Ask HN: In what ways can the FBI abuse ML models trained on its DB of faces?

[badrabbit]

This week the GAO told the US House oversight conmittee [1] that the FBI has access to a database of 640 million faces known as the "Interstate Photo System of mugshots"

What I didn't hear in this discussion with congress was how ML can be abused when trained against this database of faces and associated information such as what you would find in a drivers license,passport and the like.

I personally only know about ML at a high level and wanted to ask HN what sort of realistic fears and concerns citizens and lawmakers alike should harbor in addition to simple cross reference and realtime recognition of faces?

[1] https://news.yahoo.com/watchdog-says-fbi-access-640m-photographs-173926319--politics.html

[pizza]

If a widespread facial recognition system were deployed, the false positive rate would probably be a lot greater than would be acceptable. For what it's worth, this paper I found claims that humans are 97.53% accurate, whereas one particular deep learning algorithm was 99.80% accurate [0]. If you ask me, 99.8% accuracy means large cities will pretty regularly experience situations where there will be serious negative consequences from misidentification.

[0] https://arxiv.org/pdf/1804.06655.pdf