Study shows facial recognition biased towards certain skin tones
The systems struggled with darker-skinned men, while darker-skinned women seemed to be more difficult for the algorithm to read, with a 35% inaccuracy rate.
CAMBRIDGE, MASSACHUSETTS — A study conducted by MIT's media lab shows something shocking about facial recognition — the systems are better at identifying light-skinned males than darker-skinned females.
MIT researcher Joy Buolamwini built a data set of 1,270 faces from different countries that included a large number of females in public office.
The faces included three African nations with predominantly dark-skinned populations, and three Nordic countries with mainly light-skinned people.
To see how well the systems are at identifying faces, three different facial recognition systems made by Microsoft, IBM, and Megvii of China were put into the trial.
Each face was assigned with a rating for skin type based on the Fitzpatrick rating system, a six-point rating system which dermatologists use for classifying different shades of skin.
The results showed the systems had an easy time with lighter-skinned men, only misidentifying the gender in about 1% of all photos, with about a 7% inaccuracy rate for lighter-skinned women.
The systems struggled with darker-skinned men with 12% inaccuracy, while darker-skinned women had about a 35% inaccuracy rate.
NEXT ON TOMONEWS
How to escape from the world's fastest predators