KNOWLEDGE BASED SYSTEM FOR RELIABLE PERIMETER

PROTECTION USING SENSOR NETWORKS

Constantin Voloşencu, Daniel-Ioan Curiac, Ovidiu Banias

Automatics and Applied Informatics Department, “Politehnica” University of Timisoara

Bd. V. Parvan nr. 2, 300223 Timisoara, Romania

Alex Doboli

Electrical and Computer Engineering Department, State University of New York, NY 11794-2350, USA

Octavian Dranga

School of Engineering, James Cook University, Townsville, QLD 4811, Australia

Keywords: Perimeter protection, wireless sensor networks, redundancy, knowledge based systems, neural networks,

perceptron.

Abstract: This paper provides a strategy for perimeter protection using sensor networks with hardware and analytical

redundancy. The sensor network reliability is augmented using a knowledge-based system, which implicates

the analysis of the trustworthiness of each sensor. For this, we used two stratagems: one that relies on

hardware redundancy based on the Confidence Weighted Voting Algorithm and one that relies on analytical

redundancy based on a neural perceptron predictor that uses past and present values obtained from

neighbouring nodes. This solution can be also a way to discover the malfunctioning nodes that were subjects

of an attack and it is localized at the base station level being suitable even for large-scale sensor networks.

1 INTRODUCTION

Sensor networks have proved their huge viability in

the real world in a variety of domains. Advances in

miniaturization, decreasing of their cost and power

and improvements in wireless networking and

micro-electro-mechanical systems have led to

research for large-scale deployment of wireless

sensor networks and formation of a new computing

domain. In the last years the deployment of small-

scale sensor networks in support of a growing array

of applications has become possible (Akyildiz,

2002), (Pottie, 2000). A lot of applications,

including seismic disturbances, contaminant flows

and other ecological or environmental disasters,

battlefield control, disaster management and

emergency response, which involve sensor

networks, will also be possible in the near future.

Detecting targets moving inside a field of interest is

one of the applications of wireless sensor networks

(Li, 2002), (Cao, 2005). These networks consist of

hundreds or thousands of heterogeneous disposable

sensor nodes, capable of sensing their environment

and communicating with each other via wireless

channels, coordinating and monitoring large areas.

Individually nodes possess properties such as

functionality and inter-node cooperation, under

limited energy reserves and technological

limitations. There are applications where the sensors

were generally bulky devices wired to a central

control unit whose role was to collect, process, and

act upon the data gathered by individual sensors. A

network of sensors could be developed with small

motion detectors, metal detectors, pressure detectors,

and vibration detectors, deployed around a valuable

asset. When the sensors were able to classify

“intruders”, a human reasoning to decide what to do

in response was necessary. The vision of the smart

dust program of wireless sensor network research

was to make machines with self-contained sensing,

computing, transmitting, and powering capabilities

so small and inexpensive that they could be released

into the environment in massive numbers. Sensor

Volo¸sencu C., Curiac D., Banias O., Doboli A. and Dranga O. (2007).

KNOWLEDGE BASED SYSTEM FOR RELIABLE PERIMETER PROTECTION USING SENSOR NETWORKS.

In Proceedings of the Second International Conference on Wireless Information Networks and Systems, pages 51-56

DOI: 10.5220/0002146500510056

 SciTePress

networks are expected to play an important role in

hybrid protection infrastructures when combined

with robots and human decision makers. In such

cases a knowledge-based system is a powerful way

of solving the problems. Redundancy in sensor

networks (hardware and analytical) can provide

higher monitoring quality (Gao, 2003) by employing

the adjacent nodes in order to discern the rightness

of local data. When a sensor malfunction appears

and the hardware redundancy is lost, the problems

can be solved using the analytical redundancy.

Redundancy increases data accuracy, system

reliability and sensor network security to provide

protection against service interruptions.

The rest of the paper is organized as follows.

Section 2 presents the related work in the domain.

Section 3 contains our strategy for perimeter

protection. Section 4 describes a case study for our

security strategy. Section 5 presents the conclusions.

2 RELATED WORK

There is relatively little work in the area of securing

sensor networks based on redundancy. A useful

survey for initiation in the domain of wireless sensor

network is presented in (Akyildiz, 2002).

