NETWORK LEARNING AND TRAINING OF A CASCADED LINK-BASED FEED FORWARD NEURAL NETWORK (CLBFFNN) IN AN INTELLIGENT TRIMODAL BIOMETRIC SYSTEM

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.9, No.6, November 2018
DOI : 10.5121/ijaia.2018.9603 27
NETWORK LEARNING AND TRAINING OF A
CASCADED LINK-BASED FEED FORWARD NEURAL
NETWORK (CLBFFNN) IN AN INTELLIGENT
TRIMODAL BIOMETRIC SYSTEM
Benson-Emenike Mercy E.1
and Ifeanyi-Reuben Nkechi J.2
1
Department of Computer Science, Abia State Polytechnic Nigeria.
2
Department of Computer Science, Rhema University Nigeria.
ABSTRACT
Presently, considering the technological advancement of our modern world, we are in dire need for a
system that can learn new concepts and give decisions on its own. Hence the Artificial Neural Network is
all that is required in the contemporary situation. In this paper, CLBFFNN is presented as a special and
intelligent form of artificial neural networks that has the capability to adapt to training and learning of new
ideas and be able to give decisions in a trimodal biometric system involving fingerprints, face and iris
biometric data. It gives an overview of neural networks.
KEYWORDS
CLBFFNN, Learning, Training, Artificial Neural Network, Trimodal, Biometric System
1. INTRODUCTION
With ever growing field of artificial intelligence, there are numerous intelligent systems
providing algorithms to solve any particular problem which require human intelligence [1][2].
These artificial intelligent systems led to development of a more creative, knowledgeable and
exceptional system that helps computers learn [3][4]. CLBFFNN is an intelligent Artificial
Neural Network having interconnecting network of artificial neurons which uses some algorithms
and the mathematical formulas for the processing of the data or information on the computational
purposes. It uses modern technique of “Learning with examples” which the corresponding
response seems to come from a kind of natural intelligence rather than using Artificial
Intelligence.
2. OVERVIEW OF NEURAL NETWORKS
Neural Networks (NNs) are an abstraction of natural processes, which can be likened to the
operations and functions of biological neurons [5]. NN has been incorporated into a broad and
integrated computational model of adaptive systems, since early 1990s, in other to utilise their
learning power and adaptive capabilities, they have the ability to learn and adjust to new
incoming patterns, create adaptation to a dynamic environment much more effective and efficient
[2]. Combined with evolutionary algorithms (EAs) they can be regarded as a general framework
for adaptive systems i.e. systems that excel in changing their architectures and learning rules
adaptively without human intervention.

28
2.1 Types of Neural Networks (NN)
Neural Networks are broadly divided into Feed-forward and Recurrent or feedback networks.
Multilayer perceptron, single-layer perceptron and Radial Basis Function nets are grouped under
Feed-forward networks while Hopfield network, Competitive networks, Kohonen’s SOM and
ART models are grouped under Recurrent or feedback networks[6].
2.1.1 Multilayer Perceptron [MLP]
MLP is one of the most largely used neural networks [7]. MLP, given two sets of data (i.e.
input/output pairs) is capable of developing a precise nonlinear mapping, using a learning
algorithm by adjusting the network weights. It has been proved that a two-layer MLP can
effectively approximate any nonlinear mapping [8].MLP training uses majorly Back-propagation
(BP) algorithm [7], where a steepest descent gradient approach and a chain-rule are adopted for
back-propagated error correction from the output layer [1]. Significant efforts have been put into
improving the speed of convergence, generalization performance, and the discriminative ability
of MLP. To accelerate the BP algorithm, several heuristic rules have been proposed to adjust the
learning rates or modify error functions [9]. Acceleration of training MLP can also be achieved
by the use of other modifications to the standard BP algorithm, such as conjugate gradient BP,
recursive least-square-based BP, and the Levenberg-Marquardt algorithm. To verify the
generalization ability of MLP, the independent validation method can be used by dividing the
available data set into a number of sub sets for training, validation and testing [10]. To improve
the discriminative capability of MLP when applied to a classification task, a discriminative MLP
learning rule was proposed which is more suitable for pattern classification tasks [11].
2.1.2Hopfield Neural Networks
Associative memory networks include linear associative memory and Hopfield associative
memory. Linear associative memory is an effective single-layer network for the retrieval and
reduction of information [12]. Given a key input pattern U = [u1, u2, ...,uk] and the corresponding
output V = [v1, v2, ..., vk], associative memory learns the memory matrix P to map the key input
u1 to the memorized output 1 . There are a number of ways to estimate the memory matrix. One
estimate of the memory matrix W is the sum of the outer product matrices from pairs of key input
and memorized patterns
= (1)
To further reduce the memorized error, an error correction approach has been introduced to
minimize the error function
E(P) = ∥ - P ∥2
(2)

