Freni - Ph.D. Defense Slides

Biometric SystemTemplate Editing

Template Replacement

Template Editing & Replacement: novel methodsfor Biometric Template Selection & Update

Biagio Freni

Advisor: Prof. Fabio Roli

Pattern Recognition and Application GroupDept. Electrical Electronic Engineering - University of Cagliari

05 March 2010

Biagio Freni Template Editing & Replacement in Biometric



Biometric SystemOverviewTemplate RepresentativenessState-of-the-Art: Template Selection & Update

Template EditingClustering AlgorithmsEditing AlgorithmsExperimental Comparison

Template ReplacementSemi-Supervised Template UpdateTemplate Update with ReplacementResults




Overture

20 January 2010 . . . A man has been founded dead in a Dubai’shotel.. . . couple of days later . . . Local Police discovered that 11 mainsuspects got into the country illegally using forged passports ofEuropean citizen. Police found out that pictures in the documentswere different from legitimate owner’s pictures.

. . . 14 January 2010 just a week before the Dubai affair, EUdelegates approved — 594 vs 51, while 37 abstained — the launchof Biometric Passport including owner’s fingerprint and face.




OverviewTemplate RepresentativenessState-of-the-Art: Template Selection & Update

What’s Biometric?

Biometric refers to the use of physiological or behaviouralcharacteristics, “unique” for each person, with the aim ofestablished people’s identity.

Core of Biometric System is represented by Templates.





Template Selection & Update

The issue of template selection and update, in biometricrecognition systems, is twofold and is related to:

I Selection during Enrollment regarding the effective creationof representative template gallery of client populations,keeping the number of templates as small as possible at thesame time.

I Update during Authentication regarding the need of adaptover time templates, in order to capture the variations, in thebiometric traits not presented in the time of enrollment.

Selection & Update are different problems that share the commonnotion of “best representative” templates.





State-of-the-Art: summary

State-of-the-Art can be summurized by following modalities 1:

I Supervised: requires human intervention to labeling data.

I Semi-Supervised 2: queries labelled by the system are usedfor the task.

I Offline: a bunch of semi-labelled data are stored during thesystem authentication, later, they are used to update system’stemplates when the system itself is not operative.

I Online: each coming query is evaluated by the system duringauthentication phase, template adaptation is performed online.

1A. Rattani, B. Freni, G.L. Marcialis, F. Roli, Template Update Methods inAdaptive Biometric Systems: A Critical Review, ICB09, pp 847-856.

2B. Freni, G.L. Marcialis, and F. Roli, Online and offline fingerprinttemplate update using minutiae: an experimental comparison, AMDO08, July,9-11, 2008, Eds., Springer LNCS 5098, pp. 441-448.





PhD work

This PhD work explored the whole S-o-A and new methods havebeen proposed and published :

I S-o-A: Template Update Methods in Adaptive BiometricSystems: A Critical Review, al. et Freni, ICB09.

I Supervised: Template Selection by Editing Algorithms: acase of Study in Face Recognition, Freni et al., S+SSPR08.

I Semi-Supervised

I Offline: Online and offline fingerprint template update usingminutiae: an experimental comparison, Freni et al., AMDO08.

I Online: Replacement algorithms for fingerprint templateupdate, Freni et al., ICIAR08.

For sake of time just two works are addressed in this talk Editingmethods for Template Selection and Replacement algorithms forTemplate Update.




Clustering AlgorithmsEditing AlgorithmsExperimental Comparison

Template selection in Biometric

I Problem statement Given a set of N templates for a givenperson, select K templates that “best” represent the owner’sidentity.

I State-of-the-Art Derived from the clustering theory,consisting in exploring each template gallery according withtwo criteria: maximum similarity among templates (MDIST),maximum variation among them (DEND).

I Main Cons1. The procedure is not fully automatic since it requires the

manual insertion of parameter K .2. All the template gallery are resized to the same dimension K ,

without taking into account “intrinsic” difficulty of each client.





