An Evaluation of Score Level Fusion Approaches for
Fingerprint and Finger-vein Biometrics
Kamer Vishi and Vasileios Mavroeidis
Department of Informatics, SecurityLab
University of Oslo (UiO), Norway
{kamerv, vasileim}@[Link]
arXiv:1805.10666v1 [[Link]] 27 May 2018
Abstract
Biometric systems have to address many requirements, such as large pop-
ulation coverage, demographic diversity, varied deployment environment, as
well as practical aspects like performance and spoofing attacks. Traditional
unimodal biometric systems do not fully meet the aforementioned require-
ments making them vulnerable and susceptible to different types of attacks.
In response to that, modern biometric systems combine multiple biometric
modalities at different fusion levels. The fused score is decisive to classify
an unknown user as a genuine or impostor. In this paper, we evaluate com-
binations of score normalization and fusion techniques using two modalities
(fingerprint and finger-vein) with the goal of identifying which one achieves
better improvement rate over traditional unimodal biometric systems. The
individual scores obtained from finger-veins and fingerprints are combined
at score level using three score normalization techniques (min-max, z-score,
hyperbolic tangent) and four score fusion approaches (minimum score, maxi-
mum score, simple sum, user weighting). The experimental results proved
that the combination of hyperbolic tangent score normalization technique
with the simple sum fusion approach achieve the best improvement rate of
99.98%.
Keywords: multibiometrics, biometric fusion, fingerprint, finger-vein,
authentication systems, identity management, privacy, security
1 Introduction
The most important practical consideration for a biometric system is its accuracy. The
combination of multiple different samples or multiple biometric modalities is a natural
approach for improving the performance of a biometric system. This approach depends
on specific methods for data fusion during the analytical process. Biometric fusion can
be performed at different levels such as sensor level, feature (template) level, score level
and decision level. Currently, the most popular and suitable way of biometric fusion is at
Proceedings of the 10th Norwegian Information Security Conference (NISK 2017), Oslo,
Norway, November 27-29, 2017.
score level since all commercially available biometric sensors and feature extractors do
not provide access to their feature extraction algorithms.
Some of the limitations of the aforementioned unimodal biometric systems are
overcome by combining multiple biometric modalities. These systems are known as
multimodal biometric systems and are more secure due to the integration of multiple
independent pieces of evidence [1]. For instance, the chance of getting a valuable
biometric system increases with the number of involved biometric traits. In addition, the
integration of multiple traits increase security since it is more difficult to spoof multiple
biometric modalities of real users [2]. Advantages of multimodal biometric systems over
unimodal systems have been discussed by Ross et al. [3].
In this paper, we present an evaluation of normalization and fusion techniques
using finger-vein and fingerprint biometrics and their potential application as biometric
identifiers. The individual scores obtained from finger-vein and fingerprints are combined
at score level using three score normalization techniques (Min-Max, Z-Score, Hyperbolic
Tangent) and four score fusion approaches (Minimum Score, Maximum Score Simple
Sum and User Weighting). The fused-score classifies an unknown user as genuine or
impostor.
The rest of the paper is organized as follows: Section 2 gives an overview of related
work, Section 3 describes the normalization and fusion techniques, Section 4 presents the
experiments conducted and the data analysis, Section 5 presents the experimental results,
and lastly, Section 6 rounds off the paper with a discussion and indication of future work.
2 Related Work
Different score fusion approaches have been proposed for fusing scores received from
different biometric modalities.
Raghavendra et al. [4] proposed a multimodal biometric score level fusion scheme
using Gaussian Mixture Model and Monte Carlo Method. The authors fused face, speech
and palmprint modalities and showed that their method gives higher performance than
other fusion schemes, such as Simple sum rule, Weighted sum rule, Fishers Linear
Discriminate Analysis (FLD), and Likelihood Ratio (LR) methods.
