Yesterday at the CeBIT trade fair in Hanover, Germany, Alibaba chairman Jack Ma demonstrated Alipay’s facial recognition technologies – using his face as payment ID to buy souvenir stamps for the mayor of Hanover. Ma presented his face to a smartphone facial recognition technology (FRT) interface, which confirmed his identity and allowed the payment. Claiming that buying things online is ‘always a big headache’, Ma observed, “You forget your password, you worry about your security. Today we show you a new technology,”

The beta version of Alipay’s FRT system is called ‘Smile to pay’ [Chinese language], and is being developed together with Ant Financial Services (which runs Alipay) and Beijing-based tech company Megvii’s Face++ technology.

Facial recognition algorithms use a number of factors to identify or confirm the identity of an individual, most of which involve the relative disposition of features to each other, calculated to acceptable tolerances of error (and at one point the Federal Bureau of Investigation declared itself ready to accept an error rate of 20%).

New York City University’s Center for Catastrophe Preparedness and Response (CCPR) published an interesting report [PDF] in 2009 detailing some of the more pervasive problems of FRT systems. Taking some of the report’s research to its logical conclusion, it seems that wide take-up of a single FRT system – and most particularly of a situation where multiple systems use a central resource pool – would dramatically lower the accuracy rate of FRT idents within that system, even when the participating end-users are cooperative with the aims of the system and using identical software on near-identical devices. The report reads:

As the size of the identification database increases, the probability that two distinct images will “translate” into a very similar biometric template increases. This is referred to as the biometric double or twin. Obviously, biometric doubles lead to a deterioration of the identification system performance as they could result in false positives or false negatives.’

Ninety per cent of the factors individuating one face from another occur in just 10% of the facial area, and as the ‘enrolled’ database of volunteered face images grows, the differences can become trivial enough to obtain either a false ID (the wrong person succeeds in identifying as someone else) or a double-match (the ‘probe’ picture matches more than one face in the database).

smartgate-melbourne[1]

Neither of the latter two scenarios are fantastical – Australia’s SmartGate FRT technology (pictured right), designed to speed passengers more quickly through airport security, was duped by two similar-looking journalists (Ibid) when first launched at Melbourne Airport in 2002. The two similarly-featured journalists had previously succeeded in duping other facial recognition systems, and successfully got through the SmartGate system after swapping passports.

According to the report, FRT accuracy is ‘very sensitive to the aging effect’, stating: “For 18 to 22 year-olds, the average identification rate for the top systems was 62%, and for 38 to 42 year-olds, 74%. For every ten-year increase in age, performance increases on average 5% through age 63,”

array-of-faces[1]

Additionally women were found to be 10% less likely to be identified accurately by FRT, and Caucasian faces proved to be the hardest to individuate, according to a research study from 2002, which asserted: “Asians are easier [to recognize] than whites, African-Americans are easier than whites, other race members are easier than whites, old people are easier than young people, other skin people are easier to recognize than clear skin people,”