Aug 10, 2021 12:00:00

Researchers who proved that Twitter's image cropping algorithm 'whiter / thinner / younger / more feminine has higher priority' wins a bounty

At Twitter's Contest to Discover Bias in Image Crop Algorithms, Swiss researchers prove that 'whiter / thinner / younger / more feminine priorities' Won the championship. The contest also proved that 'people with gray hair / wheelchairs /

people who hide their heads like Islamic hijabs have lower priority,' said NBC News, a major American news agency. It is also discriminatory, discriminatory against people with disabilities, and disgusting with Islam. '

Twitter's racist algorithm is also ageist, ableist and Islamaphobic, researchers find
https://www.nbcnews.com/tech/tech-news/twitters-racist-algorithm-also-ageist-ableist-islamaphobic-researchers-rcna1632

Twitter just led the first-ever 'bug bounty' for AI bias
https://www.morningbrew.com/emerging-tech/stories/2021/08/09/twitter-just-led-firstever-bug-bounty-ai-bias

In 2018, Twitter implemented a system that uses AI technology to optimize automatic cropping of images. It was mentioned that this system uses AI technology to achieve faster and more optimal trimming than before, but it has been pointed out that there is a 'bias due to racial discrimination' on the human face, which is judged to be the center. In fact, Twitter verified that there was a tendency to prioritize white faces over blacks to a degree of about 4% overall. Thus, in May 2021, Twitter made the decision to 'stop auto-trimming.'

Twitter announces 'stop automatic trimming to eliminate racism' --GIGAZINE

To fix this algorithm, Twitter held a 'contest to discover biases in the image cropping algorithm'. Twitter published the code of the image algorithm in question on Github and conducted a public offering about what kind of bias exists. We ranked by the content of the expected harm, the type of community affected, the number of users, creativity, etc., and promised a reward of $ 3,500 (about 380,000 yen) for the first place.

Twitter holds 'Contest to discover bias of image cropping algorithm', bounty up to 380,000 yen --GIGAZINE

And Twitter announced the winner of the above contest at the hacker festival DEF CON held from August 5th to 9th, 2021. Bogdan Kulynych, a graduate student at the Swiss Federal Institute of Technology Lausanne, won first place.

Kulynych's research can be viewed below.

GitHub --bogdan-kulynych / saliency_bias
https://github.com/bogdan-kulynych/saliency_bias

Kulynych used AI to make various edits to the subject and scrutinized what Twitter's algorithm would make. As a result, the score of salience (PDF file) referred to when deciding how to crop the image was higher for whiter / thin / younger / more feminine skin. It was confirmed in 37% of the total that whitening the skin color improves the saliency score ...

It was confirmed in 18% of the total that the score increases as you lose weight.

It was also 18%, and the younger the score, the higher the score.

The more feminine the score, the higher the score, which was 25% of the total.

Also, Professor Parham Aarabi, who studies artificial intelligence bias at the University of Toronto, which won second place, proves that 'people with gray hair / wheelchairs / hiding their heads have lower priority'. did.

Regarding this result, contest management Rumman Chowdhury said, 'It is very difficult for a single team to identify all the wrong algorithms that are open to the public, and honestly it will not be feasible. We want to set a precedent on Twitter and the industry to positively and collectively identify the negative effects of algorithms. '