Derawi et al. [5] fused gait and fingerprint traits for user authentication on mobile
devices. The authors used four different methods to normalize the scores (min-max,
z-score, median absolute deviation, tangent hyperbolic) and four fusion approaches
(simple sum, user-weighting, maximum score and minimum score). The fusion results
of fingerprint and gait recognition showed an improved performance over other methods
and brought as a step closer to the advancement of user authentication on mobile devices.
Peng et al. [6] fused finger vein, fingerprint, finger shape and finger knuckle print of
a single human finger based on triangular norm (t-norm). The results showed that score
level fusion using triangular norm obtains a larger distance between genuine and imposter
score distribution, as well as lower error rates.
Conti et al. [7] and Lau et al. [8] proposed a multi-instance fusion approach. Their
approach is based on fusing two comparison scores using fuzzy logic rules from two
different fingerprints. Experimental results showed an improvement of 6.7% using the
comparison score level fusion rather than a unimodal authentication system.
Kisku et al. [9] proposed a multimodal biometric system using face and ear
biometrics. They used Gaussian Mixture Model (GMM) with belief fusion for
the estimated scores characterized by Gabor responses, and the proposed fusion is
accomplished by Dempster-Shafer (DS) decision theory. It has been proved that DS
provides increased accuracy and significant improvements over the existing classical
fusion rules.
3 Score Level Fusion
Most of the multimodal biometric systems integrate data at score level due to the strong
trade-off between the ease in combining the data and better information content. Besides,
it is a relatively straightforward way to combine scores generated by different comparators
(matchers)1 . Therefore, score level fusion is the preferred approach for integrating
biometric data. Figure 1 illustrates this process.
Fingerprint
Database
Fingerprint
(FP) image FP Score
FP Feature Fingerprint FP Score Score
Extractor Matcher Normalization threshold
FUSED
Score SCORE Match
Fusion Decision?
Non-match
Module
FV Feature Finger-vein FV Score
Extractor Matcher Normalization FV Score
Finger-vein
(FV) image
Finger-vein
Database
Figure 1: Score Level Fusion Diagram for Fingerprint and Finger-vein.
At score level, fused scores of each modality are combined with a variety of techniques
to produce a new score that is compared with the threshold in the decision module. There
are two key approaches in use today for consolidating comparison scores; classification
and combination. In classification we can construct a vector with individual scores
that are classified into accept or reject classes. A classification approach might use a
decision tree, Support Vector Machine (SVM), or Linear Discriminate Analysis (LDA)
to classify the vector as an impostor or genuine. In combination approach, individual
scores are combined to generate a single scalar score to render the final decision [11].
The combination approach for consolidating scores has compiled a superior performance
record versus the other levels.
Score level fusion has two main steps. The first step is score normalization, where
calculated scores are created by a certain comparator Si and are mapped onto a new
0
score scale or domain Si . For instance, if a comparator X produces scores on a domain
[1, 100] and a comparator Y generates scores on a domain [1, 2500] these scores need to
be normalized and mapped at a common scale. The second step of fusion at score level
is fusion. There are many score fusion techniques, however, in this paper, we have used
only some of the fusion techniques that are recommended by ISO standards - "ISO/IEC
TR 24722:2007 - Multimodal and other Multibiometric Fusion"[12] and are described in
Section 3.2.
1 Note
that the term "matching" as a synonym for "comparison" has been deprecated in the ISO SC37
Harmonized Biometric Vocabulary [10].
3.1 Score Normalization
Score normalization processes have been researched extensively [11]. In this section, we
describe some fundamental normalization concepts used in biometrics to facilitate easier
interpretation of our paper. The score normalization process is performed to transform
the comparator’s parameters and data types into a common domain. Commonly, score
normalization techniques are evaluated based on their robustness and efficiency.
The most used score normalization techniques that are also used and evaluated in this
paper are Min-Max (MM), Z-Score (ZS), and Hyperbolic Tangent (TanH). These methods
are discussed below.
