to a file
or sends a message to the requesting coordinator.
Although a multi-thread system enhances the detection system's performance, it
also brings up an important concern--concurrency control. Because there are shared
variables and memory addresses among the threads previously referred to, it is necessary
to create basic synchronization objects to implement concurrency control. The simplest
synchronization object is the mutex semaphore, which is used to allow only one thread to
access to a resource at a time.
In this thesis work, a simple mutex object is applied to control the concurrent
accesses in the multi-thread system because insertion and deletion of the data buffer will
occur mostly during detection. As a result, the critical section, data buffer, can be
accessed by only one thread so that there will be no data inconsistency or erroneous
This detector is devised as the design required. The use of named pipe makes
intrusion detection close to real time so that its use prevents the log message from being
modified by an intruder before defection takes place. Also, the detector enhances
surveillance and response performance. The configuration file and user profile facilitate
configuration and adaptation to the system and users. Reconfiguration is possible as
well. The relatively small size of the data buffer ensures a minimal system overhead. Its
sliding window's aging policy enhances data buffer utilization. If the detection system
crashes, the data in the compressed file can rebuild the detection system from the
breakpoint. Successful login detection, bolstered by user profiles, makes the detector
more knowledgeable. Finally, with the help of a multi-threaded system and a
client/server model, the designed detection system can run continuously without
supervision. Meanwhile, the coordinator can control this detection system by using a
message passing mechanism.
How efficiently the data buffer performs will determine the overall detection and
response performance, the testing is mainly focused on the data buffer testing.
Testing on the data buffer is accomplished on the host of 126.96.36.199
(tiger.cons.cise.ufl.edu) whose system configuration is RedHat Linux 7.0 on Pentium II
266 processor with 96 MB RAM. A sample log file containing 300k log data with fair
redundancy is created as an input sample. The performance test is conducted by inserting
10k, 30k, 50k, 100k and 300k bytes data into the data buffer respectively. Then, the
deletion and searching testing is applied to those data as well. Since the minimum time
scale is one second, a looping testing code is applied to increase the total running time so
that the testing time is much longer than the minimum time scale. As a result, the system
errors are reduced as much as possible. Beside the total amount of test data, the value of
hashing key in data buffer tree is chosen as another important variant on the detection
Fast searching and deletion performance: the testing results on the performance of
the deletion and searching are shown in Figure 5.1 and 5.2.
S 0.6 -/- 100000
o 0.4 30000
0 -- ^-------------
D =1 D=3 D=5 D=7
Figure 5.1 Performance of the deletion of the data buffer
U 0.2 -
D =1 D=3 D=5 D=7
Figure 5.2 Performance of the searching on data buffer
Both of the results indicate that a larger hashing key improves the performance.
However, it is found that when the hashing key is bigger than three, the enhancement of
the performance is no longer significant. A fast searching ability is helpful to collect the
information on a specific IP address and a fast deletion ability is good for performing the
sliding window efficiently. Because the shared data part is accessed by many threads, a
quick deletion or searching action will not block others to access the data buffer too long.
Fair fast insertion: There is a trade off between the performance of inserting and
searching/deletion. A more complicated data structure will be favorable to a better
searching/deletion performance. However, the original initialization and insertion will
pay for that. Figure 5.3 illustrates the result on insertion performance. The results
indicate that a larger hashing key is very beneficial for insertion performance because it
reduces the time of searching while inserted, especially when the total number of nodes
in the buffer is large. Actually, when the hashing key is equal to 1, it is similar to a
linked list. This proves that a tree-like buffer has a better performance than a circular list
buffer. Since the insertion time is directly related to the detection response, choosing a
hashing key is one of the most important factors.
Q 0.7- --300000
S0.6 -- 100000
C. 0.5 -
0. 3 30000
) 0.2 -- 10000
D=1 D=3 D=5 D=7
Figure 5.3 Performance of the insertion on data buffer
Memory Utilization: Another trade off in the data buffer is between hashing key
selection and memory utilization. Figure 5.4 depicts the memory utilization under
different hashing keys and different total amount of data. The data in Figure 5.4 has
been normalized to utilization with D = 1. As the total number of nodes increases in the
buffer, the utilization becomes better. However, with the growth of the hashing key, the
utilization decreases linearly. For the first phenomenon, a larger number of data
increases the chance of data redundancy. Consequentially, each node in buffer can
represents several different events with the same source IP address. As for the second
one, it is because more hashing buckets occupy more memory space. Therefore, the
utilization percentage is lowered in that case.