In (Nowak, 2003) a technique for edge detection

of a phenomenon within a wireless sensor network is

proposed. The approach involves a hierarchical

processing strategy, where nodes collaborate, into a

non-uniform rectangle, adapted to the phenomenon

partition of the sensor field.

Research into authentication and confidentiality

mechanisms of sensor network protocols have been

started in order to identify the problems and to

propose technical solutions (Avancha, 2003),

(Intanagonwiwat, 2000). Some threats to these

applications were identified and a security model

operating on the base station level was proposed.

The application mentioned requires mitigation

against traffic analysis, relying solely on broadcasts

of end-to-end encrypted packets. Nodes adjacent to

the base station are utilised as intermediary hops.

The model corrects some classes of aberrant node

behaviour.

Using wireless sensors networks for tracking

moving objects is discussed in (Cao, 2005), where

an analysis of their performances is done. The

authors provide analytic formulae for the mean delay

until a target is detected, when moving on a straight

line at a constant speed. The authors consider a

system model where sensors are randomly

distributed within a field of interest, with each

sensor having identical sensing areas that follow the

unit disk model.

In (Clouqueur, 2002) the authors propose

collaborative detection models, where sensors

collectively arrive at a consensus about the presence

of a target. Sensors are assumed to be randomly

deployed within the field of interest and the sensing

capability of each sensor is assumed to decay with

distance, with all sensors having identical sensing

areas. They formulate the target detection problem

as an unauthorized traversal problem and propose

deployment strategies for minimizing the cost of the

network that achieves the desired target detection

probability.

These highly localized results of redundancy in

sensor networks can be aggregated by methods such

as (Xu, 2001) to provide higher data reliability for

requesting applications such as event/target

detection (Li, 2002), (Brooks, 2003).

In (Aslam, 2005), a network with binary sensors

is used for tracking a moving object. This is an

elementary case for our solution of using a

perceptron as the model for a binary sensor network.

3 PROPOSED STRATEGY

3.1 Sensor Network Assumptions

We make the following assumptions related to the

sensor network:

a) The sensor network is static, i.e., sensor nodes

are not mobile; each sensor node knows its own

location.

b) The sensor nodes are similar in their

computational and communication capabilities and

power resources to the current generation sensor

nodes. Moreover, because they have to sense if an

intruder is in their proximity, they can provide only

two values, which we assumed to be 0 (for

inexistence of an intruder in their proximity) and 1

(for existence of an intruder in their proximity).

c) We rely on wireless cellular network

architecture (Feng, 2002). In this architecture, a

number of base stations have already been deployed

within the field. Each base station forms a cell

around itself that covers part of the area. Mobile

wireless nodes and other appliances can

communicate wirelessly, as long as they are within

the area covered by one cell.

d) The base station, sometimes called access

point, acting as a controller and as a key server, is

assumed to be a laptop class device and supplied

with long-lasting power. We also assume that the

base station will not be compromised.

With the purpose of solving the problem of a

reliable perimeter protection, we rely on two very

WINSYS 2007 - International Conference on Wireless Information Networks and Systems

important properties: a) inherent redundancy, which

is an important natural feature of sensor networks;

and b) the determinism of the measured values

provided by sensors related to their past recordings.

3.2 Redundancy in Sensor Networks

and Its Benefits

One important natural feature of the sensor networks

employed by our strategy is the inherent

redundancy. We use both hardware and analytical

redundancy in order to increase the reliability of our

perimeter protection approach.

Hardware (physical) redundancy ensures the

reliability in sensor networks (Gao, 2003),

(Clouqueur, 2001) and implies the use of

supplementary sensors (deployed in the field, due to

the necessity of covering the area in case of

malfunctioning of some sensor nodes) and selection

of data that appears similarly on the majority of

sensors.

Analytical (functional) redundancy is based on

the determinism of the measured values provided by

sensors. The information from different sensors is

built on the fact that actual sensor value is related

with past values provided by the same sensor. The

use of analytical redundancy is done through a

process of comparison between the actual sensor

value and the expected/estimated sensor value. This

approach is based on a mathematical model that can

predict the value of one sensor by taking into

consideration the past and present values of

neighbouring sensors or of the implied sensor itself.

