Biohashing and Fusion of Palmprint and Palm Vein Biometric Data Modris Greitans, Arturs Kadikis, Rihards Fuksis Institute of Electronics and Computer Science Dzerbenes 14,
Download ReportTranscript Biohashing and Fusion of Palmprint and Palm Vein Biometric Data Modris Greitans, Arturs Kadikis, Rihards Fuksis Institute of Electronics and Computer Science Dzerbenes 14,
Slide 1
Biohashing and Fusion of Palmprint
and Palm Vein Biometric Data
Modris Greitans, Arturs Kadikis, Rihards Fuksis
Institute of Electronics and Computer Science
Dzerbenes 14, Riga, Latvia
e-mail: [email protected]
International Conference on Hand-based Biometrics
November 17-18, Hong Kong
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 2
Motivation
Multimodal Palm
Biometrics
Provides:
Easy enrolment
Unique parameters
Hard to falsify
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 3
Image Acquisition (I)
White LEDs
In visible light spectrum using white LEDs
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 4
Image Acquisition (II)
IR LEDs
In infrared light spectrum using IR LEDs
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 5
Image processing (I)
Cross section of the ridge
Cross section of the vessels
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 6
Image processing (II)
Complex 2D Matched Filtering:
• Based on the matched filtering
• Improved processing speed
• Obtains vectors: magnitude – matching rate; angle - orientation in the image
CMF ( x , y )
Cross section of the ridge
l
e
jK l
G ( x , y , l )
Cross section of the vessels
For further information:
M.Greitans, M.Pudzs, R.Fuksis. „Object Analysis in Images Using Complex 2d Matched Filters”, Proceedings of
the IEEE Region 8 Conference EUROCON 2009. Saint–Petersburg, Russia, May, 2009., pp. 1392-1397.
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 7
Image processing (III)
Feature
extraction
Filtering result
Vector set
Most significant vectors are extracted to describe the object.
The result is a data set of 64 vectors (256 bytes)
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 8
Raw biometric data comparison
Vector set A
Vector set B
Vector set from the database
Acquired vector set
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 9
Vector comparison
v p ( A) vq ( B )
Rihards Fuksis
Magnitudes:
v p ( A) vq ( B )
International Conference on Hand-based Biometrics
Slide 10
Vector comparison
v p ( A) vq ( B )
Magnitudes:
Angles:
Rihards Fuksis
v p ( A) vq ( B )
cos ( v p ( A ); v q ( B ))
International Conference on Hand-based Biometrics
Slide 11
Vector comparison
v p ( A) vq ( B )
Magnitudes:
Angles:
Distance:
Rihards Fuksis
v p ( A) vq ( B )
cos ( v p ( A ); v q ( B ))
d ||2
exp 2 exp
||
d 2
2
International Conference on Hand-based Biometrics
Slide 12
Vector comparison
v p ( A) vq ( B )
s p ,q
Magnitudes:
v p ( A) vq ( B )
Angles:
cos ( v p ( A ); v q ( B ))
Distance:
d ||2
exp 2
||
d 2
exp 2
d ||2
v p ( A ) v q ( B ) cos ( v p ( A ); v q ( B )) exp 2
||
d 2
exp 2
Dot product
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 13
Vector set comparison
Similarity index of two vectors:
s p ,q
d ||2
v p ( A ) v q ( B ) cos ( v p ( A ); v q ( B )) exp 2
||
d 2
exp 2
Similarity of two vector sets:
s ( A, B )
s
p
p ,q
q
Similarity index is normalized so that S(A,B) is in the [0;1]
s ( A, B )
Rihards Fuksis
s ( A, B )
s ( A, A) s ( B , B )
International Conference on Hand-based Biometrics
Slide 14
Security of raw biometric data usage
• It is unsecure to use raw biometric data
• Therefore encryption must be introduced
10 11 00 01 01 11 10 10 00 10
Encrypted data
Raw biometric data
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 15
Biohash
du
Pixels
Vectors
dv
Vector Set
CMF
(u,v)
Palm image
Inner
product
Inner
product
Random
number
matrix
...
u1
v1
R-th vector
uR
vR
Inner
product
du1 dv1
...