SoA MDIST: maximum similarity among templates

apply to all client’s gallery

1. Compute distance betweenN templates

2. For each templatecompute the averagedistance with the other(N − 1)

3. Choose K templates withsmallest average distanceas new selected gallery





SoA DEND: maximum variation among templates

apply to all client’s gallery

1. Generate a NxN dissimilaritymatrix DM

2. Apply Complete Link Clusteringto DM in order to generate aDendrogram D, using D toidentify K clusters

3. For each K cluster select thecenter

4. The set of templates selected in 3.represent a new selected gallery





Novel Proposal: Template Editing for Biometric

Editing AlgorithmsEditing algorithms belong to the K −NN classifier theory. K −NNuse a set of prototype to perfom classification. A pattern isclassified according to the majority of “K ” prototypes close to it.

Biometric could be seen as a “1− NN” classifier where templatesare prototypes.

Editing consist in generating from a given Template Set T a subsetE able to maintain the same classification accuracy on T itself.

Characteristics of Editing Algorithms:

1. the procedure is completly automatic

2. build up variable length galleries accordingly with the“difficult” of each client

3. a superior generalization ability is expected





CNN: Condensed Nearest Neighbour 3

1. E ← x1, ..., xC , C number of clients, T template set, Eedited set and x1..xC are templates randomly selected from T

2. T ← T − E

3. classify T using E

4. Y set of misclassified templates in T

5. if Y 6= φ then

5.1 E ← E ∪ Y5.2 T ← T − Y5.3 repeat from point 4

6. Stop

3P.E. Hart, The Condensed Nearest Neighbor Rule, IEEE Transactions onInformation Theory, 14, 515-516.





RNN: Reduced Nearest Neighbour 4

1. E ← T

2. for each x ∈ E

2.1 E ← E − x2.2 classify T using E2.3 Y set of misclassified templates in T2.4 if Y 6= φ then

2.4.1 E ← E ∪ x

3. Stop

4G.W. Gates, The Reduced Nearest Neighbor Rule, IEEE Transactions onInformation Theory, 18 (3) 431-433.





Data sets, Protocol and Perfomance

Data setsResults are carried out over Equinox, public Faces Dataset.50 clients have been randomly choosen from the dataset. Each onemade up of 100 samples. A total of 5000 faces images.

ProtocolAll the images have been grouped in two equal size sets, T and t.T has been used as Template Set and t as a complete separatedtest set to assess performance.

PerformanceSystem’s performance has been evaluated as identificationaccuracy : number of correct identified queries over total numberof submitted queries.Results are showed as mean and (std) over six runs.





Results: Accuracy 5

Accuracy over a test set obtained by different selection methods.

Gallery #instances×class TEST

TRAIN 50 (0) 99.62 (0.14)

CNN 7 (3) 97.6 (0.45)SNN 4 (3) 73.66 (3.31)

RNN 17 (9) 98.43 (0.53)

ENN 49 (1) 99.35 (0.27)MDIST 6 (0) 94.15 (0.68)

MDIST 9 (0) 96.56 (0.58)

DEND 6 (0) 89.11 (1.39)

DEND 9 (0) 94.03 (0.70)

5B. Freni, G.L. Marcialis, and F. Roli, Template Selection by EditingAlgorithms: a case of Study in Face Recognition, S+SSPR08, SpringerLNCS5342, 755-764.





Results: Client’s Accuracy

Difficult clients

Gallery #classes #instances TEST MDIST 9 DEND 9

CNN 8 12 96.18 93.78 87.56

RNN 41 20 98.58 96.05 93.25

ENN 50 49 99.35 96.56 94.03

Easy clients

Gallery #classes #instances TEST MDIST 6 DEND 6

CNN 21 4 98.04 96.89 93.67

SNN 31 3 60.70 94.33 90.89

RNN 4 3 98.08 98.96 95.76





Summary

Editing algorithms have been showed as a good alternative to theState-of-the-Art Template Selection techniques.

Results pointed out main characteristics of Editing algorithms:

1. Completly automatic procedures, no futher intervention isneeded by supervisor.

2. Capability to build up variable length galleries, according toclient intrinsic difficulty.

3. Superior identification accuracy.

As a step futher a combined use of both techniques could beinvestigated.




Semi-Supervised Template UpdateTemplate Update with ReplacementResults

Template Update in Biometric

I Problem statement The problem is quite intuitive andconsists in making adaptive the biometric recognition systemsover the time.

Templates collected during enrollment tend to be nonrepresentative by the time, due by the large intra-classvariation.

Performing several enrollment sessions is expensive.

I State-of-the-ArtSemi-supervised paradigms exploit unlabelled samplessubmitted to the system in order to find out “highly genuine”to adapt system’s templates.