The computation implied in this approach is done at

the base station level (laptop class device), where all

requirements are satisfied.

3.3 Knowledge Based System for

Reliability Improvement

Our strategy to improve the reliability of the data

provided by the perimeter protection sensor network

relies on the knowledge-based system (KBS)

presented in figure 1, which contains four

components: a) Confidence Weighted Voting

(CWV) Block, b) Neural Network Block, c)

Knowledge Base Block, and d) Decision Block.

Figure 1: Knowledge-based system architecture.

3.3.1 Confidence Weighted Voting Block

This component relies on hardware redundancy and

is based on a variant of Majority Voting algorithm

(MV) known as Confidence Weighted Voting

(CWV) (Sun, 2005). This algorithm gives higher

weights to those sensors that are more likely to be

correct (i.e. with higher confidence of correctness).

The confidence value of each sensor can be

determined in a distributed manner by comparing its

sensing results with its sensing neighbours that share

overlapping coverage area. The confidence value of

sensor i is then defined as:

∑

⋅Δ

ijij

iconf

)(

(1)

where m represents the total number of the sensors

within the sensor network,

⎩

⎨

⎧

=Δ

resultsamethereportjandisensorif;1

resultsdifferentreportjandisensorif;0

(2)

and

⎩

⎨

⎧

overlappedisjandisensoroferagecovtheif;1

overlappe

notisjandisensoroferagecovtheif;0

(3)

The reliable value, obtained using CWV

algorithm for sensor S, having the in-field position

represented in Cartesian coordinates by the pair

(x,y), is the value

}1;0{k

∈

corresponding to:

() () ()

∑

δ=

jkj

y,xCjconfmaxy,xCWV

(4)

where

⎩

⎨

⎧

=δ

kisjsensorfromvaluereporttheif;1

knotisjsensorfromvaluereporttheif;0

(5)

()

⎩

⎨

⎧

jsensorbyeredcovis)y,x(intpoif;1

jsensorbyeredcovt'isn)y,x(intpoif;0

y,xC

(6)

This CWV Block is an active block in our

strategy only for sensors included in the coverage

zones of other neighbouring sensors, for example

sensor B from figure 2.

KNOWLEDGE BASED SYSTEM FOR RELIABLE PERIMETER PROTECTION USING SENSOR NETWORKS

Figure 2: Sensor coverage diagram.

3.3.2 Neural Network Block

In order to assure a higher reliability for the

information provided by the sensor network, even in

the case of low hardware redundancy, we developed

a neural network structure that provides an estimated

value for each sensor, based on the past values

provided by adjacent sensors. This estimate is

compared with the actual sensor value deciding if

this actual value is reliable or not. The neural

network is based on a perceptron, with a number of

binary neurons equal to the number of the network

sensors. The sensor network is perceived in a static

and also a dynamical way. Each neuron is

considered as a binary model for a sensor. It receives

at the sampling moments the weighted and biased

iterative values of the adjacent sensors (neurons) and

computes the estimates. A relevant architecture of

this block is depicted in section 4.

3.3.3 Knowledge Base Block

Based on some assumptions and on past in-field data

concretised in valuable rules, a knowledge base is

established. This knowledge base includes

information like: a) possible values of intruder’s

speed in the sense that detecting an intruder with a

speed higher than a limit value must not be

considered; b) the impossibility for an intruder to be

detected by an inside sensor until the intruder’s

detection by an outside sensor has been reported.

This knowledge base is used only for validation of

the results provided by CWV and Neural Network

Block.

3.3.4 Decision Block

The Decision Block is implemented in our strategy

by the following pseudo-code:

For (every moment t and every sensor S)

{

/* follows the implementation of

/* Neural Network Block

Result1(S,t)=ComputeNNB(past values

of neighbouring sensors)

If (hardware redundancy is present)

then

{

/* follows the implementation of

/* CWV Block

Result2(S,t)=ComputeCWV(actual

values from the sensors)

ReliableResult(S,t)=Validate1(

Result1(S,t),Result2(S,t),

rules from Knowledge Base)

}

else

{

ReliableResult(S,t)=Validate2(

Result1(S,t),

rules from Knowledge Base)

}

4 CASE STUDY

In this section, static and dynamical models for the

sensor network are proposed, based on the possible

trajectories of a strange object between sensors. A

basic structure of the perceptron implementing the

static and dynamical models of the sensor network is

developed, trained and tested.