Hopfield associative memory is a nonlinear content-addressable memory for storing information
in a dynamically stable environment [13]. The Hopfield network is a single-layer recurrent
network which contains feedback paths from the output nodes back into their inputs. Given an
input x(0), the Hopfield network iteratively updates the output vector by
U(n + 1) = f ( ( ) – θ), (3)

29
until the output vector become constant, where f(·) is the activation function. Associative memory
is able to deduce and retrieve the memorized information from possibly incomplete or corrupted
data.
2.1.3Modular Networks
There are strong biological and engineering evidences to support the fact that the information
processing capability of NNs is determined by their architectures [4]. Much work has been
devotedto finding the optimal or near optimal NN architecture using various algorithms,
including EAs [14]. However, many real world problems are too large and too complex for any
single ANN to solve in practice. There are ample examples from both natural and artificial
systems that show that an integrated system consisting of several subsystems can reduce the total
complexity of the entire system while solving a difficult problem satisfactorily. NN ensembles
adopt the divide-and-conquer strategy. Instead of using a single large network to solve a complex
problem, an NN ensemble combines a set of ANNs that learn to decompose the problem into sub-
problems and then solve them efficiently. An ANN ensemble offers several advantages over a
monolithic ANN [15]. First, it can perform more complex tasks than any of its components (i.e.,
individual ANNs in the ensemble). Second, it can make the overall system easier to understand
and modify. Finally, it is more robust than a monolithic ANN, and can show graceful
performance degradation in situations where only a subset of ANNs in the ensemble performs
correctly. There have been many studies in statistics and ANNs that show that ensembles, if
designed appropriately, generalize better than any single individuals in the ensemble. A
theoretical account of why and when ensembles perform better than single individuals was
presented in [16].
2.2Overview of Artificial Neural Networks
Structurally, ANN is a computational model that adopts human nervous system technique with
back propagation algorithm for training-testing phase [17][18]. It is an adaptive system capable of
transforming its arrangement according to input and output sequence that flows in the network
[19]. Neural network provides rational success [20], removes the shortcomings of traditional
methods and can be trained from examples. Its characteristics include robustness, less data
requirement, self-organizing, fast computation, adaptive learning, parallelism, ease of software
and hardware implementation, high rate of error tolerance and broad view [1][2]. However, ANN
is prone to over-fitting, has an empirical nature of model development and a greater
computational problem. The following are the various types of artificial neural networks: feed-
forward neural networks, radial basis function (RBF) networks, Kohonen self-organizing
networks, recurrent networks, stochastic neural networks, modular neural networks, dynamic
neural networks, cascading neural networks, and neuro-fuzzy networks [21].
Artificial neural network is made up of a collection of neurons that are connected to each other
mainly for processing of information [22][3]. Figure 1 shows an instance of a neuron. Neurons
have many weighted inputs such as (x1, x2, x3… xn) with their corresponding weighting, (w1j, w2j,
w3j… wnj), according to its importance. Every input is a scalar quantity and stands for the data.
For instance, in a face recognition system having 12x12 pixel image, a 144 input representation
ranging from X1 to X144 and weightings of w1 to w144, equivalent to 144 pixels in the input image,
will be achieved.

30
Figure 1: A Neuron[18]
Artificial Neural Network structure comprises of input layers, hidden nodes, output nodes and an
activation function, as seen in Figure 2.
i) Number of input layers: The input layers offer broad view ability to the network.
ii) Number of hidden nodes: The best method for selecting the most appropriate
number of hidden neurons is not in existence. One could use Kolmogorov theorem
for computing number of hidden neurons. This is given as 2i+1 neurons, where i is
the number of inputs.
iii) Number of Output Nodes: Single output neural network is preferable to networks
that produce numerous outputs because they produce more superior results.
iv) Activation or Transfer Function: This is a mathematical formula that determines
the output of a processing node. Every input node computes and adds activation to its
net input. The essence of the transfer function is to ensure that the output does not
assume a large value because this can obstruct neural network training and adversely
affect its performance.
Figure 2: The Structure of an ANN[1]

