3.5 Performance and Experimentation Results
3.5.1 Experimentation Data and Requirements
For the purposes of this investigation, 300 pseudo-users were created from five separate, genuine biometric databases, taking the corresponding values from each database and deeming them to derive from a given virtual user. These databases are FVC2004 represented fingerprint samples [91], CASIA-IrisV3 presented iris samples [92] and FERET presented face samples [93]. For each person, three samples are employed for enrolment, and for testing, each user provided five samples. In the authentication phase, then, five tests were created for each user, and for each test, five samples were provided for all 300 pseudo-users, as shown in Table 3.3.
Table 3.1 Example enrolment data.
User id Samples Face
Fingerprint Iris
Thumb Index Finger Right Left
1
S1
S2
S3
Passive forgeries were also investigated by requiring each of the 300 users to try to gain access by falsely claiming the identity of each of the other 299 users, which was
adopted to check the system‘s susceptibility to the attack. Table 3.1 shows the enrolment of a genuine user. There are eight samples per user for each modality (three for training and five for testing), chosen randomly.
98
Table 3.2 Samples for Testing
Table 3.2 presents an example of the testing method, in which five tests were provided for each user; each test could cover five different modalities (face, thumb, index finger, right iris, left iris). The performance of the system for genuine users (Table 3.3) shows five tests for each user, with each test requiring five samples to be provided.
Table 3.3 Sample performance; √ indicates success for given modality or polynomial order; X indicates failure.
A set of thresholds for each component modality classifier—chosen randomly to test the results of FAR and FRR for each biometric modality—was used to extract the threshold for the polynomial cross curve between FAR and FRR, known as the equal error rate (ERR) (Table 3.4).
User id Test Face
Fingerprint Iris
Thumb Index finger Right left
1 1 2 3 4 5
User ID Test Face
Fingerprint Iris Secret Share Techniques
Thumb Index
finger Right Left
Linear equation Quadratic equation Cubic equation 1 1 X √ √ √ √ √ √ √ 2 X √ √ √ √ √ √ √ 3 X √ √ √ √ √ √ √ 4 √ X √ √ √ √ √ √ 5 √ X √ √ √ √ x x
99
Table 3.4 FAR and FRR results for each biometric modality separately.
Biometric Modality FRR FAR ERR
Fingerprint 5.9% 3.2% 4.89%
Face 12.9% 6.2% 10.76%
Iris 1.79% 0.034% 0.25%
3.5.1.1 SSS Algorithm Requirement
The major goal of this project was to develop a working implementation of Shamir‘s SSS Algorithm [11], representing the cryptographic aspect of a multimodal biometric system. The SSS Algorithm can be used to generate a number of secret key shares from a chosen secret (which could be any number), and these key shares are then bound to each biometric modality. Successful verification of each individual biometric modality should release the appropriate key share. A combination of the minimum key shares required (two) should reconstruct the original secret, leading to authentication. If the required number of key shares is not successfully combined, the secret cannot be reconstructed, and the user will be rejected by the system.
Five different biometric modalities should be used to facilitate implementation of the SSS algorithm. Using the linear equation technique, the algorithm should be incorporated into the system in such a way that a user can only be authenticated following successful verification of at least two biometric modalities, based on their chosen options. If a user fails to verify at least two biometric modalities, they should be rejected by the system (and so on for the other techniques).
3.5.1.2 Verification Algorithms Requirement
In conjunction with the SSS algorithm for recognising a user, three individual verification algorithms should be implemented for use by multimodal biometric systems. These should include a face recognition algorithm, an iris recognition
100
algorithm and a fingerprint recognition algorithm. The algorithms should be able to identify a user by comparing a sample presented by the user against a stored template, making a decision on the basis of degree of similarity.
3.5.1.3 Database Requirement
A database should be created to accommodate storage of biometric templates for different users. Secret key shares created by the SSS algorithm should also be stored in the database and linked to the respective individual biometric modalities in the database. During the verification process, user samples presented to the system should be compared against their respective templates stored in the database, and the appropriate key shares should be released if verification was successful.
3.5.1.4 Testing Requirements
The system should be tested for technical performance of the SSS algorithm and of the individual verification algorithms. From these tests, overall performance of the system can be deduced and, most importantly for authenticating users, false acceptance rates and false rejection rates of the system can be tested.
3.5.1.5 Software implementation of feature extraction algorithms
In this work, three feature extraction algorithms have been used, named as Eigneface algorithm for face feature extraction, 2D-Gabor filter for iris feature extraction, and minutiae based feature extraction for fingerprints. The Shamir secret scheme algorithm also has been used for key generation indirectly form the multimodal biometric system used in this work. All these algorithms have been implemented using the software MATLAB. The codes for implementation of these algorithms have been generated by the researcher himself. In addition, in order to gain the expert knowledge about the
101
coding for feature extracting, the following sources have been consulted: AT&T Laboratories Cambridge (online source) implemented by Luigia Rosa [176] for the implementation of face recognition features and the fingerprints recognition implemented by Wuzhili [177]. However, the implementation of 2D-Gabor filter algorithm has been implemented in MATALB by researcher from the scratch.