Let us consider a field of interest with NxM

binary sensors for perimeter protection. Each sensor

from the field has other 8 adjacent sensors S

adj,i

i=1,…,8, as it is illustrated in figure 3. A cell with 9

sensors is taken into consideration.

A strange object could pass through the cell by

many different directions, each with two senses:

Di,j, i=1,…16, j=1, 2 and combinations of them.

The static model for the sensor network is

illustrated as follows. If an object is situated in a

point P

, at an intersection of two directions, a set of

sensor value results: S

=[S

adj1

, S

adj2

, S

adj3

, S

adj4

, S

adj5

, S

adj6

, S

adj7

, S

adj8

]

. For example, if the object is

in the point of intersection between directions D3

and D6 the set of sensor values is [0, 0, 0, 1, 1, 0, 1,

1, 0]. At each intersection of two directions four

adjacent sensors of the intersection are activated,

based on the hardware redundancy. A table of sensor

value sets is created for all points of intersections.

The dynamical model of the sensor network is

illustrated as follows. The values of the sensors are

available at the sample times S(kh), where h is the

sample period. When a strange object passes trough

the network the sensors are activated one after

WINSYS 2007 - International Conference on Wireless Information Networks and Systems

another. So, for a dynamical description of object

movement between sensors, a train of impulses

results. As an example, in figure 4 we represented

the impulse train for trajectory D3,1-D7,1-D4,2.

Figure 3: The cell structure.

Figure 4: Examples of sensor impulse trains.

For a binary sensor modelling we use a neuron

with two values 0 and 1. The neuron is trained to

learn the impulse trains for all the possible

trajectories between sensors. The sensor network

may be modelled as a perceptron with N×M binary

neurons, applying at the neuron inputs the measured

values from the adjacent sensors.

The structure of the neuron for the dynamical

model is presented in figure 5,

Figure 5: Neuron structure for the dynamical model.

where:

∑∑

−

+−=

iadjj,i

p)jt(Swa

)a(fS

(7)

For the static position the following relation

defines the neural model:

∑

−

iadji

pSwa

(8)

The neuron is using the hard-limit transfer

function f

(a), which returns 0 or 1. Each input S

adj-

(t-j) is weighted with an appropriate weight w

i,j

i=1,…8, j=1, 2. The sum a of the weighted inputs is

sent to the hard-limit transfer function f

(a), which

also has an input with a value equal to 1, biased by

p. The neuron produces a result, based on the

measured values provided by its adjacent sensors.

The hard-limit transfer function gives the perceptron

the ability to classify input vectors by dividing the

input space into regions. Specifically, outputs will be

0 if the net input a is less than 0, or 1 if the net input

a is 0 or greater.

We can estimate the value of the sensor S

at the

moment t, based on the measured values of the

adjacent sensors at the previous two time moments

(t-1) and (t-2).

A supervised learning rule is used as a procedure

to modify the appropriate values of the weights w

and bias p of the perceptron (Hagan, 1996). The

training of the perceptron is made on all possible

trajectories through the sensor network, the

behaviour being summarized by a set of input-output

pairs (u;y) = (S

adj,1

, … , S

adj,8

; S

(t)). The

corresponding target y of the perceptron is formed

by the values of the sensor S

. The objective of the

neural network training is to reduce the error ε,

which is the difference between the target vector and

the neuron response (the estimate):

S −=ε

(9)

The desired changes to the perceptron's weights

Δw and bias Δp are calculated, given an input vector

u and the associated training error ε:

KNOWLEDGE BASED SYSTEM FOR RELIABLE PERIMETER PROTECTION USING SENSOR NETWORKS

ε+=

oldnew

uww

(10)

The above perceptron rule is proven to converge

on a solution in a finite number of iterations.

The error obtained after iterative trainings is

presented in figure 6.