31
The model of a neural network model is of three types:
i. Feed-Forward Network: Here, there is a sequential connection of neurons from one
layer to another. There is a continuous connection of neurons and cannot connect
backwardly [6].
ii. Recurrent Network: In a recurrent network, there must be at least one backward
connecting loop or feedback loop. It is mainly used for memory relationship and
computation of optimization [1].
iii. Self-Organization Networks: This network is based on unsupervised learning. In this
network target output is not known to the network and is mainly used for cluster analysis.
[23],[24].
2.3 Network Training Types
The two permissible actions that can take place in a neural network are training and no-training.
A non-training scenario occurs when the network is only run with a specific amount of weights.
In a training scenario, we first run the network with a given set of weights; modify some or all the
weights to obtain an entirely new set of weights, then run the network again with the new set of
weights. This procedure continues until some pre-set target is achieved. The network assumes the
initial weight assigned to it and subsequently adjusts it based on the computed strategy. After this
training, the network will be expected to exhibit ‘intelligence’ by performing what it has been
trained to do, based on the input patterns. The obtained information is stored in secure memory or
database, as in the human brain.
Every neural network must be trained by means of parameter adjustment. Neural networks can be
trained using various methods such as supervised, unsupervised and reinforcement training [17].
a) Supervised Training: This kind of training involves presenting to the neural network a tester
set of inputs and their corresponding desired outputs. After training, when any given set of input
is given to the network, the neural network will in turn map the said input to its equivalent trained
output.
Weights are determined to allow the network to produce answers as close as possible to the
known correct answers. The back-propagation algorithm belongs into this category.
b) Unsupervised Training: This method is used when the output or outcome for a given set of
inputs is unknown. This type of training does not require a correct answer associated with each
input pattern in the training set. The neural network, in this case, is fed with enough data and is
allowed to try, learn, organize, modify network’s weights and biases and derive conclusions itself
without outside intrusion[3]. Explores the underlying structure in the data, or correlations
between patterns in the data, and organizes patterns into categories from these correlations. The
Kohonen algorithm belongs to this category.
c) Reinforcement Training: It is a learning technique that stores insufficient information
concerning the target output. It is similar to supervised learning; but gives less accurate result
[19].
3. AN OVERVIEW OF CLBFFNN
CLBFFNN is also a Feed-Forward Neural Network. Three processing layers are encompassed in

32
a feed-forward neural network. They include the input layer which is the first layer, the output
layer which is the last layer and the hidden layer which is the middle layer. Each processing
element makes self-determining computations on data that it receives, calculates the weighted
sum of its inputs and conveys the results to another layer [25]. These processing elements will
jointly produce the output from the network. These elements are similar in operations to human
brain neurons, hence, referred to as artificial neurons (see Figure 3).
Figure 3: A Brain Neuron[2]
In the output layer, authentication of a neuron output is done using a threshold function. Synapses
are connections between neurons. The artificial neurons and the synapses are respectively
denoted by nodes and edges of a directed graph, as seen in figure 4. Each synapse has an assigned
accompanying weight and the choice of these weights affects the operation and the classification
of the neural network.
CLBFFNN is an intelligent multi-biometric system classifier. When a multi-biometric system is
trained to have an artificial intelligence to be able to take decisions based on several inputs, then
such system is said to be ‘intelligent’. This is achieved by the application of an Artificial Neural
Network (ANN), Fuzzy logic, Neuro-fuzzy method etc. System accuracy cannot be fully
achieved when only one matching algorithm is applied [26].
After pre-processing, feature extraction and quality assessment, the next stage is verification or
authentication which is done by a CLBFFNN classifier. CLBFFNN is an ordered cascade of two
neural networks CLBFFNN(1) and CLBFFNN(2), as seen in figure 6. Their designed
arrangement guarantees reduction in computation cost, increases system’s detection accuracy and
efficiency. The network components are interconnected with layers and flow of information via
these network components is not interrupted in any way. Apart from learning and training of
CLBFFNN with back propagation algorithm, it makes decisions on final association between its
inputs. It is trained with a back propagation algorithm.
Functional
signals
Figure 4: Functional and Error Signals in CLBFFNN

