4.5 Typical false match rates
A false match rate is expressed as a statistic such as '1 in 1000'. A rate of 1 in 1000 means that if a sample of biometric data is compared with random selections of other data, a false match occurs, on average, once every 1000 comparisons. False match rates vary widely between different biometric systems. The following figures are taken from Mansfield and Rejman-Greene (2003).
Using good-quality fingerprints, a false match rate for single prints can be around 1 in 100,000. This rate can be improved by using more than one finger. Low-quality prints give worse rates.
Face recognition gives a false match rate of around 1 in 1000.
For single-eye iris recognition, Mansfield and Rejman-Greene (2003) quote a false-match rate of 1 in 1,000,000. Using two eyes improves the rate considerably.
These rates do not allow you to predict how many false matches there will be in any single small-scale test. For instance, a system with a false match rate of 1 in 1000 when tried on a group of 10 or 100 people might or might not produce false matches. The outcome is unpredictable. However, with a large group the result becomes more predictable. For instance, a system with a false match rate of 1 in 1000 when tried on a group of 1,000,000 people will produce about 1000 false matches. This figure comes from dividing the group size, 1,000,000, by 1000, which comes from the false match rate of 1 in 1000. However, this method is only an approximate one, and can be used only with large groups. What we mean by 'large group' or 'small group' depends on the false match statistic. For instance, if the false match rate is 1 in 100, a group size of 10 000 is quite large; but with a false match rate of 1 in 100,000, the same group size of 10,000 is small.
Activity 19 (self-assessment)
The false match rate of a particular biometric system is 1 in 100. A sample of data is compared against 1,000,000 other randomly chosen pieces of data. Approximately how many false matches can we expect?
Compared to the false match rate of 1 in 100, the group size of 1,000,000 qualifies as large, so we can use the approximate method described above. To find the answer we need to divide 1,000,000 by 100. This gives 10,000. This is approximately the number of false matches we can expect.
The lesson to draw from the last activity is that if the number of comparisons is sufficiently large, then a lot of false matches can result, even if the false match rate appears to be quite low. This needs to be borne in mind when thinking about identification systems for a whole population.
Deciding which biometric system to use is not just a matter of picking the one with the best false match statistics. The figures given above are laboratory figures, and whether the same performance can be obtained in practice, when conditions are not ideal and the participants may be impatient or uncooperative, is another question. Also, there is the question of how a system would work in practice. Would it require people to stop and do something, potentially causing bottlenecks at busy places such as airports, or could it operate on people while they were doing something else, without their being aware that an identity check was being carried out, as with face recognition? Considerations like these play a part in deciding which system to use.