1. Indexing and Binning Large Databases

2. Abstract Problems with large databases
Biometric identification (1:N Matching) does not scale well with size
No established way to organize high dimensional biometric data
Proposed Solution
Reduce search space before 1:N matching
Divide the database using Clustering Techniques
Contributions
We analyze the effect of implementing a binning scheme on search performance and accuracy
We present binning and pruning approaches using multiple biometrics
Using hand geometry and signature, we have achieved a search space reduction of 95% without any FRR

3. Background
Only biometric identification (1:N matching) can prevent duplicate enrollments, double dipping
Biometrics are being deployed for immigration and national ID applications
US-VISIT program
Voter ID and national ID programs[3]
Potential size that can run into millions
Current research is focused only on accuracy
Apart from accuracy, scalability, speed and efficiency also become important at this scale
Conventional security systems rely on PINs, passwords and other token or key based methods for authentication and identification of users. Though these systems are easy to use, they are insecure as the tokens can be lost, stolen or used by more than one person. With each service requiring different form and means of identification, the multiplicity of authentication schemes becomes difficult to manage.

4. Challenges Textual/Numeric Data Data is scalar(1D)
Textual/numeric data can be linearly ordered and therefore easily indexed
Biometric Data
Biometric templates are high dimensional
No linear ordering or sorting methods exists for biometric data

5. Search space analysis
As number of stored templates increases, template density (TD) also increases
Nc = Number of Clusters
For all the clusters – calculate the distance of all points from the mean of the cluster they belong to
The average of all such distances = Dic – Average Intra cluster distance: NOTE: the averaging is done by the NUMBER OF USERS
The 2 colored plots: The right plot shows extra elements added to the plot on the left  It shows that as the number of templates increases, the probability of finding 2 templates very close to each other also increases.
Template density simply tries to quantize the tightness of each cluster  closeness of samples within same cluster
Template density TD = (Nc / Dic)  As the number of templates increase, the optimal number of clusters needed to cluster them increases. After increasing the number of clusters, the resulting clusters are such that all templates within the same cluster are very close to each other  hence the average intra-cluster distance mentioned above decreases
The increase in the number of clusters is given by Nc.
The decrease in the average intra-cluster distance is given by Dic
Thus the template density defined by Nc/Dic increases as the numerator Nc has increased and denominator Dic has decreased
EXPLAIN FORMULA IN DEPTH

6. Identification problem
Number of false positives grows geometrically with the size of the database
Let FAR and FRR be the False Acceptance Rate (probability) and False Reject Rate (probability) for 1:1 matching
For a 1:N matching,
The total number of False Accepts is given by
FAR = False Acceptance probability in 1:1 case
(1-FAR)  Probability of NO False Acceptance in 1:1 case
(1-FAR)N  Probability of NO False Acceptance in N 1:1 cases
1 – (1-FAR)N  Probability of at least 1 False Acceptance in N 1:1 cases  Probability of FAR for Identification task
False Reject happens when template belonging to same person is not recognized as actually belonging to him
Since there exists only 1 corresponding template in database per user, the FRR would be generated when this one particular template is missed/not recognized as belonging to the test-template  For both Verification and Identification, the FRR is generated in the same case
Thus FRRN = FRR
 EXPLAIN FORMULA IN DEPTH

7. State of the Art Biometrics State of the art Research Problems
Fingerprint
0.15% FRR at 1% FAR
(FVC 2002)
Fingerprint Enhancement
Partial fingerprint matching
Face Recognition
10% FRR at 1% FAR
(FRVT 2002)
Improving accuracy
Face alignment variation
Handling lighting variations
Hand Geometry
4% FRR at 0% FAR
(Transport Security Administration Tests)
Developing reliable models
Identification problem
Signature Verification
1.5%(IBM Israel)
Developing offline verification systems
Handling skillful forgeries
Voice Verification
<1% FRR
(Current Research)
Handling channel normalization
User habituation
Text and language independence

8. State of the Art Biometrics State of the art Research Problems
Fingerprint
0.15% FRR at 1% FAR
(FVC 2002)
Fingerprint Enhancement
Partial fingerprint matching
Face Recognition
10% FRR at 1% FAR
(FRVT 2002)
Improving accuracy
Face alignment variation
Handling lighting variations
Hand Geometry
2.6% FRR at 0.02% FAR
(CUBS, SUNY-Buffalo)
Developing reliable models
Identification problem
Signature Verification
1.5%(IBM Israel)
Developing offline verification systems
Handling skillful forgeries
Voice Verification
<1% FRR
(Current Research)
Handling channel normalization
User habituation
Text and language independence

9. Identification problem (contd.)
Even if FAR = %, False accepts = 1 in 10 for N=100000(lower bound) in the identification case.
No single biometric is capable of meeting this security requirement individually
Ways to reduce identification errors:
Reduce FAR
FAR is limited by feature representation and the recognition algorithm
Cannot be indefinitely reduced
Reduce N
Classify or index the biometric database. (e.g Henry classification system for fingerprints)
Index the records based on meta-data
Can we do better?

