4.4 False matches and false non-matches
I mentioned earlier that one of the essential components of an identification system was a reliable procedure for comparing data. Ideally, one person's biometric data would never be mistaken for another's, and one person's biometric data would always match another sample from the same person. Unfortunately such perfection cannot be achieved in practice. Errors happen because samples of biometric data taken from someone on different occasions are almost certain to be different, just as two specimens of my signature are different. This means that establishing someone's identity by looking for an exact match between biometric samples is not possible.
The practical solution is to look for a certain level of similarity rather than exact sameness. To count as a match, two samples of biometric data need to be sufficiently similar rather than identical. What are the implications of this fact? We can get an insight into their implications by thinking about Figure 7. The two faces represent samples of biometric data, and the arrow represents a comparison. Clearly the samples are different, but are they from the same person?
Let us suppose that these samples are from the same person. To be fairly sure of getting a 'Yes' verdict in the comparison, we should set the threshold of similarity at a fairly low level. That is, our comparison system should be quite tolerant of differences between samples of biometric data. That increases the likelihood of the right verdict: 'Yes. These are from the same person.'
But suppose the two samples of data in Figure 7 are from different people. If the comparison is tolerant of differences, these samples will be judged to be from the same person, which is the wrong answer (in this case). To increase the likelihood of the correct verdict when the samples are from different people, the comparison needs to be less tolerant of differences.
As you can see, there are conflicting requirements here. To ensure that samples of data from the same person match we need a fairly relaxed criterion of similarity. But to ensure that difference is detected, we need a strict criterion.
These two types of error are known as false match and false non-match. They are illustrated in Figure 8.
A false match is when two pieces of biometric data from different people are judged to be from the same person, as in Figure 8(a). This type of error is sometimes called a false accept. It results from a comparison that is too tolerant of differences.
A false non-match is when two pieces of biometric data from the same person are judged to be from different people, as in Figure 8(b). This type of error is sometimes called a false reject. It results from a comparison that is too intolerant of differences.
Activity 18 (self-assessment)
Which type of error is increased by relaxing the matching criterion, and why?
Which type of error is increased by making the matching criterion stricter, and why?
False match. Relaxing the matching criterion means making the matching process more tolerant of difference. This increases the likelihood that biometric data from different people will be judged to be from the same person.
False non-match. Making the matching criterion stricter means that the comparison is less tolerant of difference. This means that we are more likely to decide that samples of data from the same person are from different people.
False matches and false non-matches are not caused by the use of IT-based systems or by the use of biometric data. They can happen when any type of data is assessed for similarity by any type of inspector – human or otherwise.
The performance of biometric systems is gauged by various statistics. One of these statistics is the false match rate. The following section looks at some typical false match rates.