Figure 6: The training error.

The neural network was tested with impulse

trains as test sets. The output accurately estimates

the impulse trains for simulated trajectories.

An important result is that the neural network

could be generalized for different possible

trajectories. If the sensor node A is attacked, it is

possible for its output value S

to be different from

the estimate. So, the sensor’s estimated value,

predicted by the neural network, differs from the

actual value of the malicious sensor A, proving that

something wrong happened to sensor A. In these

circumstances, the decision block will exclude the

sensor A from the network.

5 CONCLUSION

The goal of our research was to design a secure

architecture for a sensor network used for perimeter

protection. For this, we used a knowledge-based

system based on hardware and analytical

redundancy. Considering the detection of anomalies

and intruders in binary sensor networks to be a very

important issue, we relied on two coupled

stratagems: a) a CWV based algorithm; and b) a

perceptron predictor based on the past values of

neighbouring sensors to solve this problem. After

detection, the sensor network can take decisions to

investigate, find and remove malicious nodes if

possible. Being localized on a base station level,

with a reduced amount of computation our method is

suitable even for large-scale sensor networks.

REFERENCES

Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci,

E., 2002. Wireless Sensor Networks: A Survey. In

Computer Networks, 38(4), March.

Aslam, J., Butler, Z., Constantin, F., Crespi, V., Cybenko,

G., Rus, D., 2005. Tracking a Moving Object with a

Binary Sensor Network. In Proceeding of the 1

International Conference on Embedded Networked

Sensor Systems.

Avancha, S., Undercoffer, J., Joshi, A., Pinkston, J., 2003.

Secure sensor networks for perimeter protection. In

Computer Networks 43, Elsevier Press.

Brooks, R., Ramanathan, P., Sayeed, A., 2003. Distributed

target classification and tracking in sensor networks,

In Proceedings of the IEEE, vol. 91, no. 8.

Cao, Q., Yan, T., Abdelzaher, T., Stankovic, J., 2005.

Analysis of Target Detection Performance for

Wireless Sensor Networks, In Proceeding of the

International Conference on Distributed Computing in

Sensor Networks, CA.

Clouqueur, T., Ramanathan, P, Saluja, K.K., Wang., K.C.,

2001. Value fusion versus decision-fusion for fault

tolerance in collaborative target detection in sensor

networks, In Proceedings of Fusion 2001, Montreal.

Gao, Y., Wu, K., Li, F., 2003. Analysis on the

Redundancy of Wireless Sensor Networks, In ACM

WSNA Proceedings, San Diego, USA.

Feng J., Koushanfar F., Potkonjak M., 2002, System-

Architectures for Sensor Networks Issues,

Alternatives, and Directions, In Proceedings. of the

2002 IEEE International Conference on Computer

Design (ICCD’02), Freiburg, Germany.

Hagan, M.T.; Demuth, H.B.; Beale, M.H., 1996. Neural

Network Design, PWS Publishing, Boston.

Intanagonwiwat, C., Govindan, R., Estrin, D., 2000.

Directed Diffusion: A Scalable and Robust

Communication Paradigm for Sensor Networks. In

ACM International Conference on Mobile Computing

and Networking (MOBICOM’00).

Li, D., Wong, K., Hu, Y.H., Sayeed, A.M., 2002.

Detection, classification, and tracking of targets in

distributed sensor networks. In Signal Processing

Magazine, IEEE, Vol. 19, No. 2.

Nowak, R., Mitra, U., 2003. Boundary Estimation in

Sensor Networks: Theory and Methods. In

Proceedings of the First International Workshop on

Information Processing in Sensor Networks.

Pottie G.J., W.J. Kaiser, W.J., 2000. Wireless Integrated

Network Sensors. In Communications of the ACM,

vol. 43.

Sun, T., Chen, L.-J., Han, C-C., Gerla, M., 2005, Reliable

Sensor Networks for Planet Exploration, ICNSC.

Xu, Y., Heidemann J., Estrin, D., 2001. Georgraph-

informed Energy Conservation for Ad Hoc Routing, In

Proceeding of ACM Mobicom, Rome, Italy.

WINSYS 2007 - International Conference on Wireless Information Networks and Systems