10. Fingerprint Features
Fingerprints can be classified based on the ridge flow pattern
The corrugated surface of the fingerprint is made up of ridges and valleys cover that the entire palmer surface of the hand. The flow pattern of these ridges and valleys are unique to each individual. These patterns that are used for identification and authentication. The image below shows the image of a fingerprint along with the distinguishing features on the print.
The flow of the ridges and patterns has been classified into 5 broad classes. This classification is used to catalog the fingerprints and also in authenticating two prints. Henry systems follow an elaborate classification scheme of cataloging and filing forensic prints. Fig1 shows the different classes of ridge flows. This methods cannot be used to distinguish between two fingerprints.
Fig 2 shows the distinguishing features on the fingerprint. These features are discontinuities or anomalies in the normal flow of ridges on the surface of the finger. These features are termed as minutiae (small details). There are eighteen different types of minutiae. Fig1 shows the most commonly encountered ones and their names. Fig 2 shows a thumbprint captured on paper. These are typically the kind of images that forensic AFIS (Automatic Fingerprint Identification Systems) have to deal with. The quality of the print not only deteriorates during capture but also during storage and hence AFIS systems are more sophisticated than their biometric counterparts.
Fingerprints can be distinguished based on the ridge characteristics
65% of fingerprints belong to the Loop class

11. Henry Classification of Fingerprints
[Ratha et al,1996] used Henry Classification on database of 1800 templates, tested on 100 templates
Search Space: 25%; FRR: 10%
[Jain, Pankanti,2000] similar experiment on database of 700 templates achieved FRR: 7.4% (Focus on classification only)
State-of-art Fingerprint classification system [Capelli,Maio,Maltoni,Nanni,2003] has FRR 4.8% for 5 class problem and 3.7% for 4 class problem
Though natural class exists, still classification is non-trivial
Natural classes do not exist for biometrics like Hand Geometry
Need more sophistication for partitioning database

12. Analysis of search space reduction
We can improve performance by reducing the search space during identification
Let PSYS – Penetration rate [between 0.0 and 1.0]
Penetration rate is the average fraction of the database searched during identification
Effective size = N*PSYS
For a 1:N matching,
NOTE: The State of the art Fingerprint system Psys refers to the Psys obtained at 0% FRR, the previous slide Fingerprint system achieved Psys of 0.25, but at a high FRR
Add 1 slide before this one - stating the work people have done using Henry Classification for Fingerprints
The total number of False Accepts is given by
State of the art fingerprint systems has PSYS=0.5

13. Effect of binning on accuracy
For PSYS < 0.2, the false accepts are almost constant
Query response time improves by a factor of PSYS
Capabilities of a low FAR system
Will allow us to screen immigrants at airports
Will make biometric systems more user-friendly by eliminating the need to remember PINs and IDs

14. Binning Binning can be used to achieve a smaller PSYS
Partition the feature space
Each bin is represented by a cluster center CK
Records are compared with only NB cluster centers
Bin representatives are computed offline during training
Challenges
How to handle clustering of large databases?
How to handle additions and deletions?
Put previous diagram in here .. After the 1st Main Point

15. Tradeoff
Although binning reduces search space, it introduces another source of identification error : Bin Miss
If the bin in which the user record exists is not searched, then FRR is generated no matter how good the matcher is
If P(B) is the probability of getting the correct bin
Binning increases the probability of False Rejects
Not tolerable in security and screening applications
Solution:
Use K-means clustering to find K bins
Check Ns nearest bins for the record, such that P(B) = 1
(1 – P(B)) = Probability of Missing a Bin
P(B)*FRRPsysn = Probability of a False Reject even after getting into the correct bin
 EXPLAIN FORMULA IN DEPTH

16. Formal definition of Binning
In general a biometric template may be represented as a vector
Vectors are represented into N distinct clusters; each represented by a ‘code book vector’
The code book vectors divide the feature space into N distinct Voronoi regions
Every template is closest to the mean (codebook vector) of the region it belongs to
Each template is closest to the mean (codebook vector) of the region which it belongs to
The Regions are such that they are non-overlapping  Their Intersection is a NULL set
The Union of the Regions form the entire data space
Explain Lower Formulations

17. Search Space Partition: Voronoi Regions

18. Hand Geometry Template
Feature extraction stages
Image capture
Binarization
Contour Extraction
Noise Removal
35 Features are extracted
25 directly measured features
10 ratio and perimeter features

19. Signature Template Regression Constants b0,b1 Connected Components
11 Features Extracted
Regression Constants b0,b1
Compactness
Signature Length
Major Stroke Length
Major Stroke Angle
Connected Components
Hole Count
Hole Area
Stroke Count
Signing Time

20. Results 11 – Dimensional Signature data
Best Penetration: 35.57% for 6 bins
FRR = 0%
35 – Dimensional Hand Geometry data Best Penetration: 35.8% for 6 bins
FRR = 0%
Dataset 250 Training Set & 250 Testing Set

