handwritten signatures authentication using anns committee machines m.heinen, f. osório and p....

22
October 2007 1 dwritten Signatures Authentication using ANNs Committee Machines M.Heinen, F. Osório and P. Engel Handwritten Signatures Authentication Using Artificial Neural Networks Committee Machines Milton Roberto Heinen Milton Roberto Heinen - Applied Computing / Unisinos and II / UFR - Applied Computing / Unisinos and II / UFR Prof. Dr. Prof. Dr. Fernando Fernando S. Osório - S. Osório - Applied Computing / Unisinos Applied Computing / Unisinos Prof. Dr. Paulo M. Engel - Informatics Institute / UFRGS Prof. Dr. Paulo M. Engel - Informatics Institute / UFRGS Applied Computing - PIPCA / Unisinos Applied Computing - PIPCA / Unisinos Informatics Institute - II / UFRGS Informatics Institute - II / UFRGS UNISINOS University and UFRGS University - Brazil UNISINOS University and UFRGS University - Brazil Web: http://inf.unisinos.br/~osorio/ Web: http://inf.unisinos.br/~osorio/ NeuralSignX NeuralSignX CLEI 2007 San José - Costa Rica - October 2007

Upload: brittney-shana-knight

Post on 27-Dec-2015

219 views

Category:

Documents


0 download

TRANSCRIPT

October 2007

1

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

Handwritten Signatures Authentication Using Artificial Neural Networks

Committee Machines

Milton Roberto HeinenMilton Roberto Heinen - Applied Computing / Unisinos and II / UFRGS- Applied Computing / Unisinos and II / UFRGS

Prof. Dr. Prof. Dr. Fernando Fernando S. Osório -S. Osório - Applied Computing / UnisinosApplied Computing / Unisinos

Prof. Dr. Paulo M. Engel - Informatics Institute / UFRGSProf. Dr. Paulo M. Engel - Informatics Institute / UFRGS

Applied Computing - PIPCA / UnisinosApplied Computing - PIPCA / UnisinosInformatics Institute - II / UFRGSInformatics Institute - II / UFRGS

UNISINOS University and UFRGS University - Brazil UNISINOS University and UFRGS University - Brazil Web: http://inf.unisinos.br/~osorio/Web: http://inf.unisinos.br/~osorio/NeuralSignXNeuralSignX

CLEI 2007 San José - Costa Rica - October 2007

October 2007

2

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

Handwritten Signatures Authentication Using Artificial Neural Networks

Committee Machines

Milton Roberto HeinenMilton Roberto Heinen - Applied Computing / Unisinos and II / UFRGS- Applied Computing / Unisinos and II / UFRGS

Prof. Dr. Prof. Dr. Fernando Fernando S. Osório -S. Osório - Applied Computing / UnisinosApplied Computing / Unisinos

Prof. Dr. Paulo M. Engel - Informatics Institute / UFRGSProf. Dr. Paulo M. Engel - Informatics Institute / UFRGS

Applied Computing - PIPCA / UnisinosApplied Computing - PIPCA / UnisinosInformatics Institute - II / UFRGSInformatics Institute - II / UFRGS

UNISINOS University and UFRGS University - Brazil UNISINOS University and UFRGS University - Brazil Web: http://inf.unisinos.br/~osorio/Web: http://inf.unisinos.br/~osorio/NeuralSignXNeuralSignX

CLEI 2007 San José - Costa Rica - October 2007

Presented by Cássia Nino - Applied Computing / Unisinos

October 2007

3

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

Presentation Topics

Agenda:

1. Motivation and Context

2. Handwritten Signature Authentication

3. NeuralSignX System

a) Modules: Acquisition, Pre-Processing, ANN Classification

b) Specialists Committee

4. Description of Experiments

5. Signature Authentication Results

6. Conclusions

October 2007

4

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

1. Motivation and Context

Motivation: Security

How to guarantee a legitimate access to resources:Data access, Bank access, Passport, Credit Card, etc.

Do you really are who you claim to be ??

Context: Biometric Systems Signature

Identify: who you are Off-LineVerify: authenticate one user Scanned Signature (static)

Physical features: On-Line Fingerprints, hand, eyes, ... Digital Pen (dynamic)Behavioral features: Signature, keyboard typing, voice, ...