D = 1 D=3 D=5 D=7
Figure 5.4 Memory Utilization
Based on the testing results, it indicates that the detector detection and response
time is in a desired range. Overall detection performance of the tree-like data buffer is
better than that of a circular list buffer, though it has a little higher memory usage. When
the data has a high redundancy frequency, like in the case of login failure, its memory
utilization will be comparable with the circular list buffer. An important decision is
choice of a fair hashing key. A range between three and seven will be good enough,
because a bigger one will cause a lower utilization of memory and a smaller one will
cause inefficient performance.
CONCLUSIONS AND FUTURE WORK
This thesis implements a near real time host-based and event-based intrusion
detector. It satisfies the desirable characters of an intrusion detection system. The
detector configuration file allows the coordinator easily to reconfigure the detector so that
the coordinator can run the detector as needed. The use of the user profile contributes a
knowledge system to the detector. After a period of "learning," the detector can obtain
more accurate knowledge of local users.
The relatively small size of the data analysis buffer not only ensures a small
system overhead but also provides quick detection and response. The aging policy
implemented by the sliding window enhances the efficiency of the data analysis buffer
and the capability of its continued running. The compressed file makes periodic polling
analysis and system recovery possible.
A multi-threaded system allows the detector to perform different tasks
concurrently. A distributed detection system is feasible by using a client/server model.
Finally, an object-oriented programming technique can more easily extend the current
data structures for further development.
6.2 Future Work
Future work can be in following fields:
The basic data structure can be made to be reusable and flexible. Vectors of
counters in the data indexing tree can be built in the tree. Each vector has different
threshold and alarm function. The key of tree node, hashing function and the depth of the
tree can be varied. For the successful login data structure, the key of each entry can be
parameterized so that the data structure can be configurable.
The designed detector has good scalability and adaptability. Multi-level analysis
based on the data buffer and the compressed file is possible. Any future specific
searching request can be conducted as a thread to access the existing data analysis buffer.
Machine learning and artificial intelligence research can enhance the detection on
local users' abuse attacks. Accordingly, the user profile could be built more accurately.
 Bill Boni. "Hackers, Crakers, Lawyers and Other Dangers," Network Security,
4: 19-20, 2000.
 L. Todd Heberlein, Gihan V. Dias, Karl N. Levitt, "A Network Security
Monitor," Proceedings of the IEEE Symposium on Research in Security and
Privacy, pages 296-304, 1990.
 B.Harris, R. Hunt, "TCP/IP Security Threats and Attack Method," Computer
Communication, 22: 885-897, 1999.
 CNET News.com Staff, "How a "Denial of Service" Attack Works,"
 CERT Coordination Center, "Denial of Service Attacks," Software Engineering
Institute, Carnegie Mellon University,
http://www.cert.org/tech tips/denial of service.html, 1999.
 Brooke Paul, "DDoS: Internet Weapons of Mass Destruction,"
 Guy Helmer, Johnny Wong, Subhasri Madaka, "Anomalous Intrusion Detection
System for Hostile Java Applets," The Journal of Systems and Software, 55: 273-
 B. Mukherjee, T.L. Heberlein, K.N. Levitt, "Network Intrusion Detection," IEEE
Network, 8 (3): 26-41, 1994.
 Eugene H. Spafford and Diego Zamboni, "Intrusion Detection Using Autonomous
Agents," Computer Networks, 34(4): 547-570, October 2000.
 M. Wood, M. Erlinger, Intrusion Detection Message Exchange Requirments,
IETF (Internet Engineering Task Force) model, Internet Security Systems,
 Herve Debar, Marc Dacier, Andreas Wespi, "Towards a Taxonomy of Intrusion-
Detection Systems," Computer Networks, 31: 805-822, 1999.
 Sandeep Kumar and Eugene H. Spafford, "A Pattern Matching Model for Misuse
Intrusion Detection," Proceedings of the 17th National Computer Security
Conference, pages 11-21, Baltimore, MD, USA, October 1994.