Table 1: Symbols Used for Score Normalization Expressions.
Statistical Measures Genuine Distribution Impostor Distribution Both
Minimum Score G
SMin I
SMin B
SMin
Maximum Score G
SMax I
SMax B
SMax
Mean G
SMean I
SMean B
SMean
Score Standard Deviation G
SSD I
SSD B
SSD
Min-Max Normalization (MM) performs a linear transformation of the original data.
This is one of the simplest normalization techniques and is appropriate when the
limits of the produced scores are known. MM is considered efficient and provides
adequate performance; however, it may not yield completely accurate results if the
data contain outliers. MM maps raw scores to the [0,1] range and given scores SBMax
and SBMin designates the endpoints of the score range.
B
S − SMin
S0 = B − SB
(1)
SMax Min
Z-Score Normalization (ZS) is one of the most commonly used normalization tech-
niques. It uses an arithmetic mean and a standard deviation to normalize the data;
therefore, a priori knowledge regarding the average score and score variances of
the comparator is needed. ZS is considered efficient and tends to work exception-
ally well if the scores of each modality follow a Gaussian distribution; however,
this technique may not achieve similar accuracy if the data used contain outliers
(the mean and standard deviation are sensitive to outliers). ZS normalization trans-
forms the scores to a normal distribution with an arithmetic mean SIMean of 0 and a
standard deviation SISD of 1.
I
S − SMean
S0 = I
(2)
SSD
Hyperbolic Tangent Normalization (TanH) is efficient (when the parameters are se-
lected carefully), provides adequate performance, and is very robust in handling
outliers. TanH maps the raw scores to the (0,1) range, where SG B
Mean and SSD are the
mean and standard deviation estimation of the score distribution, respectively.
G
0.01(S − SMean )
S0 = 0.5 · tanh · B
+1 (3)
SSD
Figure 2, illustrates distribution graphs for the aforementioned fingerprint score
normalization techniques; where x represents raw scores before normalization and y
represents normalized scores.
Figure 2: Distributions of Genuine and Impostor Comparison Scores.
3.2 Score Fusion Techniques
In general, score fusion techniques fall under two categories: classification and
combination. Classification techniques formulate the problem by dividing the decision
space into two classes: genuine and impostor. The reliability and effectiveness of this
method are dependent on the large amount and quality of the input data that are used to
train the classifier. In addition, the comparison scores do not need to be homogeneous;
hence the normalization step is not required. Some of the classification methods that have
been researched are neural networks, nearest neighborhood algorithms, and tree-based
classifiers.
Combination is the most common and effective technique for combining biometric
scores. This technique combines scores from multiple comparators and generates a single
score. Combination requires score normalization before fusing the scores. This paper
analyses the following score level fusion techniques: Maximum Score, Minimum Score,
Simple Sum, and Weighted Sum.
Maximum Score: the max rule estimates the mean of the posteriori probabilities by the
maximum value.
max(i = 1 to N) Si0 (4)
Minimum Score: the min rule sets the minimum value of posteriori probabilities.
min(i = 1 to N) Si0 (5)
Simple Sum: it is a weighted average of the raw scores. Comparison scores are summed
without the benefit of normalization routines. It simplistically assumes that the
raw scores supplied by the biometric methods used have a comparable scale,
distribution, and strength. It can be used in ambiguous classifications resulting
from high-level noise.
∑(i = 1 to N) Si0 (6)
Weighted Sum: this method computes the combined score as a weighted sum of the
comparison scores. The motivation behind the idea of user-specific weights for
computing a weighted sum of scores is that some biometric traits cannot be reliably
obtained from some people (e.g., individuals with faint fingerprints). Assigning
a lower weight to a fingerprint score and a higher weight to a finger-vein score
reduces the probability of a false rejection.
