LivDet competitions

Pattern Recognition
and Applications Lab

LivDet 2013
Fingerprint Liveness Detection Competition
2013
Gian Luca Marcialis, Stephanie Schuckers, Fabio Roli
Organizing Committee
{marcialis,roli}@diee.unica.it, sschucke@clarkson.edu

Universit
di Cagliari, Italia

ICB-2013 - Madrid, Spain

Dipartimento di
Ingegneria Elettrica
ed Elettronica

http://pralab.diee.unica.it

Competition Overview

Fake creation (cooperative method)

Algorithms
Software-based systems
Four data sets (sensors)
2,000 live and fake fingers for training
2,000 live and fake fingers for test
Four basic materials (gelatine, silicone, plasticine, latex)

Three optical sensor (Biometrika, Italdata, Crossmatch) and one swipe sensor
Biometrika and Italdata
Test on non-consensual (latent) fake fingerprints

Crossmatch and Swipe

Test on consensual fake fingeprints

Due to problems with CMT data set we only present Biometrika, Italdata and
Swipe results

Systems
Hybrid-based systems (even hardware-based fingerprint liveness detection)
1,000 live and spoof (fake) attempts
additional attempts by 50 fake fingerprints per material brushed on saline
solution

http://pralab.diee.unica.it

http://pralab.diee.unica.it

Fake creation (cooperative method)

http://pralab.diee.unica.it

Fake creation (non-cooperative method)

Fake creation (non-cooperative method)

http://pralab.diee.unica.it

Part 1: Algorithms - Protocol

A Win32 console application:
LIVENESS XYZ.exe [ndataset] [inputfile] [outputfile]
[ndataset]: id number of the data set to analyze.
[inputfile]: text file with the list of images to analyze.
[outputfile]: text file with the output of each processed
image, in the same order of [inputfile].

http://pralab.diee.unica.it

http://pralab.diee.unica.it

Part 1: Algorithms - Protocol

Part 1: Algorithms - Results

Overall Accuracy rate

Four data sets created with four different sensors

4000 or more images for each of the data set
Nine participants (academic and industrial
institutions), three of them preferred to be anonymous
Eleven different algorithms

http://pralab.diee.unica.it

Biometrika Italdata

Swipe

Average

Dermalog

98.30%

99.20%

96.47%

97.99%

Anonym1

98.00%

98.85%

N.A.

N.A.

ATVS

94.95%

50.00%

53.55%

66.17%

Anonym2

98.20%

99.40%

94.19%

97.26%

UniNap1

95.30%

96.50%

85.93%

92.58%

UniNap2

93.45%

90.55%

73.15%

85.72%

UniNap3

93.45%

96.50%

85.93%

91.96%

Anonym3

94.30%

97.20%

94.75%

95.42%

HZ-JLW

67.05%

86.85%

84.81%

79.57%

Itautec

91.00%

96.10%

N.A.

N.A.

CAoS

95.40%

54.10%

N.A.

N.A.

http://pralab.diee.unica.it

Part 1: Algorithms Results

Part 2: Systems - Protocol

100,00
90,00

System requirements:
Running on either a Windows XP 32-bit or 64-Bit
system.
An .exe or similar executable file.
The use of either a USB or Firewire connection.
To output the collected image if the image is
considered a live image and a liveness score in the
range of 0 and 100 (100 is the maximum degree of
liveness, 0 spoof)

80,00
70,00
60,00
50,00
40,00
30,00
20,00
10,00
0,00
Consensual (LivDet 2011)
Average accuracy
http://pralab.diee.unica.it

Non-consensual (LivDet 2013)

Average standard deviation
http://pralab.diee.unica.it

Part 2: Systems - Protocol

Part 2: Systems - Results

FerrLive and FerrFake for submitted systems

Participants were asked to ship a fingerprint system

which captures a fingerprint image as well as outputs a
liveness detection score.

Submitted
Systems
Dermalog
Morpho

FerrFake
0.6%
0.0%

FerrLive
11.8%
1.4%

FerrFake
Known
0.3%
0.0%

Three spoof recipes and methods were made available

to the participants (the materials were Playdoh, Gelatin
and Ecoflex).
In addition the systems were tested with two
unspecified methods (the materials were Modasil and
Latex).
http://pralab.diee.unica.it

Conclusions

http://pralab.diee.unica.it

Many thanks to

The number of participants, from both academic and

industrial institutions, is growing with respect to
previous editions.
Part 1 showed that both consensual and nonconsensual fake fingerprint were difficult to recognize,
leading to a noticeable error rate.
In Part 2, a high FerrLive value, and thus a high rate of
live fingerprints misclassified as fakes, could still be a
problem when integrating fingerprint liveness
detection methods on verification systems

http://pralab.diee.unica.it

Stephanie Bauer
Luca Ghiani
Valerio Mura
Simona Tocco
David Yambay

http://pralab.diee.unica.it

FerrFake
Unknown
1.0%
0.0%