Twitter holds 'Contest to discover bias of image cropping algorithm', bounty up to 380,000 yen
Twitter announced on July 30, 2021 that it will hold a 'contest to discover biases in image cropping algorithms'. According to Twitter, this is the industry's first 'algorithm bias bounty contest,' and the winner of the contest will be paid a $ 3,500 bounty.
Introducing Twitter's first algorithmic bias bounty challenge
Twitter announces first algorithmic bias bounty challenge | ZDNet
https://www.zdnet.com/article/twitter-announces-first-algorithmic-bias-bounty-challenge/
In 2018, Twitter introduced a function to optimize the trimming of images posted by users using AI technology and display them as thumbnails according to the screen layout. However, there were voices pointing out that there is a racial bias for algorithms that should cut out mainly human faces.
When Twitter actually verified the cropping of the image by the algorithm, it turned out that the algorithm tends to prioritize the white face over the black by about 4% as a whole. In response to this result, Twitter has announced that 'the algorithm of the automatic trimming function will be phased out.'
Twitter announces 'stop automatic trimming to eliminate racism' --GIGAZINE
And on July 30th, Twitter announced on its official blog that it will hold a 'contest to discover biases hidden in image cropping algorithms'. 'It's hard to find bias in machine learning (ML) models, and companies sometimes learn about unintended ethical harm after it's already open to the public,' Twitter said. We are planning to hold this contest with the hope of changing the situation.
Twitter has already published the code for the problematic image cropping algorithm on GitHub, allowing others to reproduce the problem with the algorithm. 'We would like to invite the community to incentivize and take this task one step further so that we can identify potential algorithmic harms beyond what we have identified,' Twitter commented.
The found bias is evaluated by various items such as the content of the expected harm, the type of community affected, the number of users, and creativity. The amount of the bounty is as follows.
・ 1st place: $ 3,500 (about 380,000 yen)
・ 2nd place: 1000 dollars (about 110,000 yen)
・ Third place: 500 dollars (about 55,000 yen)
· Most innovative: $ 1000
• The most generalizable (that is, the one that can be applied to the most types of algorithms): $ 1000
The entry period for the contest is from 9:01 on July 30, 2021 to 23:59 on August 6, 2021 Pacific Time (PT), and the winners will be the Security Conference in Nevada, USA. It will be announced on August 8th at DEF CON's AI Village Workshop. Twitter says it will invite the winners to the workshop.
Details of the contest can be found on the following page published by HackerOne, a platform where companies pay incentives to hackers who discover vulnerabilities. Rewards from Twitter will also be paid through HackerOne.
Twitter Algorithmic Bias --Bug Bounty Program | HackerOne
https://hackerone.com/twitter-algorithmic-bias
Twitter product manager Jutta Williams also told foreign media ZDNet, 'I hope that people in the broader community will share a lot of views on what is harmful to them. Commented. The contest isn't just about identifying algorithmic biases, it's also about learning what people think is 'harmful.'
The contest was conceived by Rumman Chowdhury, who belongs to Twitter's 'Ethics, Transparency and Accountability (META) Team in Machine Learning' , after discussing with the organizers of AI Village. 'This wasn't done just for our benefit. I wouldn't have sweated personally unless it was for the ultimate in transparency,' Williams said. Said.
Related Posts:
in Mobile, Software, Web Service, Posted by log1h_ik