33
CLBFFNN(1) is a neural network that uses Back Propagation algorithm. It is composed of an
input layer, a single hidden layer and an output layer with a sigmoid activation function. The
feature vectors from pre-processed multi-biometric data form the input into CLBFFNN(1). The
CLBFFNN(1) has three cascaded stages as seen in figure 5.
The first stage ensures that the feature patterns that are entering into the network are life-scanned
features, i.e. facilitating further liveness detection. The second stage filters biometric noises such
as the non-face from face feature, distorted or incomplete minutiae from fingerprints and blurred
iris pattern from iris images. Lastly, the third stage determines the image width and height and
calculates the time complexity of the biometric features. These sub windows are passed through
each section of image and minimizes the false positive rate (i.e. non-faces for example) and
detects the face. This method reduces false rejection and false acceptance errors to their barest
minimum and increases system verification accuracy. The simulated result from CLBFFNN(1) is
fed into CLBFFNN(2) for the function approximation.
Network learning and training, template matching, comparison and decision making take place in
CLBFFNN(2). Output from CLBFFNN(1) is approximated since it is connected with
CLBFFNN(2). CLBFFNN(2) is a universal approximator and has a very compact topology.
CLBFFNN(2) is also composed of an input layer, a single hidden layer and an output layer with
fast locally regulated neurons; and also has a feed-forward design.The output from CLBFFNN(2)
determines the recognition result.
T
T
FF
1 2
CLBFFNN(2) for Further
Processing
Sub-windows containing Biometric Noise such as non-face,
distorted minutiae and blurred images are rejected
F
T
All Sub-windows
Figure 5: The Three Cascaded Stages of CLBFFNN(1)
[Source: [27]]
3

34
3.1 CLBFFNN Learning Using Back Propagation Algorithm
Back propagation algorithm is a supervised or monitored learning process used for the design of a
Feed Forward Neural Network (FFNN). Learning in CLBFFNN is accomplished by back
propagation algorithm and it has to do with the modification of weights and other networks
parameters to ensure proper training of the biometric features. Learning is done here by giving
instances. When instances of what is to be achieved; is giving to the algorithm, it obtains the
necessary output for the particular input through series of network’s weight adjustments and
training. Back propagation algorithm in
CLBFFNNhas two passes for each input-output vector pair:
i)forward pass: Here, the input propagates through the net and an output is produced.
Biometric Feature
Vector1
CLBFFNN (1)
CLBFFNN (2)
Biometric Feature(1)
Decision Module
Authentication
Classifier
Fusion
Decision Module
Decision Module
Matching
Score Normalization
Biometric Feature
Vector2
Vector
Biometric Feature
Vector3
Figure 6: CLBFFNN in a Trimodal Biometric
System

35
ii)backward pass: Here, the difference (error) between the desired and obtained output is
evaluated with an error function. This error function is minimised towards the weights of the net,
starting backwards from the output layer to the input layer. Back propagation algorithm is easy to
implement and is capable of solving all manners of problems.
Network learning process begins by submitting input patterns, whose output has already been
determined and known by the network. The weighted sum of the neurons is calculated and
transferred by each layer to the next layer following it through an activation function. The
acquired outcome is used as an input to the subsequent layer. Finally output layer generates an
output pattern which is then compared to the target pattern. The input and its corresponding target
are called a Training Pair. A mean squared error is computed based on the dissimilarity between
the result obtained and the expected output. The learning rate of the network and its efficiency is
measured by the obtained dissimilarity or error [28]. The back propagation algorithm aims at
minimizing this error to its lowest point. The following is a back propagation algorithm designed
for this purpose:
BP Algorithm
1. Give initial values to weights, learning rate, momentum and error precision.
2. Enter the input vector and the desired output associated with it.
3. Compute the weighted sum and the output of the network.
4. Compute the mean square error. This is obtained by evaluating the contrast between the result
obtained and the expected output.
5. Adjust the weights by back propagating them from the output layer [29].
3.2CLBFFNN Training and Learning
Network training and learning go hand in hand. At the commencement of training and learning,
the acquired feature matrices for the biometric images are used as input data set to the
CLBFFNN. This is done through a supervised training algorithm. After the training, the trained
pairs are supplied to the neural network. The training of CLBFFNN is done using both positive
and negative set of the biometric data because after the network training, it is expected to
recognize negative set of data such as non-face, distorted and blurred image features. Therefore,
the input to CLBFFNN could be aggregates of face and non-face training data; correct and
distorted or incomplete minutiae for fingerprint; or correct and distorted or blurred image features
for iris.
The network will be trained to provide an output ‘positive’ for good and complete image features
of biometric features while with negative, distorted or blurred image features, the output will be
‘negative’. We fix a target output for each feature based on the average feature value of each of
the images. Then a program is developed to train the neural network for biometric feature
recognition.
When the training begins, all the weights of the neurons are aggregated and stored in a weight
file. After the first set of data is trained, an error is calculated, which is based on the difference
between the computed output and the target output, as seen in figure7. This aims at reducing this
error and ensuring that the final output obtained is equal to the target output; by series of network
training and weight adjustments.