October 2007

5

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

1. Motivation and Context

Motivation: Security

How to guarantee a legitimate access to resources:Data access, Bank access, Passport, Credit Card, etc.

Do you really are who you claim to be ??

Context: Biometric Systems Signature

Identify: who you are Off-LineVerify: authenticate one user Scanned Signature (static)

Physical features: On-Line Fingerprints, hand, eyes, ... Digital Pen (dynamic)Behavioral features: Signature, keyboard typing, voice, ...

October 2007

6

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

1. Motivation and Context

Motivation: Security

How to guarantee a legitimate access to resources:Data access, Bank access, Passport, Credit Card, etc.

Do you really are who you claim to be ??

Context: Biometric Systems Signature

Identify: who you are Off-LineVerify: authenticate one user Scanned Signature (static)

Physical features: On-Line Fingerprints, hand, eyes, ... Digital Pen (dynamic)Behavioral features: Signature, keyboard typing, voice, ...

ThisWork

October 2007

7

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

2. Handwritten Signature Authentication

Signature Authentication

Binary Classification:- True: Authentic [ not exactly the same as a previously known ]- False: Random, Traced, Skilled

On-Line Signatures

- Dynamical and Temporal Information- Easy to access input devices... tablet + pen Becoming more usual nowadays:

Tablet PCsPalm TopsCell phonesPortable GamesLow Cost Digitizing Tables

October 2007

8

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

October 2007

9

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

LOGIN=MILTONLOGIN=MILTON116 478 0 21:42:23:821116 478 0 21:42:23:821116 478 1 21:42:23:831116 478 1 21:42:23:831125 467 1 21:42:23:842125 467 1 21:42:23:842134 456 1 21:42:23:852134 456 1 21:42:23:852145 441 1 21:42:23:873145 441 1 21:42:23:873168 444 0 21:42:23:883168 444 0 21:42:23:883198 373 1 21:42:23:894198 373 1 21:42:23:894

Position XPosition YPen Up/DownTime Stamp

October 2007

10

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

Collected Data...

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

Signatures Database

TOTAL: 2440 Signatures

1800 authentic signatures (30 per user, 60 users)320 false (traced forgeries)320 false (skilled forgeries)

Select 1 user to authenticateAll other signatures can be used as non-authentic

The user signatures are divided into: training set / validation setTypical authentic user training data set is composed by 10 to 15 signatures

October 2007

11

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

Position and Scale Adjust

October 2007

12

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

Features Extraction

- Signature elapsed time (1 value)- Number of pen lifts (1 value)- Signature length (1 value)- Medium and maximum pen velocity (2 values)- Number of pen X, Y direction changes (2 values)- Cardinal points measure (8 values)- Pseudo-vectors total length (8 values)- Signature density grid (48 values)- Vertical and horizontal intersections (26 values)- Sequential sampling (16 values)- Symmetry measure (2 values)

Total: 117 values! (or even more)

October 2007

13

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

Features Extraction

- Signature elapsed time (1 value)- Number of pen lifts (1 value)- Signature length (1 value)- Medium and maximum pen velocity (2 values)- Number of pen X, Y direction changes (2 values)- Cardinal points measure (8 values)- Pseudo-vectors total length (8 values)- Signature density grid (48 values)- Vertical and horizontal intersections (26 values)- Sequential sampling (16 values)- Symmetry measure (2 values)

Total: 117 values! (or even more)

Large Input Space:Reduced using PCA and other techniques presented in some of our previous worksIEEE/IJCNN 2006, SBIA/SBRN/WCI 2006 and SBC/ENIA 2005

October 2007

14

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning

- Classification

Features Extraction

- Signature elapsed time (1 value)- Number of pen lifts (1 value)- Medium and maximum pen velocity (2 values)- Cardinal points measure (8 values)- Signature density grid (48 values)- Vertical and horizontal intersections (26 values)

Total: 86 input values

In this work:we selected these 6 features represented by 86 values

NeuralsignX pre-processing module is available at:http://inf.unisinos.br/~osorio/NeuralSignX/nsx.html

October 2007

15

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module ANN - Artificial Neural Networks