 Mandar Raje, "Behavior-based Confinement of Untrusted Applications,"
Technical Report TRCS99-12, April 1999. Computer Science Department,
University of California at Santa Barbara.
 Internet Security System, Network- vs. Host-based Intrusion Detection, A Guide
to Intrusion Detection Technology, http://secinf.net/info/ids/nvh ids/, 1998
 Aaron Schwartzbard and Anup K. Ghosh, "A Study in the Feasibility of
Performing Host-based Anomaly Detection on Windows NT," 2nd International
Workshop on Recent Advances in Intrusion Detection--RAID99.
 Yang, Bo, "Design and Implementation of a Generic Port-based and Event-based
Detection System," Master of Science Thesis, Computer and Information
Sciences and Engineering, University of Florida, Gainesville, 1998.
 G. Vigna and R.A. Kemmerer, "NetSTAT: A Network-based Intrusion Detection
System," Journal of Computer Security, 7(1): 37-71, 1999.
 M. Crosbie, E. Spafford, "Defending a Computer System using Autonomous
Agents," Proceedings of the 18th National Information Systems Security
Conference, pages 549-558, Baltimore, MD, October 1995.
 Jai Balasubramaniyan, Jose Omar Garcia-Femandez, E. H. Spafford, and Diego
Zamboni, "An Architecture for Intrusion Detection using Autonomous Agents,"
Coast TR 98-05, COAST Laboratory, Purdue University, West Lafayette, IN,
 Guy Helmer, Johnny S. K. Wong, Vasant Honavar, and Les Miller, "Intelligent
Agents for Intrusion Detection," Proceedings of the IEEE Information
Technology Conference, pages 121-124, Syracuse, NY, September 1998.
 P. Neumann and P Porras, "Experience with EMERALD to Date," 1st USENIX
Workshop on Intrusion Detection and Network Monitoring, pages 73-80, Santa
Clara California, April 11-12, 1999.
 P. A. Porras and P. G Neumann, "EMERALD: Event Monitoring Enabling
Responses to Anomalous Live Disturbances," Proceedings of the 20th National
Information Systems Security Conference, pages 353-356, Baltimore, Maryland,
Oct. 7-10 1997. National Institute of Standards and Technology/National
Computer Security Center.
 Ulf Lindqvist and Phillip A. Porras, "Detecting Computer and Network Misuse
through the Production-Based Expert System Toolset(P-BEST)," Proceedings of
the 1999 IEEE Symposium on Security and Privacy, pages 146-161, Oakland, CA,
 S. Staniford-chen, S. Cheung, R. Crawford, M.Dilger, J.Frank, J.Hoagland, K.
Levitt, C. Wee, R. Yip, and D. Zerkle, "GrIDS--A Graph-Based Intrusion
Detection System for Large Networks," Proceedings of the 19th National
Information Systems Security Conference, volume, pages 361-370, Baltimore,
Maryland, October 1996.
 S. Cheung, R. Crawford, M. Dilger, J. Frank, J. Hoagland, K. Levitt, J. Rowe, S.
Staniford-Chen, R. Yip, and D. Zerkle, "The Design of GrIDS. A Graph-Based
Intrusion Detection System," Technical Report CSE-99-2, Computer Science
Department, The University of California, Davis, January 1999.
 James A. Hoagland, Raju Pandey, and Karl N. Levitt, "Security Policy
Specification using a Graphical Approach," Technical Report CSE-98-3,
Department of Computer Science, The University of California, Davis, CA, July
 Seetharaman Balasubramanian, "An Architecture for Protection of Network Hosts
from Denial of Service Attacks," Master of Science Thesis, Computer and
Information Sciences and Engineering, University of Florida, Gainesville, 2000.
 CERT Coordination Center, "Configuring and Using Syslogd to Collect Logging
Messages on Systems Running Solaris 2.x," Software Engineering Institute,
Carnegie Mellon University,
Jin Chen was born in Kaifeng, Henan, China. He received the Bachelor of
Science and Master of Science degrees from Nankai University, Tianjin, China, in July
1995 and New Mexico Highlands University, Las Vegas, in August 1999 respectively.
He will receive his Master of Science degree in computer and information science and
engineering from the University of Florida, Gainesville, in August 2001. His research
interests include computer network and database.