36
3.2.1 The Back Propagation Algorithm in CLBFFNN Training
Back Propagation algorithm is adopted for CLBFFNN training. It is a feed-forward supervised
training network with the capability of automatically updating network weights using defined
update procedures. This is done by taking an incomplete derivative of the error function to
examine how each of the network weight contributes to the error. The learning and training is
done from one area of the neuron to another.
There are different stages in training CLBFFNN with back propagation algorithm, as seen in
Figure 7. The first stage is the initialization of all the network weights which involves assigning
initial values to the network weights in the range of -1 ≥ x ≤ 1. The second stage involves the
feeding of the input pattern into the network through the input neurons. The input neurons send
the patterns to each of the hidden neurons; which will compute the activation function and
transfer the result to the output neurons. The output neurons will calculate the activation, the
output and the neuron error. The error is computed by the difference between the set target and
the neuron output, as shown in equation (4).
Neuron Error = Set Target – Neuron Output (4)

37
Figure 7: CLBFFNN Training
N
Y
Initialize the Network
weights, NeuronError and
TotError
Apply and Train the
initial set of train
Pattern
TotError = TotError +
│NeuronError│WHO(new) = WHO + (ErrorO
x OutputH)
Is
TotError<Tar
getError?
More Pattern to
be Trained?
Y
Load Next
Pattern
And train
N
Calculate the Error of each Neuron
NeuronError = SetTarget -
NeuronOutput
Stop
Start

38
In the third stage, the NeuronError gotten is propagated back to the hidden layer neurons (O) to
obtain the output neuron error, as given by equation (5).
ErrorO= OutputO (1-OutputO)(TargetO – OutputO) (5)
The next stage has to do with weight adjustment, in which case, the neuron weights assume new
values, as given by equation (6).
WHO(new) = WHO + (ErrorO x OutputH) (6)
Where
H = The hidden layer neuron,
O = The output neuron,
WHO = The weight of the synapse between the hidden layer and the output
layer, and which stands for the initial neuron weight,
WHO(new) = The new or trained neuron weight.
After the weight adjustment, the hidden neuron errors are calculated. Finally, the total error is
compared with the target error; if it is less than the target error the training stops, otherwise the
whole process repeats. The network will be well trained so as to recognize input patterns and give
out output accordingly.
3.2.2 Face Learning and Training in CLBFFNN
Figure 8 is the flowchart for learning and training procedure for face feature in CLBFFNN.
Where
= input layer neuron vector
= hidden layer neuron vector
P = output layer neuron
= the weight matrix between the input and the hidden layer
= the bias of the hidden layer neurons
= weight matrix connecting the hidden and the output layers
b0 = the bias of the output layer neurons
= the error vector for output neurons
= the error vector of each hidden layer neuron
= the desired output vector
LR = learning rate
MF = momentum factor
The sigmoid activation function is defined by
F( ) = ! (" )
(7)

39
Figure 8: Flowchart for Face Learning and Training in CLBFFNN
N
o
Ye
Is
MSE< #?
Compute Hidden layer neuron activation
= f(∑ . + )
Output layer Neuron activation
P= f(∑ . + )
Initialize Hidden output weights and bias randomly
Compute
a) Output layer error: = P(1 - P)(P - )
b) Hidden layer error: = (1- )∑ .
Adjust weights and bias
(new) = – LR * * +MF * bo
bo (new) = – LR * *
(new) = – LR * * +MF *
w0(new) = – LR * *
Calculate Error
for (k = 0; k< output; k ++)
{MSE = ∑(' ( − ()2
;
If [(' ( − ()> 0 Break]
MSE =√MSE
Calculate the output of the network and
select the maximum output:
=f(∑ . + )
P= f(∑ . + )
Start
Stop

40
3.2.3 Fingerprint Learning and Training in CLBFFNN
The flowchart in Figure 9 gives the algorithm for the fingerprint learning and training in
CLBFFNN. The following learning rule is adopted in the training of the network:
Let V(x)(a) be the input vector,
X = Mr. V’s fingerprint
A = pattern matrices, 16x16 array (a = 0, 1,2,3…., 256)
G(x)(a) = The target output
Wab and Wbc= The weight vectors
Ixa= unit a input for pattern x.
Ixb=The output of the respective PE which is calculated to be either 1 or 0
using a sigmoid function
Nxb= The weighted Sum
atvb = The activation function
yfb= the hidden threshold weight for bth PEs
C1 = called the spread factors
yIc= output threshold weight for cth output
NTW = new threshold weight