∑(i = 1 to N)Wi∗ · Si0 (7)
When it is implemented correctly, score level fusion can improve accuracy, thwart
fraudsters, and increase usability. On the other hand, in case it is implemented incorrectly,
a multibiometric system might experience performance degradation in comparison to a
unimodal system. Furthermore, multimodal biometric systems have some disadvantages.
They could potentially have a higher cost of ownership, they could increase user
inconvenience, they could decrease user acceptance, and also exacerbate privacy issues.
4 Experiment Setup
In this section we present details regarding the database, datasets and modifications
needed to conduct our experiments.
4.1 Database
Fingerprint and finger-vein experiments in this paper are conducted over four different
datasets collected during the summer of 2010 by the Machine Learning and Applications
(MLA) Group at Shandong University, in China [13].
The MLA group named this database ”SDUMLA- HMT: A Multimodal Biometric
Database”. The database includes 106 subjects; 61 males and 45 females between 17
and 31 years old [13]. The database consists of face images captured from seven different
view angles, finger vein images of six fingers, gait videos from six view angles, finger-
vein images from a finger-vein sensor and fingerprint images captured by five different
sensors. Figure 3, shows some sample images from SDUMLA-HMT database.
In addition, modifications were needed over the acquired data to conduct our
experiment. In particular, we have reduced the number of participants, fingers, and
impressions as follows: we use two index fingers of both hands out of 6 fingers provided
in the dataset, five impressions for each fingerprint out of eight that were captured, and we
reduced the number of participants from 106 to 100. These modifications were demanded
because in the provided database there were available only finger-vein images of the index
fingers. Consequently, we have 2( f ingers)x100(sub jects)x5(attempts) = 1000 images
Figure 3: Sample Images of Face, Finger-vein, Gait, Iris and Fingerprint from SDUMLA-
HMT Database [13].
per dataset. In other words, we used 2000 fingerprint images from two datasets (DS2
and DS3) out of a total of 25,440 fingerprint images. Furthermore, we decided to follow
the naming convention recommended in ISO 19794-2 [14]. To do that we requested the
original finger-code positions from the MLA Group and we renamed them accordingly.
An illustration of finger-code positions (names) from the MLA Group and ISO 19794-
2 is given in Figure 4.
Figure 4: Illustration of Finger-code Positions (names).
4.2 Fusion Experiments
Based on a quality assessment conducted over the database, the best datasets are named
with suffix 1, while the worst datasets are named with suffix 2.
• Fingerprint Datasets:
– Fingerprint, best quality dataset (SDUMLA-HMT DS2) is named as FP_DS1,
– Fingerprint, worst quality dataset (SDUMLA-HMT DS3) is named as
FP_DS2.
• Finger-vein Datasets:
– Finger-vein, best quality dataset (SDUMLA-HMT Finger-vein-Lamp) is
named as FV_DS1,
– Finger-vein, worst quality dataset (SDUMLA-HMT Finger-vein) is named as
FV_DS2.
We have defined four fusion scenarios; fusion with the best and the worst dataset as
follows:
1. Fusion of FP_DS1 and FV_DS1
2. Fusion of FP_DS1 and FV_DS2
3. Fusion of FP_DS2 and FV_DS1
4. Fusion of FP_DS2 and FV_DS2
5 Experimental Results
We have retrieved both low and high Equal Error Rates (EERs) from unimodal biometric
systems. In table 2, we can see that TanH gives better performance than MinMax and Z-
Score. Additionally, Simple Sum gives better results than Maximum Score, Minimum
Score and User Weighted Sum. Specifically, Simple Sum yields 99.98 % average
verification accuracy, making it the most stable fusion framework.
Table 2: Sample Comparison Results for Different Normalization and Fusion Techniques.
EER (%)
Normalization Technique
Fusion Technique
MinMax(MM) Z-Score(ZS) TanH
Minimum Score (MinS) 3.81836% 7.32851% 1.01881%
Maximum Score (MaxS) 0.11591% 0.11206% 0.07837%
User Weighting (UW) 0.08949% 0.15935% 0.01763%
Simple Sum (SS) 0.08281% 0.10955% 0.00010%
5.1 Fusion Scenarios
In the following sections we summarize the results of four fusion scenarios (fusion with
the best and the worst dataset).
