CEO of Facial Recognition Company Kairos Argues that the Technology's Bias and Capacity For Abuse Make It Too Dangerous For Use By Law Enforcement (techcrunch.com)

Brian Brackeen, chief executive officer of the facial recognition software developer Kairos, writes in an op-ed: Recent news of Amazon's engagement with law enforcement to provide facial recognition surveillance (branded "Rekognition"), along with the almost unbelievable news of China's use of the technology, means that the technology industry needs to address the darker, more offensive side of some of its more spectacular advancements. Facial recognition technologies, used in the identification of suspects, negatively affects people of color. To deny this fact would be a lie. And clearly, facial recognition-powered government surveillance is an extraordinary invasion of the privacy of all citizens -- and a slippery slope to losing control of our identities altogether.

There's really no "nice" way to acknowledge these things. I've been pretty clear about the potential dangers associated with current racial biases in face recognition, and open in my opposition to the use of the technology in law enforcement. [...] To be truly effective, the algorithms powering facial recognition software require a massive amount of information. The more images of people of color it sees, the more likely it is to properly identify them. The problem is, existing software has not been exposed to enough images of people of color to be confidently relied upon to identify them.

Comments

If it can be

[oldgraybeard]

made it will be and those in power will use it to expand and protect their power

What?

[Anonymous Coward]

Facial recognition technologies, used in the identification of suspects, negatively affects people of color.

Surely only if the suspect is a person of color.

Re:What?

[Immerman]

>Not sure I know enough about how that works to understand why less data would mean more false positives.
More training data means it needs to learn to recognize more subtle distinctions to be able to correctly identify an image. Without that subtly it will tend to overlook the differences and misidentify images.

It's actually very similar to the "X all look alike to me" effect. Let's take an extreme example: Imagine you live somewhere where pretty much everyone is white. You've only ever seen a handful of black people in your life, and Fred is the only black guy you personally know. Cool guy - you like him, grab beers after work, etc. And since we identify people by recognizing the differences between them and everyone else, "dark skin", "wide nose", "full lips", etc. are some of the big features you use to identify Fred. And why not? Nobody else you encounter has those features, so they really stand out to identify him from everyone else you see.

Then one day you're walking down the hall and see a black guy coming your way - similar build to Fred, with the same dark skin, wide nose, full lips, etc. And so you identify him as Fred, ask him how his project is going, and if he wants to grab a beer after work. And a totally confused Steve tries to figure out why the hell some complete stranger is acting like an old friend. Then Fred walks up, and seeing them stand side by side you start noticing the differences you didn't see initially - Fred has way more wrinkles around his eyes, Steve's cheeks are considerably rounder, etc. And, with a bit of practice you get good at telling them apart. Then you go to a conference where almost everyone is black - and once again you keep losing track of Fred, because there's a sea of faces around you, all bearing features superficially similar to Fred's, and you've really only learned to identify the small subset of obvious differences between Fred and Steve. You'll get better at it eventually, but in the meantime you just haven't yet recognized enough of the normal range of variance to make a clear distinction even between not-all-that-similar-looking people that share the same obvious features.

Re:Racist much?

[Anonymous Coward]

Facial recognition technologies, used in the identification of suspects, negatively affects people of color

This statement is outright saying that black people are mostly criminals, the only case in which facial recognition identifying suspects "negatively affects people of color".

I think it's time we put an end to subtle racism like this, just because someone is black does not mean they are a criminal.

He's saying that the ML training dataset for people of color is too small. The machine needs to see more black people to identify them properly. This has nothing to do with black crime rate.

Re:Racist much?

[fish_in_the_c]

but does it give more false positives ? It seems like a %accuracy calculation would be most appropriate with any computer identification system, so the humans on the ground can behave differently with a 85% likelihood vs a 99% likelihood vs a 20% are these systems not able to generate that kind of data?

why?

[cascadingstylesheet]

I'm afraid you are going to have to show your work here.

The problem is, existing software has not been exposed to enough images of people of color to be confidently relied upon to identify them.

Are you sure? And if so, why hasn't it?

This isn't the 1960s. Who exactly is biasing facial image databases, in 2018? Noted hotbeds of racism like universities and tech companies? How are they doing so?

Please look up what Bias means

[FeelGood314]

The technology doesn't routinely make judgement calls that are inaccurate in a specific direction. It is however much less accurate but lack of accuracy does not mean bias.

Second, it is the policies around how it is used that negatively affect non-white people. This is a policy problem not a technology problem. I'm really not keen on being tracked and scanned by facial recognition or any of the other ways organizations track me but please don't exaggerate and play the racism card just to get clicks. In the end it numbs us to real abuse.

Re:Please look up what Bias means

[Drethon]

A large number of technological methods have bias ( https://cals.arizona.edu/class... ) and the facial recognition algorithms are usually machine learning I believe, they can indeed have quite a bit of bias built in. This bias can be created by the developers not training the system with properly balanced data, which is a technological issue. That bias can be due to actual bias in the world (as you mention) so here the model is right, it is just reflecting real world bias. Understanding the cause is very important.

Machine learning bias: https://towardsdatascience.com...

Re:Devil's advocate: This technology will save liv

[arth1]

Scenario 4: There are a few hundred thousand people who will trigger the detection routines over and over again, because that's just how they look. So they get apprehended and arrested over and over again, and are unable to lead normal lives.

No, thanks, we do not need this.

The Police Have Always Used Facial Recognition

[PastTense]

The police have always used facial recognition--both the police recognizing criminals from previous knowledge and mugshots and witnesses recognizing criminals--from the criminals who attacked them to photos they see on TV.

This recognition has always had inaccuracy problems--and a lot of people have wrongly suffered. The (partial) solution has always been to use it in conjunction with other evidence.

So there is no basic difference from facial recognition from software vs facial recognition by people--and with time the software recognition will be much better.

I'm just waiting.....

[bev_tech_rob]

...for the Pre-Cogs to show up in the news and then we are in deep sh*t....