Neural Learning:Neural Learning: Select 1 user as authentic

Learn how to classify as

- Authentic Signature

- Non Authentic Signature

Classification:Classification: Committee of three specialists (5 Neural Nets each)

Inputs: 86 valuesOutput: 1 binary (Authentic or Not)

October 2007

16

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

3. NeuralSignX System

NeuralSignX Modules:

1. Acquisition Module

2. Pre-Processing Module

3. ANN Module

- Learning: > MLP with RProp

- Classification: Main goals

> Improve generalization

> Reduce FP (False Positive)

FP: False Positive - False signature classified as authentic (accepted)

FN: False Negative - Authentic signature classified as false (refused)

Inputs: 86 values are divided into 3 Esps

Output: 1 combined binary output

Simulation Tool: SNNS Simulator

Committee of three specialists

October 2007

17

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

4. Description of Experiments

Experiments:

Signatures Database

1800 Authentic

320 Traced

320 Skilled=================

2440 Signatures (usually from 10 to 30 different signatures per user)

10 Users were authenticated10 Folds Cross-Validation 10 Runs each folder (≠ initializations)1000 experiments

Results:HIT = % of correct answerFPR = False Positive Rate (accepted)FNR = False Negative Rate (refused)

Single ANN: 86 inputs, 5 hidden, 1 output versusCommittee: 3 specialists (5 ANN in each one)

Experiment Results:

October 2007

18

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

5. Signature Authentication Results

Results: One Single ANN

MSE = Mean Squared ErrorHIT = % of correct answerFPR = False Positive Rate (accepted)FNR = False Negative Rate (refused)

Top:

FPR = 0.14 %FNR = 23.00 %

Global mean

HIT = 99.86 %FPR = 0.05 %FNR = 8.03 %

Single ANN

High FNR error rate

October 2007

19

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

5. Signature Authentication Results

Results: Specialists committee

MSE = Mean Squared ErrorHIT = % of correct answerFPR = False Positive Rate (accepted)FNR = False Negative Rate (refused)

Top:

FPR = 0.00 %FNR = 13.33 %

Global mean

HIT = 99.95 %FPR = 0.00 %FNR = 4.00 %

Committee

Lower FNR error rate

October 2007

20

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

5. Signature Authentication Results

Comparing Results: Single x Specialists committee

MSE = Mean Squared ErrorHIT = % of correct answerFPR = False Positive Rate (accepted)FNR = False Negative Rate (refused)

Top:

FPR = 0.00 %FNR = 13.33 %

Global mean

HIT = 99.95 %FPR = 0.00 %FNR = 4.00 %

Committee

Top:

FPR = 0.14 %FNR = 23.00 %

Global mean

HIT = 99.86 %FPR = 0.05 %FNR = 8.03 %

Single ANN

Specialists committee performs better!

October 2007

21

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

6. Conclusions

Conclusions and Future Work

We proposed: Authentication using ANN Committee Machines

We obtained: a more robust and secure authentication system

- Pre-Processing: Good Features = Good Results- Classification: Specialist Committee are able to... Improve Generalization (HIT) ! Reduce Errors (FPR and FNR) !

Future Work: How to Improve Security ? Hybrid User Authentication System

"Combine SpecialistsCombine Specialists but also combine different biometric techniques"combine different biometric techniques"

Composed authentication: Signature, Eye scanning, Fingerprints, etc.

October 2007

22

Handwritten Signatures Authentication using ANNs Committee MachinesM.Heinen, F. Osório and P. Engel

CONTACT INFORMATION

UNISINOS University - BrazilUNISINOS University - Brazil

Applied Computing Research Post-grad ProgramApplied Computing Research Post-grad Program - PIPCA - PIPCA

Artificial Intelligence Research GroupArtificial Intelligence Research Group - - GIAGIA

UFRGS University - BrazilUFRGS University - Brazil

Informatics Institute - IIInformatics Institute - II

NeuralSignX Project Web Page:NeuralSignX Project Web Page:

http://inf.unisinos.br/~osorio/NeuralSignX/nsx.htmlhttp://inf.unisinos.br/~osorio/NeuralSignX/nsx.html

Contact:

Milton Heinen, Fernando Osório and Paulo Engel

Web: http://inf.unisinos.br/~osorio/

E-Mail: [email protected]