Documente Academic
Documente Profesional
Documente Cultură
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/287481134
CITATIONS READS
4 57
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by R. Raghavendra on 28 July 2016.
LFC. This procedure resulted in a print artefact database Figure 7. Illustration of artefacts generated from best focus images
with 1122 light field samples (each consisting of multiple obtained from the Lytro LFC (a) normal (or real) image (b) iPad
depth images) corresponding to 520 high resolution DSLR (c) Samsung tablet
samples. (2) Screen artefact using iPad: Here, we record
the artefacts by displaying each DSLR captured image us-
age selected from the set of multiple focus image rendered
ing iPad(4th generation) with retina display to the Lytro
by the LFC, we used the best focus image corresponding to
LFC as shown in Figure 5 (b). This artefact database con-
each subject to generate the corresponding artefact sample.
sists of 1444 light field samples (each consisting of multiple
For simplicity, here we generate the artefact sample by pre-
depth images) corresponding to 520 high resolution DSLR
senting the best focus sample using both iPad and Samsung
camera samples. (3) Screen artefact using Samsung Galaxy
tablet one at a time to the Lytro LFC. The artefact image
Note 10.1: In order to more effectively analyze the effect of
database collected using an iPad will result in a total of 1860
electronic screen attacks, we consider to collect the artefact
light field samples while the use of Samsung tablet resulted
samples using Samsung Galaxy note 10.1. Here, we capture
in a total of 1973 light field samples. Figure 7 shows the
the artefact samples by storing the high quality samples cap-
examples of the artefacts captured from LFC using iPad (7
tured using DSLR in the Samsung tablet that in turn will be
(b)) and Samsung tablet (7 (c)).
presented to the LFC. This database is comprised of 1208
light field samples including multiple depth images. Figure 3.3. Performance evaluation protocol
6 shows the examples of the artefact images captured us-
In this section, we describe the performance evaluation
ing iPad (fig:Figure6 (b)), Samsung tablet (6 (c)) and photo
protocol to effectively analyze the prominence of our gen-
print (6 (d)). While Figure 6 (a) shows the corresponding
erated visible spectrum iris artefacts and also to bench-
normal (or real) image.
mark the performance of our proposed PAD algorithm. The
Natural Image Artefact Images
whole database of 104 unique iris is divided into three in-
dependent non-overlapping groups namely: training, devel-
opment and testing. The training set consists of 4 unique
eyes, development set consist of 20 unique eyes and test-
ing set consists of 80 unique eyes. In this work, we em-
ployed the development set to tune the proposed PAD al-
gorithm and also to fix the value of the threshold which is
further used for the PAD classification. The value of the
(a) (b) (c) (d)
threshold adopted in this work corresponds to the Equal Er-
Figure 6. Illustration of artefacts generated from DSLR images (a) ror Rate (EER) calculated on the development dataset. The
Normal (or real) images (b) iPad (c) Samsung tablet (d) Photo print test dataset should be solely used to report the performance
of the proposed PAD algorithm. Further, in order to evalu-
In addition to the above mentioned artefact generation ate the vulnerability of the visible spectrum iris recognition
using high quality print-outs of the enrolled iris samples, we system for the artefacts collected in this work, we further
also consider to generate the artefacts using normal (or real) divide both development and testing dataset into reference
samples collected using LFC. This will allow one to study and probe samples. In this work, we have collected 5 sam-
the impact of the presentation attack in which an attacker ples for each subject, we choose the first four samples as
will successfully get access to the enrolled sample of the le- reference and the last sample as the probe. With this setting
gitimate user and later the same sample will be used to gen- of reference and probe, we report the baseline performance
erate the artefact to subvert the biometric sensor (i.e. LFC). with normal (or real) visible spectrum iris biometric sam-
Since our baseline system will use only the best focus im- ples. While, to study the vulnerability of the system for the
0.25
Genuine Scores
Baseline
Imposter Scores
Artefact attack Scores Baseline under attack
60
0.2 Baseline with proposed PAD
40
0.15
20
0.1
10
0.05 5
2
0
0 0.2 0.4 0.6 0.8 1 1
Score Bins
0.5
Figure 8. Overlapped comparison scores obtained on normal and
artefact (photo print) (best viewed in color) 0.2
0.1
0.050.1 0.2 0.5 1 2 5 10 20 40 60
False Match Rate (in %)
presentation attacks we employ the artefact samples (gener- Figure 9. Verification performance of the proposed PAD on photo
ated as mentioned in Section 3.2) as the probe samples. print artefact corresponding to DSLR samples (best viewed in
color)
3.4. Availability of database
The database shall be distributed for the academic situation illustrated in Figure 8 the baseline system has ac-
and research purposes via: http:/www.nislab.no/ cepted 96.78% of artefact samples as the genuine samples.
