Observations on Genuine Scores

Dr. Manfred Bromba
http://www.bromba.com/contacte.htm
Permanent address for citation: urn:nbn:de:0125-2008032564
2002-01-02; revised: 2007-07-29


Introduction

This document deals with FRR (False Rejection Rate) measurements for fingerprint recognition. Each finger is assigned to a unique ID. The personal FRR is defined as the FRR for just one ID. Fingerprint samples are defined as different measurements of the same finger, resp. ID. Score values are the result of comparisons between two fingerprints and describe the similarity. In this investigation, the scores assume values between 0 (least similarity) and 100 (most similarity).

For the determination of the personal FRR, it is obvious that the larger the number of prints (samples) of the same finger, the higher the accuracy of the FRR result. However, if a global FRR which is defined, e.g., as the mean value of personal FRRs, is to be calculated, there are two parameters to be chosen: The number of prints per finger (ID) and the number of different fingers (IDs). The question rises whether it is better to have a data basis with a larger number of IDs or a larger number of samples per ID to get best accuracy with least effort.

Theory

Assume a fingerprint data base with N different fingers (IDs) and M prints (samples) per ID. The request (query) part of this data base may be represented by a matrix of N columns and M rows, each cell containing a request fingerprint. The reference part of the data base contains N sets of reference fingerprints according to the N different IDs. After matching each fingerprint with the corresponding references, a new matrix of M by N  score values is generated. The score values show a random behavior which can be represented by the probability (distribution) function. From this probability function usually the FRR curve is calculated straight-forward, showing the probability that a score value succeeds a certain threshold. To simplify the consideration, only the mean value and the standard deviation are regarded which can be directly calculated from the score values.

There are two main reasons for the random behavior of the score values:

  1. Different finger positioning and pressure
  2. Different personal characteristics (number of minutiae, image quality, behavior, etc.)
The first type of error is nearly independent on the finger (ID) and does always happen. The second type of randomness does only happen between different IDs. As a result, the degree of randomness should be different between different scores from one ID which only shows type 1 errors and the scores between different IDs which represent both type 1 and 2 errors. To see the difference, the matrix representation of the score values has been used to calculate the mean value and standard deviation row by row or column wise. What is to be expected is that the row mean values should show a higher variance since it is affected by two error types in contrast to one with the column mean values. We regard two extremal cases:
  • Only one ID is available: If we increase the number of scores for this finger and take the mean value, only the error type 1 is reduced. If we assume M to approach infinity, this error approaches 0. However the personal effects (error type 2) are not affected at all.
  • Only one request print per ID is available: If we increase the number of scores by increasing the number of different IDs both error types are reduced simultaneously!
From this we may conclude that for the determination of the global FRR it is to be preferred to have more different fingers instead of more prints resp. samples per finger.

However, it should be noticed that it is more expensive to have more different IDs than to have more prints per ID!

Experimental results

Using the data base FGA1010x a score value matrix has been calculated with 68 columns for 68 different IDs which were successful during enrollment and 100 rows for 100 samples per ID. From this matrix the column and row mean values and standard deviations have been calculated using the corresponding MS-Excel functions "Mean" and "Variance". (The standard deviation "sigma" is defined as the square root of the variance.) Table 1 shows the upper left  part of the Excel table with results. Further results are given in Table 2, see Appendix.

What we are looking for are the mean values and standard deviations of the genuine score distribution. This has not to be confused with the measurement errors (belonging to a measurement error distribution) which also have been calculated as standard deviations and which should tend to zero as the number of score values in the test approaches infinity.

The following observations are obvious, see Table 1 and 2:

  • The row mean values show a small variation, in contrast to the column mean values (both yellow). This fact is expressed by a standard deviation (sigma) of 2.86 (sigma of Req sample mean) versus 15.73 (sigma of Ref ID mean).
  • The mean of the column mean values is equal to the mean of the row mean values (= 60.62). This is trivial.
  • The mean of the row standard deviations (22.60, estimated as the square root of the mean of Req sample variance) is significantly higher than the mean of the column standard deviations (16.52, estimated as the square root of the mean of Ref ID variance)
  • The column mean values should only slightly depend on error type 1 because this error has been eliminated by averaging. As a result, the standard deviation of the column mean values (15.73) should mainly represent the error type 2
  • The column standard deviations should solely represent the error 1 by definition. As a result, the square root of the mean of the column variances (16.52) should be a good measure for the error type 1.
  • The square root of the mean of the row variances (22.66) should represent the combined error type 1 and 2.
  • The sum of the squared type 1 and 2 errors (= variances) approaches the squared combined error:
16.52²  + 15.73² ~ 22.81² ~ 22.66²
  • The square root of the mean of the row (= Req sample) variances (22.66) only shows a small deviation to the global standard deviation which is given by 22.68
  • Suppose a test has been made with one score sample per ID (the first one in the tables) and 68 IDs. Then the actual mean value would have been 65.25 (compared to 60.62 +- 2.86) and the actual standard deviation 21.44 (compared to 22.66).
  • Suppose a test has been made with only one ID (the first one in the tables) but 100 samples. Then the actual mean value would have been 89.45 (compared to 60.62 +- 15.73) and the actual standard deviation 12.75 (compared to 16.52).
  • The last two points reveal that both scenarios deliver different distributions with personal distributions being "more narrow" than global distributions. (This does not automatically imply that a personal distribution is better with respect to FRR since the personal mean value should be as high as possible, but often is not.) The square root of the mean of the variances of the personal score distributions (16.52) is NOT equal to the standard deviation of the global score distribution (22.68)!
  • The second observation is that the measurement error (standard deviation) for the first test scenario is significantly smaller than for the second one:

  •  
    Table 0: Results for different test scenarios
    Calculation method Mean of genuine score Sigma of measurement error Variance of genuine score Sigma of measurement error
    Global average:
    6800 samples
    60.62
     
    514.17
     
    ID average:
    68 IDs,
    100 samples
    60.62
    2.86
    513.56
    77.72
    Test scenario 1:
    68 IDs;
    1 sample
    65.25
    (2.86)
    459.56
    (77.72)
    Sample average:
    68 IDs,
    100 samples
    60.62
    15.73
    272.95
    127.42
    Test scenario 2:
    1 ID,
    100 samples
    89.45
    (15.73)
    162.51
    (127.42)

Conclusions

  • Although the number of IDs is smaller than the number of prints per ID, a test based on 68 IDs and 1 sample delivers more accurate results than a test based on 1 ID and 100 samples. For the planning of tests this means that it is more advisable to have a large number of participants than a large number of samples per finger (although this is easier to achieve).
  • Due to the strong personal influences, each ID must be represented by the same number of samples when calculating global characteristics. Alternatively, the mean value of personal characteristics may be taken, provided it delivers the desired result. (Example: The average over all Ref ID means delivers an unbiased global mean value. This is not true for the average over all Ref ID variances which does not estimate the correct global variance. However, there should be no problem when calculating the genuine distribution or the FRR from the personal genuine distributions or the personal FRRs by averaging.)

Comments

All results are based on mean values and standard deviations which represent the position and the width of the score distribution, respectively. (If the distribution type were known a priori, it could eventually be determined completely by mean and standard deviation.) To obtain a low FRR, the mean of the genuine score distribution should be as high as possible and, simultaneously, the standard deviation of the genuine score distribution should be as small as possible. (This is a necessary condition, but it is not sufficient unless the distribution function has a known simple form. Especially the tails of the distribution which are most important for small FAR and FRR values, may show remarkable deviations.)

Appendix

The following tables show a part of the score value matrix together with the mean and variances of the columns (Ref IDs) and rows (Req samples).
 
Table 1 Ref ID: 10010 10013 10014 10016 10022 10048 10054 10059 10061 ...
Req sample:                      
variance   162.51 296.57 128.20 446.77 244.62 309.44 218.79 183.66 116.67 ...
  mean 89.45 57.57 29.76 76.09 64.19 51.79 68.43 83.57 56.98 ...
459.56 65.25 100 22 34 93 74 31 59 100 65 ...
563.63 61.34 20 73 45 100 98 33 95 83 70 ...
457.79 65.87 100 28 40 73 74 31 73 61 70 ...
466.60 62.03 74 31 41 71 63 76 75 40 53 ...
372.03 65.76 100 41 44 73 100 36 55 93 59 ...
542.87 65.38 98 74 28 77 67 60 78 98 78 ...
537.31 65.97 100 66 19 86 82 71 59 71 83 ...
434.19 64.85 88 66 40 78 84 54 53 90 69 ...
495.86 64.59 91 84 13 75 62 30 89 78 62 ...
574.25 61.56 93 28 35 0 68 78 58 89 67 ...
477.86 61.81 76 80 27 52 49 60 50 80 51 ...
542.73 62.66 75 85 35 56 76 79 56 86 49 ...
611.21 59.25 100 42 13 85 55 56 100 74 40 ...
619.94 61.71 83 60 28 65 79 61 79 79 39 ...
513.31 63.97 100 58 56 90 74 75 46 83 74 ...
476.46 63.15 100 28 33 65 66 71 29 82 61 ...
453.72 60.84 75 59 52 68 58 85 40 87 46 ...
483.74 58.81 82 30 34 61 62 16 53 77 44 ...
423.55 61.03 97 65 37 75 49 60 97 91 53 ...
483.77 62.41 71 95 30 87 57 20 61 84 49 ...
447.42 63.96 86 66 31 100 75 78 77 83 62 ...
340.50 60.87 83 61 40 54 82 65 86 90 69 ...
466.72 64.76 90 92 29 100 86 72 49 88 1 ...
440.97 58.96 100 63 34 41 68 46 58 83 55 ...
504.97 62.32 100 49 47 72 38 50 100 84 58 ...
483.97 61.28 81 59 11 83 69 28 61 85 31 ...
... ... ... ... ... ... ... ... ... ... ... ...

 
 
Table 2
Ref ID mean: Average of one ID over all samples
Req sample mean: Average of one sample over all IDs
Mean of Ref ID mean: Average of all Ref ID means etc.
Global score mean 60.62
Global score variance 514.17
Global score sigma 22.68
Mean of Ref ID mean 60.62
Variance of Ref ID mean 247.52
Sigma of Ref ID mean 15.73
Mean of Ref ID variance 272.95
SQRT of mean of Ref ID variance 16.52
Variance of Ref ID variance 16236.13
Sigma of Ref ID variance 127.42
Mean of Req sample mean 60.62
Variance of Req sample mean 8.17
Sigma of Req sample mean 2.86
Mean of Req sample variance 513.56
SQRT of mean of Req sample variance 22.66
Variance of Ref sample variance 6040.20
Sigma of Ref sample variance 77.72
Number of Ref IDs 68
Number of Req samples 100