Token
1st vector
...
duR dvR
Data vector consists
of 4R components
...
Thresholding
1
0
Rihards Fuksis
...
1
Biocode
Biocode consists of
4R bits
International Conference on Hand-based Biometrics
Slide 16
Biohash Advancements(I)
Filtered palm vein image
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 17
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
83
163
87
44
52
85
146
83
41
57
87
51
Extracted vector magnitudes
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 18
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Rihards Fuksis
Most intensive vector labeling
International Conference on Hand-based Biometrics
Slide 19
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Data vector
u1
v1
du1 dv1 ... uR
Most intensive vector labeling
Most intensive vector information
vR
Rihards Fuksis
duR dvR + 0
0
1
0
0
International Conference on Hand-based Biometrics
1 ... 0
Slide 20
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Data vector
u1
v1
du1 dv1 ... uR
Most intensive vector labeling
Most intensive vector information
vR
duR dvR + 0
0
1
0
0
1 ... 0
New Data vector
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 21
Biohash Advancements(II)
Person 1; Biocode No.1
1 1 0 ...
Person 1; Biocode No.2
1 0 1 ...
Person 1; Biocode No.3
1 1 0 ...
Person 1; Biocode No.4
1 1 1 ...
By looking at the values before
the thresholding in Biohash
algorithm, we can obtain the
information about the distance
from threshold value for each of
the bits in biocodes
4 3 2
Random
number
matrix
Data
vector
Dot
product
Capture this value
Calculate the distance
to the threshold
Thresholding
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 22
Biohash Advancements(II)
Bit #1
Bit #2
Person 1; Biocode No.1
1
0.61
1 0.33 0 0.04
...
Person 1; Biocode No.2
1
0.47
0 0.12 1 0.15
...
Person 1; Biocode No.3
1
0.59
1 0.47 0 0.18
...
Person 1; Biocode No.4
1
0.46
1 0.39 1 0.14
...
4
2.13
3 1.19 2 0.29
...
Distance to
the threshold
Bit #1
Bit #2
Bit #3
Rihards Fuksis
...
Bit #3
If the distance to the
threshold value is greater,
the resulting bit most likely
will not change between
one person’s biocodes
International Conference on Hand-based Biometrics
Slide 23
Biohash Advancements(II)
Distance to
the threshold
Bits
Sort bits into groups
Distance to
the threshold
4
Rihards Fuksis
3
2
International Conference on Hand-based Biometrics
Slide 24
Biohash Advancements(II)
Distance to
the threshold
4
3
2
Sort bits in every group
in ascending order
Distance to
the threshold
4
Rihards Fuksis
3
2
International Conference on Hand-based Biometrics
Slide 25
Biohash Advancements(II)
Distance to
the threshold
4
3
2
What we obtain is the indexes of the most “stable” bits in descending
order. When comparing two biocodes this information is used to
calculate weights for the errors of the bits by using exp or other function
Weight function
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 26
Biocode comparison
= 4 mistakes
Similarity:
Sb
Rihards Fuksis
l Dh
l
l – length of the biocode
Dh – Hamming distance
International Conference on Hand-based Biometrics
Slide 27
Database evaluation
• Two databases; 500 images from 50 persons
• 5 images in IR and 5 in visible light spectrum
Raw biometric data comparison results[EER]
EER [%]
Palm Veins
Palm Prints
Fused data
0.32
2.79
0.1
Biohash test results [EER]
Palm Veins
Palmprints
Fused data
Mean
14.043
12.073
6.190
StDev
1.152
1.102
0.803
Proposed Biohash test results [EER]
Palm Veins
Palmprints
Fused data
Mean
1.073
0.471
0
StDev
0.304
0.231
0
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 28
Conclusions
• Complex 2D Matched Filtering approach
speeds up the feature extraction procedure.
