Prominent facial recognition researcher scraped videos of trans people, left dataset exposed
New research finds that a controversial facial recognition dataset of trans people remained available online for years after the initial controversy of its existence, reports AlgorithmWatch. The academic responsible is a U.S. government advisor.
Public scrutiny goes back to 2017 when The Verge investigated the work of Karl Ricanek, professor at the University of North Carolina Wilmington, notes AlgorithmWatch. But his research goes back even further, to 2013.
Ricanek built a biometric dataset of 10,000 images of 38 trans people, scraped from their YouTube videos documenting their hormone therapies. The dataset would improve facial recognition system accuracy in being able to establish that an individual is the same person but after hormone therapy.
In 2017, Ricanek said he had tried to acquire consent from the video posters, but had not been able to reach them all and that he only shared links to the videos, not the images themselves.
However, fresh research by Os Keyes and Jeanie Austin, published in Big Data & Society, finds that not only was the HRT Transgender Dataset still available as a Dropbox URL until April 2021, but that it was not simply video URLs, but contained the videos, many of which had since been deleted from YouTube by the posters and all of which were subject to copyright.
Keyes and Austin contacted the institutional review board at the University of North Carolina at Wilmington about the project, only to discover that Ricanek never sought ethical approval. The dataset was shared with academics who subsequently shared it with their own doctoral students and researchers, again without oversight.
Earlier this week, Ricanek spoke about democratizing face biometric technology for the transgender community, at FedID, the Federal Identity Forum and Expo, as part of the U.S. National Academies of Sciences, Engineering and Medicine knowledge-gathering project on facial recognition and the data behind it.
Ricanek said that democratizing face-based tech for the transgender community will be a challenge for algorithms as well as policymakers.
His data acquisition appears to be informing his advice. He said that even hormone therapy can alter a face significantly as changes in skin density and vascular structure can be enough to fool AI.