biometrics_lab/guc_lf_viar_db. This illustrates not only the quality of the artefact employed
in this work but also the strong need of an efficient and ro-
4. Experiments and results bust PAD algorithm.
Experimental results presented in this work is carried out Figure 9 shows the verification performance of the base-
according to the protocol presented in the Section 3.3. The line visible spectrum iris system under attack (photo print).
performance of the proposed PAD algorithm is measured It can be observed here that, despite the good performance
using two kind of errors [8] namely: (1) Attack Presenta- of the baseline system (with EER = 2.26% shown in green
tion Classification Error Rate (APCER), which is reporting color) on normal (or real) samples, when exposed to the at-
the proportion of attack presentation (with a fake or arte- tack the performance is heavily deceived (as shown in red
fact) that are incorrectly classified as normal (real) presen- color). This fact can be attributed to the strong overlapping
tation. (2) Normal Presentation Classification Error Rate of genuine and artefact comparison scores as shown in Fig-
(NPCER), which is reporting the proportion of normal pre- ure 8. Further, it can also be observed from Figure 9 that,
sentation incorrectly classified as attacks. Finally, the per- the implication of the proposed PAD algorithm that can mit-
formance of the overall PAD algorithm is presented in terms igate the presentation attacks will bring back the baseline
of Average Classification Error Rate (ACER) such that, performance to the normal level (indicated by the overlap-
ping of green and blue lines in the Figure 9). Similar results
AP CER + N P CER can also be observed with the iPad attack that uses lytro
ACER = (5)
2 samples as shown in Figure 10. We illustrated these two
cases for simplicity, however similar observation can also
The lower the values of the ACER, the better is the
be made with all five kinds of attacks that are addressed in
performance. The value of APCER, NPCER and ACER
this work.
are computed with the classifier threshold. The classifier
threshold is calculated on the development dataset such that Table 1. Performance of the proposed PAD algorithm
the threshold value corresponds to the point where APCER Presentation attacks PAD Performance Measure (%)
and NPCER are equal (nothing but an EER as mentioned in APCER NPCER ACER
the Section 3.3) on the development dataset. Photo Print 0.25 0.75 0.50
Artefacts from DSLR samples iPad 1.50 0.50 1.00
Figure 8 shows the distribution of the comparison scores
Samsung Tablet 0.75 0.50 0.62
obtained using the baseline visible spectrum iris system on
Artefacts from Lytro samples iPad 1.20 3.70 2.45
both normal (or real) and artefact (photo print attack) sam-
Samsung Tablet 1.40 0.90 1.15
ples. Here, one can observe the strong overlapping of the
artefact comparison scores with the genuine comparison
scores obtained on the normal (or real) samples. With the Table 1 presents the quantitative performance of the pro-
Baseline ples. The attacks are carried out by presenting these arte-
60
Baseline under attack facts using photo print and electronic screens (using iPad
Baseline with proposed PAD
and Samsung tablet) to the LFC that is used as the biomet-
40
ric sensor for the visible spectrum iris recognition system.
Extensive experiments carried out on our database has re-
False Non Match Rate (in %)
5 References
2 [1] Livdet-iris competition. "http://people.clarkson.
1
edu/projects/biosal/iris/index.php".
0.5
[2] Mobile iris liveliness detection competition (mobilive 2014).
"http://mobilive2014.inescporto.pt".
0.2
0.1
[3] A. Czajka, P. Strzelczyk, and A. Pacut. Making iris recogni-
0.050.1 0.2 0.5 1 2 5 10 20 40 60 tion more reliable and spoof resistant. In SPIE, 2007.
