Anti-dark skin bias of facial recognition software

Discrimination against people with dark skin is quite common in many parts of the world, especially in the United States of America. This is despite the fact that dark skin is arguably biologically superior to pale skin. Sadly, the victims of this kind of discrimination usually suffer some setbacks – such as denied opportunities, mistreatment, […]

Discrimination against people with dark skin is quite common in many parts of the world, especially in the United States of America. This is despite the fact that dark skin is arguably biologically superior to pale skin. Sadly, the victims of this kind of discrimination usually suffer some setbacks – such as denied opportunities, mistreatment, racial profiling and so on – even though the discrimination is not based on any rationale or fair logic. The act is social in nature; being rooted in vanity in most cases.

Now, what if the discriminating object is not a human being per se, but a software? This will even be more worrisome because it means discrimination has been carried to a new level; aka some notches beyond the social realm. It is of course my hope that the reported discrimination is not intentional.

The software apps in question are those used for recognising people’s faces. This Facial Recognition Technology (FRT) allows some hardware to match your face with some database. The exploding use of FRT; especially in China, is obviously a result of the technology’s immense potential usefulness, coupled with the relatively weak human rights laws in that country. In fact, some of the uses to which FRT is being applied in China will cause serious human rights issues in the West.

FRT is conceptually quite simple. At some point, a picture of your face would have been taken and specific geometric/topological features extracted and stored along with your personal data, bio data, or metadata. By searching through such a database and making comparisons, any newly captured facial feature can be evaluated to explore a match and present identification.

This column mentions some of the various uses of FRT in the 17 July 2017 article. In the most benign usage, FRT is invaluable for verifying customer identities in business transactions such as banking. Installing FRT hardware in students’ dormitories can discourage criminals. FRT is also quite useful in tailoring marketing efforts based on some database. Security and surveillance represents a critical area where FRT is also very useful. Here, the identities of people are obtained before they are allowed entry into a building or are matched with some information in a database for the purpose of catching criminals. 

Also, a few universities are installing FRT to identify ghost exam takers trying to sit exams for other students. Obviously, FRT is a law enforcer’s best friend in any country.

The above referenced article suggests that FRT is becoming a daily feature in the lives of many Chinese. Some of the more benign uses in that country include deployment on streets, in subway stations, and at airports. I have also read that FRT has been used in China to identify joggers who attempt to take shortcuts at a running course. With the extensive use of FRT in the manner just discussed, the app has to be accurate. That is, the consequences could be dire if people’s faces are wrongly identified.

Well, as it turns out, the most popular FRT apps, such as those developed by Microsoft, IBM, and Megvii of China are grossly inaccurate for identifying the faces of people with dark skin, thereby complicating the racial discourse. (The app by Microsoft is reportedly more accurate than those from the other two companies.) This is quite sad because these apps have moved from development to deployment, with extensive usage on the field as described above.

In a nutshell, and as ominous as it sounds, facial recognition is accurate if you are a white guy. That is, the darker the skin, the higher the errors. More specifically, the errors are as follows: white male (less than one per cent), white female (seven per cent), dark-skin male (12 per cent), and dark-skin female (35 per cent). Note that you do not have to be ebony-black to be of dark skin in this context! 

These findings were reported by Ms. Joy Buolamwini, an African-American graduate researcher at the Massachusetts Institute of Technology in the US. She noted that the FRT apps developed by the three companies mentioned above “worked well on her white friends but could not recognise her face at all.” 

The poor performance of these apps is believed to be rooted in the bias that is built into the technology, and Buolamwini has made it her mission to help fix the problem. Microsoft is proactively working with her, including undertaking a joint research work, to resolve the issues. IBM reportedly said in a statement to Buolamwini that the company had steadily improved its facial analysis software and was “deeply committed” to “unbiased” and “transparent” services. This month, the company said, it would roll out an improved service with a nearly 10-fold increase in accuracy on darker-skinned women. However, according to Buolamwini, “Megvii, whose Face++ software is widely used for identification in online payment and ride-sharing services in China, did not reply to several requests for comment.”

Obviously, FRT and other so-called artificial intelligence (AI) software are driven by data, such as the ones obtainable from your social media accounts. For accurate results, such data cannot be biased in any way.