S-o-A summary: Semi-Supervised Template Update

Semi-Supervised methods can be summarized by basic operations:

1. Insertion. A “highly genuine” is added into template gallery.

2. Condensing. A template gallery is “fused” in a“super-template”.

Main Cons:

1. Sistematic use of Insertion made up long galleries. For realsystems Memory and Time of Matching are constrains.

2. Condensing absolves constrains but is less representative ofthe original template galleries.

Replacement is a novel basic operation. Able to:

1. Absolve constrains of Memory and Time of Matching.

2. Assure high level of perfomance.





Novel Proposal: Replacement Algorithm

T c indicates the template gallery of client c.M is the maximum number of template slots allowed.|T c | is the length of client’s gallery.Replacement algorithm consists in the following steps:

I for each client c = 1..C

1. x ← i , i as novel input2. s = ms(x ,T c), matching score3. if s > threshold , “highly genuine”

3.1 if |T c | < M then T c = T c ∪ x3.2 else T c = replace(T c , x)

Function replace is made up according to some criteria.





Replacement criteria

I Random Novel template replaces one randomly chosen.I Naive Novel template replaces the one nearest to it.

I FIFO Template galleries are managed as a First In First Outqueue. The new element supersede the oldest one.

I LFU Template galleries are seen as a priority queue LeastFrequently Used. Less used template is substituted by novelone.

I MDIST applied to semi-supervised scenario. A new gallery iscreated adding by a novel template. MDIST is applied topruned one element from the gallery.

I DEND applied in semi-supervised scenario. A new gallery iscreated adding by a novel template, then, a Dendrogram ismade up. Based on Dendrogram an element is removed fromthe gallery.





Data sets, Protocol and Perfomance

Data setsResults are carried out over 12 public Fingerprint datasets. Eachone made up of 100 clients, 8 samples per client, a total of 800 offingerprint images for dataset.

Protocol50 clients have been selected as system’s users. Other 50 asimpostors. For each user 3 sets have been created L, U and T. Lrefers to user’s template gallery, U as unlabelled coming inputs andT as separeted test. U contains genuine and impostors.

PerformanceEqual Error Rate has been calculated over seven runs. EERrepresents the error of the verification system when a number offalse acceptances is equal to a number of false rejections.





Results: EER (M = 3)

DB FVC2002algorithm Db1 Db2 Db3 Db4

initial 7.3 6.9 12.1 6.0

update 3.9 3.8 6.5 1.7

RANDOM 5.2 3.4 6.4 2.2

NAIVE 5.0 4.4 7.6 1.5

FIFO 5.5 3.1 6.2 1.9

LFU 4.2 3.1 6.5 1.7

DEND 4.8 5.3 7.0 1.7

MDIST 3.6 3.1 6.3 2.3





Results: EER vs gallery dimension M 6

6B. Freni, G.L. Marcialis, F. Roli, Replacement algorithms for fingerprinttemplate update, 5th Int. Conf. On Image Analysis and Recognition ICIAR08,June, 25-27, 2008, Povoa de Varzim (Portugal), A. Campihlo and M. KamelEds., Springer LNCS 5112, pp. 884-893.





Summary

Results pointed out:

1. Less EER respect to update without replacement has beenshowed.

2. Perfomance differences among replacement criteria are strongwith small “M”. Which means when strong requirements ofstoring memory is a constrain.

3. MDIST outperfom other criteria, due to the fact that itperforms replacement only if it is necessary.





Conclusions

I Biometric plays a central role in the problem of security andits importance is going to grow.

I Template representativeness is the key for the success of aBiometric system. Templates that “best” represent people’sidentity must be choosen during “enrollment”, as well duringthe “authentication”, “highly genuine” must be detected inthe coming input queries.Among the whole S-o-A explored in this investigation:

1. the employ of Editing for template selection during enrollment2. the use of Replacement for template update during

authenticationI Template representativeness is crucial for other important

issues in Biometric as Sensor-Interoparability in fingerprint.This problem has been addressed too, but for sake of roomthis talk was dedicated just to Editing and Replacement.


Freni - Ph.D. Defense Slides

Technology

template adaptation

issue of template selection

template selection update

core of biometric system

replacement algorithms

launchof biometric passport

et freni

oine ngerprinttemplate