5.1.1 Scenario #1:Fusion of FP_DS1 and FV_DS1
This case shows the fusion performance of finger-vein dataset FV_DS1 and fingerprint
dataset FP_DS1 using Hyperbolic Tangent estimators (TanH) normalization and Simple
Sum (SS). The results show that the fingerprints dataset attains an EER of 0.86%, the
finger-veins dataset an EER of 0.71%, and their fusion a low EER of 0.00010%.
5.1.2 Scenario #2: Fusion of FP_DS1 and FV_DS2
This scenario shows the fusion performance of the fingerprint dataset FP_DS1 and finger-
vein dataset FV_DS2 using Hyperbolic Tangent (TanH) and Maximum Score (MaxS). The
results show that the fingerprints dataset attains an EER of 0.86%, the finger-veins dataset
an EER of 7.35%, and their fusion a low EER of 0.0320%.
5.1.3 Scenario #3: Fusion of FP_DS2 and FV_DS1
This case shows the fusion performance of the fingerprint FP_DS2 and the finger-vein
datasets FV_DS1 using MinMax normalization (MM) and Maximum Score (MaxS)
fusion rule. The results show that the fingerprints dataset attains an EER of 1.01%, the
finger-veins dataset an EER of 0.71%, and their fusion a low EER of 0.00015%.
5.1.4 Scenario #4: Fusion of FP_DS2 and FV_DS2
This case shows fusion performance of the fingerprint FP_DS2 and finger-vein datasets
FV_DS2 using Hyperbolic Tangent estimators (TanH) normalization and Maximum Score
fusion rule. The results show that the fingerprints dataset attains an EER of 1.01%, the
finger-veins dataset an EER of 7.35%, and their fusion a low EER of 0.0038%.
Figure 5, presents four graphs for different normalization and fusion techniques per fu-
sion scenario (we get an insight of the increased performance when TanH and SS is used,
in relation to other normalization and fusion techniques).
Figure 5: Performance of Fingerprint and Finger-vein Using TanH Score Normalization
and Simple Sum Score Fusion
To summarize, the results indicate that fusion of finger-vein and fingerprint improves
the performance of biometric systems. We can define improvement as the percentage
difference between the minimum EER of fingerprint or finger-vein (we choose the lowest
value between the two) minus the EER of their fusion and (divided by) the minimum EER
of fingerprint or finger-vein (we choose the lowest value between the two):
EERmin(print|vein) − EER f usion(print+vein)
Improvement = · 100 (8)
EERmin(print|vein)
Table 3 summarizes our results. The values in percentages indicate EER.
Fingerprint alone Finger-vein alone Fusion(print+vein) Improvement (method)
0.86 % 0.71 % 0.0001 % 99.98 % (TanH+SS)
0.86 % 7.35 % 3.81 % 48.10 % (MM+MinS)
1.01 % 0.71 % 0.1872 % 79.50 % (TanH+UW)
1.01 % 7.35 % 0.0038 % 89.62 % (ZS+SS)
Table 3: Multimodal Fusion Improvements of Fingerprint and Finger-vein Recognition.
6 Conclusion
Multibiometric systems are the new frontier, and this is evident from the number of
ongoing efforts in the research and commercial domains. These systems address some
of the deficiencies mentioned above of the unimodal biometric systems. Theoretically,
they present a huge potential, but research and commercial systems have yet to reflect this
promise. Furthermore, the lack of standards indicates that more work is required before
it reaches an acceptable level of maturity.