• Biohashing with proposed advancements can
be used as a method for securing the
biometric data with similar or better precision
as raw biometric data comparison gives
• Future work: Tests on larger databases and
evaluation of other biometric encryption
methods
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 29
This presentation was supported by ERAF funding under the
agreement No.2010/0309/2DP/2.1.1.2.0/10/APIA/VIA/012
Rihards Fuksis
International Conference on Hand-based Biometrics
Biohashing and Fusion of Palmprint
and Palm Vein Biometric Data
Modris Greitans, Arturs Kadikis, Rihards Fuksis
Institute of Electronics and Computer Science
Dzerbenes 14, Riga, Latvia
e-mail: [email protected]
International Conference on Hand-based Biometrics
November 17-18, Hong Kong
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 2
Motivation
Multimodal Palm
Biometrics
Provides:
Easy enrolment
Unique parameters
Hard to falsify
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 3
Image Acquisition (I)
White LEDs
In visible light spectrum using white LEDs
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 4
Image Acquisition (II)
IR LEDs
In infrared light spectrum using IR LEDs
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 5
Image processing (I)
Cross section of the ridge
Cross section of the vessels
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 6
Image processing (II)
Complex 2D Matched Filtering:
• Based on the matched filtering
• Improved processing speed
• Obtains vectors: magnitude – matching rate; angle - orientation in the image
CMF ( x , y )
Cross section of the ridge
l
e
jK l
G ( x , y , l )
Cross section of the vessels
For further information:
M.Greitans, M.Pudzs, R.Fuksis. „Object Analysis in Images Using Complex 2d Matched Filters”, Proceedings of
the IEEE Region 8 Conference EUROCON 2009. Saint–Petersburg, Russia, May, 2009., pp. 1392-1397.
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 7
Image processing (III)
Feature
extraction
Filtering result
Vector set
Most significant vectors are extracted to describe the object.
The result is a data set of 64 vectors (256 bytes)
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 8
Raw biometric data comparison
Vector set A
Vector set B
Vector set from the database
Acquired vector set
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 9
Vector comparison
v p ( A) vq ( B )
Rihards Fuksis
Magnitudes:
v p ( A) vq ( B )
International Conference on Hand-based Biometrics
Slide 10
Vector comparison
v p ( A) vq ( B )
Magnitudes:
Angles:
Rihards Fuksis
v p ( A) vq ( B )
cos ( v p ( A ); v q ( B ))
International Conference on Hand-based Biometrics
Slide 11
Vector comparison
v p ( A) vq ( B )
Magnitudes:
Angles:
Distance:
Rihards Fuksis
v p ( A) vq ( B )
cos ( v p ( A ); v q ( B ))
d ||2
exp 2 exp
||
d 2
2
International Conference on Hand-based Biometrics
Slide 12
Vector comparison
v p ( A) vq ( B )
s p ,q
Magnitudes:
v p ( A) vq ( B )
Angles:
cos ( v p ( A ); v q ( B ))
Distance:
d ||2
exp 2
||
d 2
exp 2
d ||2
v p ( A ) v q ( B ) cos ( v p ( A ); v q ( B )) exp 2
||
d 2
exp 2
Dot product
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 13
Vector set comparison
Similarity index of two vectors:
s p ,q
d ||2
v p ( A ) v q ( B ) cos ( v p ( A ); v q ( B )) exp 2
||
d 2
exp 2
Similarity of two vector sets:
s ( A, B )
s
p
p ,q
q
Similarity index is normalized so that S(A,B) is in the [0;1]
s ( A, B )
Rihards Fuksis
s ( A, B )
s ( A, A) s ( B , B )
International Conference on Hand-based Biometrics
Slide 14
Security of raw biometric data usage
• It is unsecure to use raw biometric data
• Therefore encryption must be introduced
10 11 00 01 01 11 10 10 00 10
Encrypted data
Raw biometric data
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 15
Biohash
du
Pixels
Vectors
dv
Vector Set
CMF
(u,v)
Palm image
Inner
product
Inner
product
Random
number
matrix
...
u1
v1
R-th vector
uR
vR
Inner
product
du1 dv1
...
Token
1st vector
...
duR dvR
Data vector consists
of 4R components
...
Thresholding
1
0
Rihards Fuksis
...