41
Initialize Wab, Wbc and threshold values in the
network with random numbers
Nxb =
∑WabIxa
atvb =
Nxb + yfb
∆Wbc+2C2 ,xcIxbI
xc (1 – Ixc)
-(./012) = C2
Nxc =
∑WbcIxb
atvIc
= Nxc + yIc
Ixc = 1/(1 +
3"24∗./012)
6. ,
Wab= Wab +
∆8ab
∆yf = + C (1 –
,
NTW =
yfb + ∆yfb
Enter input patterns and
desired respective output
∆yIc =
+ C ,x Ix (1 –
NTW =
yI + ∆yI
∆Wab = +aCaIxaIxb(1
– Ixa) Ʃ,xcWbc
-(./09 ) = C Ix (1 –
Figure 9: Flow Diagram for Fingerprint Learning and
Training Using CLBFFNN

42
3.2.4 Iris Learning and Training in CLBFFNN
After iris image normalization, the feature vectors are fed into CLBFFNN for network learning
and training. After series of training and weight adjustments, the network training stops when the
computed error is at its minimum. The system is evaluated by feeding the trained network with a
test data; to prove its efficiency in handling noisy set of data. Figure 10shows the NN learning
and training algorithm for iris feature.
The variables used include:
thd = threshold
S = security band
N = the numbers of characteristics
L1 and L2= learning rates
Stop = stopping criterion
C(i,j) = a comparison code
S(C) = a similarity score
ICC = the set of imposter comparison codes
GCC = the set of genuine comparison codes

43
3.3 Overtraining Check Facility
Adopting the use of the Overtraining Check Facility helps to determine when to stop network
training. This prevents network overtraining (i.e. network being extremely correct). This lowers
the efficiency of the biometric system because it will lack the capability of handling noisy
versions of data. This facility is ensured by having another set of a training data containing
negative attributes (such as distorted or blurred images). This set is called a validation set. It is
not involved in the network training rather acts as a check against network overtraining. After
each training session, the validation set is used to calculate an error. Network training stops when
validation error attains its minimum value, but when the error starts increasing, it then means that
the network is being over-trained.
Figure 4.5: Iris Learning and Training Using CLBFFNN
Feed normalized iris feature vector into
CLBFFNN
Stop =1
Is
S(score)C(i,j)
∈
Stop
Is C(i,j) ∈
GCC?
Initialize stop, N, thd, S
Enter C(i,j)
Zi = (Zi – L1)* (C(i,j) + L1) *
;( , =)
Zi = (Zi + L1)* (C(i,j) - L1) *
>(?, @)
C(I,j) ∈ ICC
Y
Y
N
N
Start

44
3.4 Comparison of CLBFFNN with Other Methods
The comparison of CLBFFNN with other methods is done in MATLAB 7.5
environment.CLBFFNN is seen to have a better recognition performance when compared with
Adaptive Principal Component Analysis (APCA) and Multilayer Perceptron (MLP) methods [30]
as seen in Figure 11. CLBFFNN approach also gives better performance than that of Long & Thai
[31]who used RVM approach, as seen in figure 12.
Figure 11: Performance Comparison MLP, APCA and CLBFFNN
Figure 12: Comparing SVM and RVM Classifiers with CLBFFNN