Experimental results show that in most cases fused performance (fingerprint and
finger-vein) was significantly improved compared to unimodal biometric. It is worth
noting that the best fusion performance is achieved by the combination of Hyperbolic
Tangent (TanH) score normalization technique and Simple Sum (SS) method for fusion,
yielding an EER of 0.00010%. In other words, a multimodal biometric system with
fingerprint and finger-vein would perform 99.98% better than a unimodal biometric
system of fingerprint or finger-vein. In addition, it has been observed that both
MinMax and Z-Score methods are sensitive to outliers. On the other hand, TanH score
normalization is both robust and efficient.
Nevertheless, our future work should investigate fusion approaches using fingerprint
and finger-vein biometrics at feature extraction level or template level, to secure biometric
templates and enhance privacy by hiding the meaning of extracted features (points) from
finger-vein and fingerprint in the stored template.
Acknowledgement
This research is supported by the Research Council of Norway under the Grant
No.: IKTPLUSS 248030/O70 and 247648 for SWAN and Oslo Analytics projects,
respectively. This research is also part of the SecurityLab of University of Oslo.
References
[1] L. I. Kuncheva, C. J. Whitaker, C. A. Shipp, and R. P. W. Duin. Is Independence
Good for Combining Classifiers? In Proceedings of the 15th International
Conference on Pattern Recognition, volume 2, pages 168 –171 vol.2, 2000.
[2] B. Bhanu and V. Govindaraju. Multibiometrics for Human Identification. Cambridge
University Press, 2011.
[3] Anil K. Jain and Arun Ross. Multibiometric Systems. Communications of the ACM,
47(1):34, 2004.
[4] R. Raghavendra, A. Rao, and G. Hemantha Kumar. A Novel Approach for
Multimodal Biometric Score Fusion Using Gaussian Mixture Model and Monte
Carlo Method. International Conference on Advances in Recent Technologies in
Communication and Computing. ARTCom ’09., pages 90–92, 2009.
[5] M. O. Derawi, D. Gafurov, R. Larsen, C. Busch, and P. Bours. Fusion of gait and
fingerprint for user authentication on mobile devices. In 2010 2nd International
Workshop on Security and Communication Networks (IWSCN), pages 1–6, May
2010.
[6] Jialiang Peng, Ahmed A. Abd El-Latif, Qiong Li, and Xiamu Niu. Multimodal
Biometric Authentication Based on Score Level Fusion of Finger Biometrics. Optik
- International Journal for Light and Electron Optics, 125(23):6891–6897, 2014.
[7] Vincenzo Conti, Giovanni Milici, Patrizia Ribino, Filippo Sorbello, and Salvatore
Vitabile. Fuzzy Fusion in Multimodal Biometric Systems, pages 108–115. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2007.
[8] CW Lau, Bin Ma, HML Meng, YS Moon, and Yeung Yam. Fuzzy Logic Decision
Fusion in a Multimodal Biometric System. Interspeech, 2004.
[9] Dakshina Ranjan Kisku, P Gupta, H Mehrotra, and J Sing. Multimodal Belief Fusion
for Face and Ear Biometrics. Intelligent Information Management, 01:166–171,
2009.
[10] ISO/IEC JTC 1/SC 37 Biometrics: SC 37 Standing Document 11 (SD 11), Part 1
Harmonization Document - Biometric Vocabulary.
[11] Kamer Vishi and Audun Jøsang. A new approach for multi-biometric fusion based
on subjective logic. In Proceedings of the 1st International Conference on Internet
of Things and Machine Learning, IML ’17, pages 68:1–68:9, New York, NY, USA,
2017. ACM.
[12] ISO/IEC TR 24722:2007, Information Technology - Biometrics - Multimodal and
Other Multibiometric Fusion, 2007.
[13] Yilong Yin, Lili Liu, and Xiwei Sun. SDUMLA-HMT: A Multimodal Biometric
Database. Biometric Recognition, pages 260–268, 2011.
[14] ISO/IEC 19794-2:2005, Information technology – Biometric Data Interchange
Formats – Part 2: Finger Minutiae Data, 2005.