Twitter will hold a contest to see if hackers and researchers can spot biases in the algorithm's image cropping. The winner teams will receive cash prizes (via Engadget). Twitter hopes that teams will be able to access its code and image-cropping model, which will allow them to identify ways the algorithm could cause harm (e.g. it cropping images in a way that stereotypes or erases subjects).Competitors will need to submit a description and a dataset to be used in the algorithm to show the issue. Twitter will assign points based upon the type of harms found and how it could possibly affect people.The $3,500 prize will go to the winning team, while $1,000 prizes are awarded to the most creative and generalizable discoveries. Some users suggested that the amount should be increased by one. To give you an idea of the context, Twitter's bug bounty program would pay $2,940 to anyone who found a bug that allowed you to perform actions for another person (like retweeting tweets or images) via cross-site scripting. $7,700 would be earned if you find an OAuth problem that allows you to take control of someone's Twitter account.Twitter can get feedback from more perspectives through a competitionTwitter did its own research on its image-cropping algorithm. In May, it published a paper that examined how the algorithm was biased. This followed accusations that its previews crops were racially motivated. Twitter has mostly eliminated algorithmically cropping previews, but it is still useful for companies like Twitter to have a good cropping algorithm.Twitter can get feedback from a wider range of perspectives by opening up a contest. The Twitter team discussed the competition in a discussion room. One member of the team mentioned that he was getting questions about the algorithm's caste-based biases. This may not have been obvious to California software developers.Twitter is also exploring ways to exploit its algorithmTwitter is also looking for subconscious algorithmic biases. This rubric includes point values for both unintentional and intentional harms. Twitter defines unintentional damage as crops that result from a user posting a regular photo on the platform. Intentional harms are cropping behavior that could be exploited to harvest maliciously-designed images.Twitter announced in its blog that the competition is separate to its bug bounty program. If you submit a report on algorithmic biases for Twitter outside of the competition the company will close it and mark it as not applicable. To learn more about the HackerOne competition, visit the HackerOne competition page. Submissions can be submitted until August 6th, 11:59 PM PT. The winners of the challenge are announced at the Def Con AI Village August 9th.