45
4. CONCLUSIONS
CLBFFNN is an intelligent multi-biometric system classifier and has input layer, hidden layer
and output layer, like every other artificial neural network. It is an ordered cascade of two neural
networks: CLBFFNN(1) and CLBFFNN(2). The use of a Cascaded Link-Based Feed Forward
Neural Network (CLBFFNN) with Back Propagation algorithms offers effective training of the
network with reduced FAR and FRR. Security enhancement of the algorithms is enforced by the
use of liveness checks, biometric data quality checks and overtraining check facilities.
5. ACKNOWLEDGEMENTS
The authors would like to appreciate the reviewers of this paper, for their useful comments and
contributions which added to the quality of this work.
REFERENCES
[1] El-Zoghabi A. A., Yassin A. H., Hussien H. H. (2013). Survey Report on Cryptography Based on
Neural Network.International Journal of Emerging Technology and Advanced Engineering. Website:
www.ijetae.com (ISSN 2250-2459, ISO 9001:2008 Certified Journal, Volume 3, Issue 12, December
2013), 456 -462.
[2] Jogdand R. M. and BisalapurS. S. (2011). Design of An Efficient Neural Key Generation.
International Journal of Artificial Intelligence & Applications (IJAIA), Vol.2, No.1, 60-69, January
2011. DOI : 10.5121/ijaia.2011.2105 60 (PDF). Available from:
https://www.researchgate.net/publication/49612116_Design_of_An_Efficient_Neural_Key_Generatio
n
[3] Dilek S., Çakır H. and AydınM. (2015). Applications of Artificial Intelligence techniques To
Combating Cyber Crimes: A Review. International Journal of Artificial Intelligence & Applications
(IJAIA), Vol. 6, No. 1, January 2015. 21-39
[4] Bagrow, J. P., Lehmann, S., &Ahn, Y-Y. (2015). Robustness and modular structure in networks.
Network Science , 3(04), 509-525. DOI: 10.1017/nws.2015.21
[5] Debashish D. and Mohammad S. U. (2013). Data Mining And Neural Network Techniques In Stock
Market Prediction:A Methodological Review. International Journal of Artificial Intelligence &
Applications (IJAIA), Vol.4, No.1, January 2013 DOI : 10.5121/ijaia.2013.4109, 117 – 127.
[6] Chiang Y., Chang L., Chang F. (2004). Comparison of static-feed-forward and dynamic-feed-
backneural networks for rainfall – runoff modelling. ELSEVIER, Journal of Hydrology 290 (2004)
297–311.
[7] Islam M. J., AhmadiM. and Sid-Ahmed M. A. (2010). An Efficient Automatic Mass Classification
Method In Digitized Mammograms Using Artificial Neural Network. International Journal of
Artificial Intelligence & Applications (IJAIA), Vol.1, No.3, July 2010, 1-13. DOI :
10.5121/ijaia.2010.1301
[8] Khaze S. R., MasdariM. and Hojjatkhah S. (2013).Application of Artificial Neural Networks in
Estimating Participation In Elections. International Journal of Information Technology, Modelling
and Computing (IJITMC) Vol.1, No.3, August 2013.DOI : 10.5121/ijitmc.2013.1303, 23-31
[9] Li Y., Wang K., and Zhang D. (2002). Step acceleration based training algorithm for feed-forward
neural networks. In IntConf Pattern Recognition (ICPR), pages 84–87.
[10] Levin E., Tishby N., and Solla S.A. (1990). A statistical approach to learning and generalization
in layered neural networks. IEEE, 78(10):1568–1574.
[11] JuangB. and Katagiri S. (1992).Discriminative learning for minimum error classification.IEEE
Trans Signal Processing, 40(12):3043–3054.
[12] Lee K. Y. and Sode-YomeA. and Park J. H. (1998).Adaptive Hopfield Neural Networksfor Economic
Load Dispatch. IEEE Transactions on Power Systems, Vol. 13, No. 2, May1998 519-526
[13] Hopfield J. (1982). Neural networks and physical systems with emergent collective computational
abilities. National Academy of Science USA, 79(8):2554 –2558.
[14] Yao X. (1999). Evolving artificial neural networks.Proceedings of the IEEE, 87:1423 – 1447.
[15] Ludmila I. Kuncheva (2004).Combining Pattern Classifiers - Methods and Algorithms.Wiley.

