TX man sues Macy's, Sunglass Hut parent over wrongful facial recognition arrest

Tried opening the link on iOS with my app’s built-in Safari browser.

‘’’ Browser Blocked We apologize, but your web browser is configured in such a way that it is preventing this site from implementing required components that protect your privacy and allow you to view and change your privacy settings. This functionality is required for privacy legislation in your region.

We recommend you use a different browser or disable the “EasyList Cookie” filter from your “Content Filtering” settings (found under “Settings” -> “Shields” in the Brave Browser). ‘’’

I do not use Brave…

Hope he gets his $. Absolutely absurd what happened to him.

So many layers of BS here (in the his treatment I mean, not the article).

First off lets deal with law enforcement:

* Why was he in jail for two weeks if he could presumably unambiguously demonstrate he was not even in the state at the time of the crime?

* Why is a single witness with no supporting evidence capable of imprisoning someone with no recourse?

* Why is it possible for someone to be raped or assaulted in a prison with no prison officials going to jail? The police are demonstrably aware of this kind of constitutional violation occurring because they're well known to use it as a threat. Simply using that threat in interrogations honestly should be sufficient to render any testimony inadmissible anyway, and threatening harm to others is a crime.

Now on to the facial recognition BS. No form of facial recognition should be considered sufficient for any kind of testimony unless you can demonstrate that it does not exhibit bias. The problem with facial recognition systems I used to read about is that they seem hell bent on focusing on true positive recognition rates and not (especially in this case) the much more relevant false positive rates. Now maybe in the years since I escape academia it's improved, but I really don't think so, as every article I see seems to repeat the same points:

* High false positive rates

* The rate of false positives is extremely biased, first by race, and then by gender

* Any hit by these programs is considered to be correct (I recall some systems including "percentage match" and such like but people not understanding that "97% match" or similar did not mean "97% likely that this is the same person")

It seems that these systems are just functionally laundering racism again. It does not matter if you're not trying to be racist, because the likelihood is your input data is itself almost certainly racially biased you have to actively audit and verify that your system is not simply recreating that bias. You get very basic things like "how are you verifying accuracy" where people _still_ do things like verify using specific types of shot that ML systems learn to mean "likely X" (where X in this case may be "in our list of mug shots", but other cases include "has cancer"). e.g. your validation for crime focus facial recognition needs to have a large corpus of mugshot style (structure, lighting, focus, etc) photos of people who are not criminal (of every race, gender, religion, hair style, tattoos, etc) and a similarly well lit "normal" photos of a similar and/or overlapping group, and your tests for match accuracy need to include a wide variety of shot types including the grainiest and most terribly flickery fluorescent security camera through to fancy fish eye shots where again a variety of different body types, sizes, genders, races, etc are expected to be matched vs not matched.

That's all expensive and so it seems everyone is still just downloading mugshots and security footage, and ML algorithms continue to just go "this is grainy footage and the training rewards returning hits when I am presented with grainy footage". This is no different from the cases where cancer detecting bots learned to recognize fluorescent lighting and rulers in photos as the indicators of cancer being present.