A number of simulations have been performed in order to determine the authentication and Identification efficacy of the proposed method in terms of some standard evaluation criteria, namely false rejection rate(FRR), genuine acceptance rate (GAR) and false acceptance rate (FAR), confusion matrix, precision, sensitivity and specificity, and identification accuracy. For performance comparison, we use some state-of-the-art methods in [21] and [14].
5.3.1 Authentication Accuracy
The objective of the proposed method is not only to identify the closest matched person but also to authenticate whether the unknown person truly exists or not in the database. For this purpose, a distance based similarity measure between the training feature vector of the database and the feature vector of the unknown person is utilized to define a threshold value, which ensures the process of authentication.
We set our threshold value empirically from the mean and standard deviation of the distance vectors of a number of trained persons in the database. For authentication, the distance measure of the person on test is compared with the threshold. If the test is passed by accepting the person on test, the next task is to find out the closest matched person in the training to show the identification efficacy.
For evaluating the authentication ability of the proposed method, the test feature vector is needed to be matched with each template of feature vectors stored in the training database. On the basis of the empirically determined threshold whose level is calculated from the mean and standard deviation of the distance vectors corresponding to all the trained persons in the database, there may several cases:
1. The testing template of a subject with its feature template already trained in the database may be truly accepted.
2. The testing template of a subject with its feature template already trained in the database may be falsely rejected.
3. The testing template of a subject with no training feature template in the database may be falsely accepted.
4. The testing template of a subject with no training feature template in the database may be truly rejected.
The phenomena as mentioned above can be explained in terms of following parameters: Genuine Acceptance Rate, False Rejection Rate and False Acceptance Rate are defined on the basis of threshold whose level is calculated from the mean and standard deviation of the distance vector of a number of the testing persons and the enrolled persons in the database.
Genuine Acceptance Rate (GAR)
GAR is a measure to authenticate legitimate subject and defi ned as the rate of truly accepting the testing template of a subject when the feature template of the subject is trained in the database.
False Rejection Rate(FRR)
FRR signifies the event of denying the identity authentication, which is the rate of falsely rejecting the testing template of a subject although the feature template of the subject is trained in the dataset.
FRR can be derived from GAR as
FRR (%) = (1 − GAR) × 100% (5.1)
False Acceptance Rate (FAR)
FAR stands for the occasion of mistakenly authenticate outsiders, thus expresses the rate of falsely accepting the testing template of a subject even if there is no template of the subject in the training database.
Genuine Rejection Rate (GRR)
GRR is the rate of defining a case of truly rejecting the testing template of a subject when there is actually no template of the subject in the training database. Therefore, as an indication of performance of the human identification system, FAR and FRR rates can be used to calculate the authentication accuracy, which is defined as
5.3.2 Identification Accuracy
Fig. 5.1: Confusion Matrix
For the performance evaluation of the proposed method, parameters considered in our simulation study are: 1) Precision 2) Sensitivity 3) Specificity 4) Identification accuracy. All the parameters as mentioned above can be derived from the confusion matrix, which is a form of representing the result from an identification exercise. The rows in the matrix stand for the actual persons to be tested and columns provide the persons identified by a method. In particular, any [row; column] entry in the confusion matrix indicates the number of cases from the test database that belongs to the person corresponding to the row but identified as the person corresponding to the column. In Fig. 5.1, a general confusion matrix for two persons is shown, where TP, FP, FN and TN are represented for person 1. In general, TPi, true positive for any person i, measures the number of testing cases, which are correctly classified as person i. FPi, false positive for any person i, denotes the number of testing cases, which are incorrectly classified
as person i. FNi, false negative for any person i, indicates the number of testing cases, which are incorrectly classified as other than person i. TNi, true negative for any person i, means the number of testing cases, which are correctly classified as other than person i.
Precision relates the number of positive testing cases, which are correctly classified to the number of identified cases for that particular person. Thus, precision for a person i can be defined as
Sensitivity relates the number of positive testing cases which are correctly identified to the number of testing cases of that particular person. Thus, sensitivity for a person i can be defined as
An identifier, which always indicates positive, regardless of the person of the testing case, provides 100% sensitivity f o r that person. Therefore the sensitivity alone cannot be used to determine the usefulness of the identifier in practice.
Specificity of a person relates the number of negative testing cases, which are correctly classifi ed to the total number of testing cases belonging to other persons rather than that person. Therefore, specificity for a particular person i can be expressed as
Identification accuracy of a person relates the number of testing cases which are correctly identified to the number of total testing cases. Therefore, accuracy for a person i can be written as
For the purpose of comparison, we use state-of-the-art methods in [21] and [14].