False Match Rate (in %)
[4] J. Galbally, S. Marcel, and J. Fierrez. Image quality assess-
Figure 10. Verification performance of the proposed PAD on iPad
ment for fake biometric detection: Application to iris, fin-
artefact corresponding to Lytro samples (best viewed in color)
gerprint, and face recognition. IEEE Transactions on Image
Processing, 23(2):710–724, Feb 2014.
posed PAD scheme on all five kinds of attacks discussed [5] X. He, Y. Lu, and P. Shi. A fake iris detection method based
in this work. Quantitative results shown in the Table 1 are on fft and quality assessment. In Chinese Conference on Pat-
tern Recognition, pages 1–4, Oct 2008.
obtained by performing 10-fold cross validation to partition
[6] X. He, Y. Lu, and P. Shi. A new fake iris detection method. In
the whole database into training, development and testing
M. Tistarelli and M. Nixon, editors, Advances in Biometrics,
set. Here, it can be observed that, the proposed PAD algo-
volume 5558, pages 1132–1139. Springer Berlin Heidelberg,
rithm has shown the outstanding performance especially on 2009.
the print attack with ACER of 0.5%. In addition, similar [7] K. Hughes and K. W. Bowyer. Detection of contact-lens-
performance can also be acknowledged on the remaining based iris biometric spoofs using stereo imaging. In 46th
four different kinds of attacks. Thus, based on the above Hawaii International Conference on System Sciences, pages
experiments the proposed PAD algorithm that explores the 1763–1772, Jan 2013.
variation of focus between multiple depth images rendered [8] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC WD 30107-
by the LFC provides a new dimension for the biometric ap- 3:2014 Information Technology - presentation attack detec-
plications. tion - Part 3: testing and reporting and classification of at-
In addition the proposed PAD algorithm can offer the tacks. International Organization for Standardization, 2014.
following advantages: (1) Since we are exploring an inher- [9] A. Jain, K. Nandakumar, and A. Ross. Score normaliza-
ent characteristics of the LFC, the proposed PAD scheme tion in multimodal biometric systems. Pattern recognition,
will work as an integral component rather than a stand-alone 38(12):2270–2285, 2005.
unit that are normally used in the available state-of-the-art [10] J.Daugman. iris recognition and anti-spoofing countermea-
schemes. (2) Overcomes the need of additional feature ex- sures. In 7th International Biometrics Conference, 2004.
tractions scheme based on LBP, quality defining parameters [11] J. Kautsky, J. Flusser, B. Zitová, and S. Šimberová. A new
wavelet-based measure of image focus. Pattern Recognition
and also complex classifiers. Further there is no need for
Letters, 23(14):1785–1794, 2002.
additional hardware components and the proposed scheme
[12] Kiran B. Raja, R. Raghavendra, F. A. Cheikh, B. Yang, and
is highly computationally efficient.
C. Busch. Robust iris recognition using light field camera. In
The Colour and Visual Computing Symposium 2013, 2013.
5. Conclusion [13] O. Komogortsev and A. Karpov. Liveness detection via ocu-
In this work, we have presented a novel PAD algorithm lomotor plant characteristics: Attack of mechanical replicas.
that exploits the variation of the depth between multiple In International Conference on Biometrics (ICB), pages 1–8,
depth images rendered by the LFC. This work explores the June 2013.
inherent characteristics of the LFC that can reveal the pres- [14] S. J. Lee, K. R. Park, and J. Kim. Robust fake iris detection
ence of the presentation attacks. To this extent, we collected based on variation of the reflectance ratio between the iris
a new artefact visible spectrum iris database by simulating and the sclera. In Biometric Consortium Conference, pages
five different kind of attacks. The artefacts are generated us- 1–6, Sept 2006.
ing both high quality visible spectrum iris samples obtained [15] A. Pacut and A. Czajka. Aliveness detection for iris biomet-
using DSLR camera in addition to the light field iris sam- rics. In Carnahan Conferences Security Technology, Pro-
ceedings 2006 40th Annual IEEE International, pages 122–
129, Oct 2006.
[16] R. Raghavendra, Kiran B. Raja, B. Yang, and C. Busch.
Combining iris and periocular recognition using light field
camera. In 2nd IAPR Asian Conference on Pattern Recogni-
tion (ACPR2013), 2013.
[17] R. Raghavendra, B. Yang, Kiran B. Raja, and C. Busch. A
new perspective - face recognition with light-field camera. In
Biometrics (ICB), 2013 International Conference on, pages
1–8, 2013.
[18] A. F. Sequeira, J. Murari, and J. S. Cardoso. Iris liveness
detection methods in mobile applications. In 9th Interna-
tional Conference on Computer Vision Theory and Applica-
tions, pages 1–5, 2013.
[19] G. Sutra, B. Dorizzi, S. Garcia-Salicetti, and N. Othman. A
biometric reference system for iris, osiris version 4.1. 2012.
[20] C. Zhang, G. Hou, Z. Sun, T. Tan, and Z. Zhou. Light field
photography for iris image acquisition. In 8th Chinese Con-
ference on Biometric Recognition, pages 345–352. 2013.