Facial Recognition Failure

Artificial Intelligence (AI) has come a long way, with one of its most notable advances being that of facial recognition software. 

Despite many years of development, the software used by big companies such as Facebook and Microsoft often still fails to accurately identify the faces of women and those with darker skin tones. The question is, what is it about dark-skinned faces that AI has trouble with?

As part of her research at MIT, computer scientist Joy Buolamwini, who herself is an African-American woman, discovered that recognition software primarily works well on faces of white males. Moreover, she found that the software mainly struggled to identify facial features of women, especially those with dark skin tones. 

In order to conduct her research, Buolamwini looked at several facial recognition software systems developed by Face++, Microsoft and IBM. She even shared her findings with IBM who, so it seems, have made improvements to their software as a response. 

According to her, this is evidence that facial recognition of dark-skinned faces is not dependent on physical features, such as the reflection of light on the skin, but rather on expanding the facial image databases.

However, this was not the first time that it became clear that facial recognition software still faced major difficulties. In 2015, Google came under scrutiny after its recognition software on Google Photos identified a Black man’s face as a ‘gorilla’. 

Unfortunately, all that Google did to fix the problem was removing the labels referring to ‘gorilla’ from their database; they have done little to address the actual underlying issue.

But how does facial recognition actually work, and what is its purpose? Facial recognition software is used by Facebook to recognise faces of people you might want to tag; to enable phones to unlock the screen by only using your face, but most importantly, it has been used by law enforcement to scan faces to identify possible criminals or terrorists. 

In order to achieve these results, the software takes facial features in a given image, such as shapes, and compares it to faces in a database. The more features which can be matched with a face on the database, the more accurate the result will be.

So why does facial recognition have such a low accuracy-rate for dark-skinned faces? First of all, facial recognition software can only compare the target image to an existing database containing images of faces gathered during the development of the software. Thus, if the majority of faces on the database are from white males and females, the software cannot accurately match facial features from dark-skinned faces to existing data. This is indeed what has happened with virtually all facial recognition software thus far. It is likely that the majority of staff members who were working on the software during its developmental stage, were white males.

In order to make facial recognition software less racially biased, Google decided to gather facial images through a greater mix of people with various skin tones, which could later be used for the face unlock feature on their new Pixel 4 smartphone. 

In theory this sounds great, however, the contractors that carried out the data collection were allegedly rushing people, including black homeless people, into signing the consent form. Most of them, especially the homeless, didn’t even know what was happening. By offering incentives such as 5$ gift cards or letting people play with the phone, which unbeknownst to them recorded their faces, the contractors were able to collect facial images. Google has now ended their data collection as people were calling for an investigation into the matter.

On the whole, AI-supported facial recognition software is a useful addition to our daily life, but we need to keep in mind that it’s still a work in progress. 

It is good to see some progress being made in addressing the racial biases of facial recognition software, but without good practice surrounding the gathering of new images, the future of facial recognition is set to remain controversial.

image source: Karelnoppe/Shutterstock