Swiss doctoral student Bogdan Kulynych developed a method that can be used to test how Twitter’s image cropping algorithms prefer certain facial features. Kulinich, who studied at the Security and Privacy Engineering Lab of the cole Polytechnique Fédérale de Lausanne (EPFL), won a cash reward of $3500.
In short, Kulynych found that Twitter’s algorithms favor female-looking, youth with lighter, warmer skin tones and thinner faces when cropping thumbnails.
The initial trigger for their experiments was a report by PhD student Colin Madelan. He had observed that the algorithm with which Twitter automatically crops photos placed his white-skinned face in the center of his photo with a black colleague. The part of the picture that was shown to his colleagues was completely cut out by the software.
$3,500 for the Bug Finder
Further experiments by other Twitter users were confirmed: the algorithm shows racist tendencies. However, not always in the same way, as in another experiment they preferred black faces. Twitter announced that it will investigate the problem more closely and want to improve the algorithm. After some time The company has turned off automatic image cropping.
As if that wasn’t enough, it also promised $3,500. reward of For those trying to find new ways to demonstrate how Twitter’s artificial intelligence inadvertently discriminates against certain groups of people. The award is now presented to Kulinich at the DEF CON IT Security conference in Las Vegas.
The source of the problem is the bias of the developers
The researchers’ findings will reveal how the bias of people selecting the data on which the algorithm was trained affects its results, AI expert Ariel Herbert-Voss saidwho participated as a jury member in awarding the award. She hopes that “more companies developing algorithm-based products will see the value of product feedback this way and build on that approach,” she wrote on twitter.
Bogdan Kulynych himself writes in his publication: algorithmic bias “can lead to the exclusion of minorities and the maintenance of conservative aesthetic standards in thousands of images.”
He praised Twitter’s unusual competition as something worth emulating. If such competitions are held regularly, with experts looking for security holes, “the software will no longer be around for years unless there is evidence of the damage it caused.”
Reader. Organizer. General creator. Zombie fanatic. Alcohol advocate. Food junkie. Bacon ninja.