1
Biocode
Biocode consists of
4R bits
International Conference on Hand-based Biometrics
Slide 16
Biohash Advancements(I)
Filtered palm vein image
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 17
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
83
163
87
44
52
85
146
83
41
57
87
51
Extracted vector magnitudes
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 18
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Rihards Fuksis
Most intensive vector labeling
International Conference on Hand-based Biometrics
Slide 19
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Data vector
u1
v1
du1 dv1 ... uR
Most intensive vector labeling
Most intensive vector information
vR
Rihards Fuksis
duR dvR + 0
0
1
0
0
International Conference on Hand-based Biometrics
1 ... 0
Slide 20
Biohash Advancements(I)
Filtered palm vein image
76
27
187
49
0
0
1
0
83
163
87
44
0
1
0
0
52
85
146
83
0
0
1
0
41
57
87
51
0
0
0
0
Extracted vector magnitudes
Data vector
u1
v1
du1 dv1 ... uR
Most intensive vector labeling
Most intensive vector information
vR
duR dvR + 0
0
1
0
0
1 ... 0
New Data vector
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 21
Biohash Advancements(II)
Person 1; Biocode No.1
1 1 0 ...
Person 1; Biocode No.2
1 0 1 ...
Person 1; Biocode No.3
1 1 0 ...
Person 1; Biocode No.4
1 1 1 ...
By looking at the values before
the thresholding in Biohash
algorithm, we can obtain the
information about the distance
from threshold value for each of
the bits in biocodes
4 3 2
Random
number
matrix
Data
vector
Dot
product
Capture this value
Calculate the distance
to the threshold
Thresholding
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 22
Biohash Advancements(II)
Bit #1
Bit #2
Person 1; Biocode No.1
1
0.61
1 0.33 0 0.04
...
Person 1; Biocode No.2
1
0.47
0 0.12 1 0.15
...
Person 1; Biocode No.3
1
0.59
1 0.47 0 0.18
...
Person 1; Biocode No.4
1
0.46
1 0.39 1 0.14
...
4
2.13
3 1.19 2 0.29
...
Distance to
the threshold
Bit #1
Bit #2
Bit #3
Rihards Fuksis
...
Bit #3
If the distance to the
threshold value is greater,
the resulting bit most likely
will not change between
one person’s biocodes
International Conference on Hand-based Biometrics
Slide 23
Biohash Advancements(II)
Distance to
the threshold
Bits
Sort bits into groups
Distance to
the threshold
4
Rihards Fuksis
3
2
International Conference on Hand-based Biometrics
Slide 24
Biohash Advancements(II)
Distance to
the threshold
4
3
2
Sort bits in every group
in ascending order
Distance to
the threshold
4
Rihards Fuksis
3
2
International Conference on Hand-based Biometrics
Slide 25
Biohash Advancements(II)
Distance to
the threshold
4
3
2
What we obtain is the indexes of the most “stable” bits in descending
order. When comparing two biocodes this information is used to
calculate weights for the errors of the bits by using exp or other function
Weight function
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 26
Biocode comparison
= 4 mistakes
Similarity:
Sb
Rihards Fuksis
l Dh
l
l – length of the biocode
Dh – Hamming distance
International Conference on Hand-based Biometrics
Slide 27
Database evaluation
• Two databases; 500 images from 50 persons
• 5 images in IR and 5 in visible light spectrum
Raw biometric data comparison results[EER]
EER [%]
Palm Veins
Palm Prints
Fused data
0.32
2.79
0.1
Biohash test results [EER]
Palm Veins
Palmprints
Fused data
Mean
14.043
12.073
6.190
StDev
1.152
1.102
0.803
Proposed Biohash test results [EER]
Palm Veins
Palmprints
Fused data
Mean
1.073
0.471
0
StDev
0.304
0.231
0
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 28
Conclusions
• Complex 2D Matched Filtering approach
speeds up the feature extraction procedure.
• Biohashing with proposed advancements can
be used as a method for securing the
biometric data with similar or better precision
as raw biometric data comparison gives
• Future work: Tests on larger databases and
evaluation of other biometric encryption
methods
Rihards Fuksis
International Conference on Hand-based Biometrics
Slide 29
This presentation was supported by ERAF funding under the
agreement No.2010/0309/2DP/2.1.1.2.0/10/APIA/VIA/012
Rihards Fuksis
International Conference on Hand-based Biometrics