46
[16] Tang E. K., Suganthan P. N., and Yao X.(2006). An analysis of diversity measures.Machine
Learning, 65:247–271.
[17] Krenker A., Bester J. and Kos A. (2011). Introduction to the Artificial Neural Networks, Artificial
Neural Networks - Methodological Advances and Biomedical Applications, Prof. Kenji Suzuki (Ed.),
ISBN:978-953-307-243-2, InTech, Available from: http://www.intechopen.com/books/artificial-
neural-networks-methodological-advances-and-biomedical-applications/introduction-to-the-artificial-
neural-networks.
[18] Shubhangi S. A. and Madhuri A. C. (2012). Artificial Neural Network Controller for Performance
Optimization of Single Phase Inverter.International Journal of Artificial Intelligence & Applications
(IJAIA), Vol.3, No.1, January 2012. DOI : 10.5121/ijaia.2012.3105. 53-64.
[19] Vijaya, S. (2012). A study on the neural network model for finger print recognition.
International Journal of Computational Engineering Research (ijceronline.com)2 (5).
[20] Le Cun, V., Bottou, L., Bengio, Y., and Haffner, P., (2012). Handwritten digit recognition
with a back propagation network. Neural Information Processing Systems, 2: 396-404.
[21] Fausett, L. (1994). Fundamentals of neural networks. New York: Prentice Hall. ACM, New
York, 89: 151– 158.
[22] Schmidhuber J. (2015) Neural Networks. ELSEVIER 61(2015) 85-117
[23] Ritu, M. G. (2014).A review on fingerprint-based identification system.International
Journal of Advanced Research in Computer and Communication Engineering 3(3).Copyright to
IJARCCE www.ijarcce.com 5849.
[24] Hazem, M. E. (2002). Face detection using neural networks and image decomposition.
Lecture Notes in Computer Science 22: 205-215.
[25] Askarunisa, A., Sankaranarayanan, K., Sundaram, R., and Sathick, M. B. (2009).Fingerprint
authentication using neural networks, MASAUM Journal of Computing, 15(2): 234.
[26] Ross, A. and Jain, A. K. (2003). Information fusion in biometrics.PatternRecognition
Letters, 24(13): 2115-2125.
[27] Benson-Emenike, M. E. &Sam-Ekeke Doris C. Trimodal Biometric Authentication System using
Cascaded Link-based Feed forward Neural Network [CLBFFNN]. International Journal of Applied
Information Systems (IJAIS) – ISSN : 2249-0868 Foundation of Computer Science FCS, New York,
USA 12 (5): 7-19 August 2017 – www.ijais.org
[28] Salim L. (2011). A comparative study of back propagation algorithms in financial prediction.
International Journal of Computer Science, Engineering and Applications (IJCSEA),1(4).
[29] Devika, C., Amita, S. and Manish, G. (2013). Recapitulation on Transformations in Neural
Network Back Propagation Algorithm.International Journal of Information and Computation
Technology 3(4): 323-328
[30] Nayak, P.K. and Narayan, D., (2013). Multimodal biometric face and fingerprint recognition using
adaptive principal component analysis and multilayer perception.International Journal of Research in
Computer and Communication Technology, 2(6).
[31] Long, B.T., and Thai, H. L., (2015). Person authentication using relevance vector machine (RVM) for
face and fingerprint.I. J. Modern education and Computer Science, 5, 8-15.
Authors
Benson-Emenike Mercy E. has a doctorate degree and Master’s degree in
Computer Science from University of Port Harcourt, Nigeria. She obtained
her Bachelor of Technology degree [B.Tech] from Federal University of
Technology, Minna, Niger state. She is a lecturer in the Department of
computer Science, Abia State Polytechnic and an adjunct lecturer in
Computer science Department, Rhema University Nigeria and National
Open University of Nigeria [NOUN]. She is a member of Computer
Professionals (Registration Council) of Nigeria (CPN) and Nigeria
Computer Society (NCS). Her publications have appeared in an
international journal and conference proceedings. Her research interests
include Artificial Intelligence, Biometrics, Operating System, and Information Technology.

47
Ifeanyi-Reuben Nkechi J. has a doctorate degree in Computer Science from the
University of Port-Harcourt Nigeria. She obtained her M.Sc. and B.Sc. in Computer
Science from the University of Ibadan, Nigeria and University of Calabar, Nigeria
respectively. She is a lecturer at the Department of Computer Science, Rhema
University Nigeria. She is a member of Computer Professionals (R egistration Council)
of Nigeria (CPN), Nigeria Computer Society (NCS) and Nigeria Women in Information
Technology (NIWIIT). Her research interests include Database, Data mining, Text
mining, Information Retrieval and Natural Language Processing.

NETWORK LEARNING AND TRAINING OF A CASCADED LINK-BASED FEED FORWARD NEURAL NETWORK (CLBFFNN) IN AN INTELLIGENT TRIMODAL BIOMETRIC SYSTEM

More Related Content

What's hot (17)

Similar to NETWORK LEARNING AND TRAINING OF A CASCADED LINK-BASED FEED FORWARD NEURAL NETWORK (CLBFFNN) IN AN INTELLIGENT TRIMODAL BIOMETRIC SYSTEM (20)

Recently uploaded (20)

NETWORK LEARNING AND TRAINING OF A CASCADED LINK-BASED FEED FORWARD NEURAL NETWORK (CLBFFNN) IN AN INTELLIGENT TRIMODAL BIOMETRIC SYSTEM