21. Multi-modal approach Resulting bins have very high template densities
A different biometric modality should be used to classify templates within a bin
Multimodal biometrics
Using multiple biometrics improves accuracy
It is difficult to forge multiple biometrics
Composite templates reduce template density
Statistical independence ensures that individual binning results are diverse
The search space (intersection of bins) is reduced due to low commonality between the individual binning results

22. Multi-Modal Approach

23. Multi-Modal Approach
Search Space: 5% original database size; FRR – 0%

24. Results of Combination
Best combined penetration rate of 5%
Dataset 250 Training Set & 250 Testing Set

25. Binning v/s Indexing
Applications can have frequent insertions of new templates
Binning works well when database is static
Insertions will require re-partitioning the entire database
Indexing can be used in both – static and dynamic database scenarios
Trees are commonly used for indexing
Extend the concept of indexing relational databases to indexing biometric databases
Much more challenging – no concept of primary key exists in biometric templates!

26. Pyramid Technique spatial hashing
Determine the Pyramid (i) within with which the template lies
Determine height (h) of template from the apex
The 1-D value = Pyramid Number (i) + Height (h)
Indexing done using B+ Trees

27. Various Indexing Techniques
Grid Files KD Tree
R Tree R+ Tree X Tree
Pyramid Technique
ALL STRUCTURES: INSERTION ORDER DEPENDENT
R TREE: Overlapping nodes
R+ Tree: Long tree – poor space utilization
Grid File: Number of cells increases super-linearly with size of data
KD Tree: Useful only if database is non-changing
Grid File – Number of cells increases super-linearly with Data, Not effective for range-querying
KD Tree – Insertion Order dependent, Not dynamic, Range Querying difficult
R Tree – Overlapping intermediate nodes
R* Tree – Overlapping intermediate nodes
R+ Tree – Longer Trees, poor space utilization
X Tree – Insertion order dependent

28. Comparative Study Method Grid File Y N R Tree R* Tree R+ Tree KD Tree
Scalable
Order Invariant
Dynamic
Range Query
No Overlap
Grid File Y N
R Tree
R* Tree
R+ Tree
KD Tree
X Tree
Pyramid Tech

29. Results of Indexing
35 – Dimensional Hand Geometry data
Best Penetration: 27%
FRR = 0%
Dataset 450 Training Set & 450 Testing Set
Parallel combination with signature will further reduce the search space

30. Multimodal Biometrics

31. 2D Biometric: Signature & Fingerprint Fusion
Impostor Score Pairs
True Match Score Pairs

32. Optimal Fusion Algorithm Signature Fused With Fingerprint
Unrealizable Performance Area
True Match Score Pairs
Optimal Fusion ROC
Fusion
Algorithm
Accuracy (1-FRR)
False Accept Rate (FAR)
Suboptimal Performance Area
Impostor Score Pairs
The ROC is the boundary between
what is possible and suboptimal performance.

33. Match Zone No-Match Zone
Optimal Fusion Algorithm Decision Regions 99.04% Specified FAR of 1 in a Million
2nd Biometric Score Axis
1st Biometric Score Axis
irregular decision region boundary due to finite sample size
the more data the smoother the boundaries
No-Match Zone
Match Zone

34. RSS Fusion RSS Fusion ROC
RSS Fusion Algorithm for Fingerprint & Signature Provides A Suboptimal Performance ROC
Optimal ROC
True Match Score Pairs
RSS Fusion ROC
RSS
Fusion
Accuracy (1-FRR)
False Accept Rate (FAR)
Impostor Score Pairs

35. Match Zone No-Match Zone
RSS Fusion Decision Regions 96.11% Specified FAR of 1 in a Million
2nd Biometric Score Axis
1st Biometric Score Axis
No-Match Zone
Match Zone

36. OR Fusion Algorithm for Fingerprint & Signature Provides A Suboptimal Performance ROC
Optimal ROC
True Match Score Pairs
OR Fusion ROC OR
Fusion
Accuracy (1-FRR)
False Accept Rate (FAR)
Impostor Score Pairs

37. Match Zone No-Match Zone
OR Fusion Decision Regions 96.85% Specified FAR of 1 in a Million
2nd Biometric Score Axis
1st Biometric Score Axis
No-Match Zone
Match Zone

38. AND Fusion AND Fusion ROC
AND Fusion Algorithm for Fingerprint & Signature Provides A Suboptimal Performance ROC
Optimal ROC
True Match Score Pairs
AND Fusion ROC
AND
Fusion
Accuracy (1-FRR)
False Accept Rate (FAR)
Impostor Score Pairs

39. Match Zone No-Match Zone
AND Fusion Decision Regions 62.91% Specified FAR of 1 in a Million
2nd Biometric Score Axis
1st Biometric Score Axis
No-Match Zone
Match Zone

40. ROC

41. Thank You