Facial recognition CEO: software is not ready for use by law enforcement

As in many things, a blanket statement of 'not ready for use' isn't really the most helpful way to look at the situation. There's a range of ways such technology could be usefully deployed, even with current technological problems, that might aid law enforcement officers in the field. It's a very wide continuum from zero to "automatically locking people away for life based on a facial match".

As the article says, there are huge concerns about how such tools might be misused, whether legally or not. So the title might more accurately stated not as 'facial recognition software is not ready for law enforcement', but that 'law enforcement is not ready for facial recognition software'. I'm fairly sure there are countries out there where citizens do not lie awake at night in fear of their government, and where such technologies might be responsibly deployed as a tool in law enforcement.

I dunno. We already had the Orwellian scare when security cameras started getting ubiquitous. We keep kicking that fear can down the road, but we don't have to. We've already seen what security cameras are really good for, watching the watchers.

It doesn't make sense to surveil regular people, but it makes a heck of a lot of sense to surveil public officials while they're doing their jobs. We went from "the police are going to sit back and watch cameras of everybody and show up at our doors to arrest us!" to "Damn cop department won't release his body cam footage because they claim it was out of battery power!"

I think facial recognition is going to follow that same sort of acceptance arc. Nobody except China is going to have the stones to build a bureaucracy around deploying it and we're going to want it to solve a way more pernicious problem once we find a killer app.

If you're really worried about government malfeasance, all you have to do is to actually read some of the Wikileaks dumps and go looking for smoking guns. Everybody else already did that and other than the NSA programs, nobody's found anything really interesting.

We all know the systems are biased, because we feed them biased information - Cathy ONeil did a great job on this with "Weapons of Math Destruction" What puzzles me is why he doesn't just provide the counterexample to scare the hell out of everyone - if he inverts the racial distributions in the training datasets, then you'd have a system where the shoe was on the other foot, and both argument and fuel for removing bias from the training sets - the recognition systems only know what they know, and they don't really understand that they might not know, they only know the confidence they have in knowing - hence this is my favorite happy place example of the problem --https://youtu.be/UFVB5rnqjyY

I agree in general with the spirit of the article, at least as I understood it: facial recognition can empower oppression and lead to even more doubtful convictions, for example it could erroneously supply additional rationale to suspect or convict a person, reinforce biases, etc.

However, I do not get the "companies unite against selling X to government" approach. It is, to me, both misguided (restricting gov't from buying something private companies can buy will simply lead to relabeling or minor redesigns of the same technology) and naive (to work, it needs a very broad agreement which IMO will not get enough traction).

An approach that could work better is to inform, not restrict: make sure that all imagery and videos used or considered for use police are public unless there is a short term, limited exception. And make police wear cameras whenever they wear uniforms and make those video streams public, too (maybe with a few hour lag in case there is a tactical need). My 2c.

>We need movement from the top of every single company in this space to put a stop to these kinds of sales.

What are some historical examples in which industry coalitions of unprovoked top private executives refused government contracts in the name of the civil rights of their fellow citizens?

Would ir LEDs built into sunglasses be effective in combatting this trend. There was an article on hackaday about a person that did that to his license plate holder to deal with red light cameras.

It seems obvious that, regardless of the legality, it will be used in parallel construction. Once these tools are made, they will be used.

https://en.wikipedia.org/wiki/Parallel_construction

Yet Axon (formally Taser International) is building systems to recognizes faces for law enforcement body cams. [1]

Especially disturbing is that these systems seem to have a much higher rate of misidentifying minorities.

[1] https://www.npr.org/2018/05/12/610632088/what-artificial-int...

>Facial recognition technologies, used in the identification of suspects, negatively affects people of color. To deny this fact would be a lie.

Since, instead of support for this statement, the CEO of this 'company' decided that shaming his reader would be more effective, I'll give you his support. It's an article that he wrote based on studies that show there are significantly higher error rates in gender and race classification in some algorithms, without ever showing why this even matters within the context of how facial recognition algorithms are or would be used by law enforcement. Nor did he show that the specific algorithms being proposed have these biases, and not just some other ones. Nor did he show that race classification is even a result that law enforcement runs.

[1] https://www.kairos.com/blog/face-off-confronting-bias-in-fac...

On a side note - I'm unable to square this fear of watchers with parents happily singing to their kids - "he knows when you've been sleeping .. he knows when you're awake .. he knows when you've been bad or good, so be good for goodness sake"

Their FR is not ready for Law Enforcement, because it is weak, poorly designed, and overly expensive. They are smucks trying to capitalize on Amazon's name with generic, low quality FR. Anyone that understands FR knows to go to the NIST web site and use the companies competing to be the best for government contracts. Their algorithm stats are all tested and the results published at NIST.

Does Kairos actually produce any useful software or does it just exist so Brackeen gets publicity?

decidethefuture.org

Why are all of these articles written under the premise that this software sends you automatically to jail? It’s just a first step in the vetting process.

lol tell that to china

"We lost the contract to Amazon"

This is a data retention/sharing issue, not a technology issue. Nobody cares if their face, license plate or cell phone is seen and mapped to their identity as long as the interaction and its metadata is soon forgotten. It's the storage of this information that causes the problem.

When police can accurately determine who they're looking for based on their appearance and can quickly determine whether the person they're bothering is that person they are more limited in how they can harass people in the present. If no records are kept big brother is more limited in his capacity to harass people in the future.

Most people don't have any warrants out, haven't recently been recorded on a security camera robbing a liquor store, etc. Being able to ID people reliably without stopping them is the last thing law enforcement wants (let's ignore the three letter agencies for a minute here) because it makes it much harder to stop people for being "suspicious" or "looking like a drug dealer" (license plates provide similar protection for cars, they can't just stop every silver Camry because a silver Camry was once stolen).

License plates and cell phones almost map 1:1 to your identity. Nobody has a problem with license plates or phones as long as they're not used to stalk people at scale. It's the creation of the data-set that can be used to stalk people. We need to stop creating these data sets in order to prevent the stalking at scale (which is the real issue here, even the article says that).

Imagine a world where the speed limits were followed like normal laws (i.e. they were reasonable enough that most people didn't have to go out of their way to follow them). The police in that world would hate radar guns because they're accurate (and can be audited for accuracy) Big brother would hate them because they don't record metadata for each reading. Facial recognition, ALPRs, need to work like that. History has proven time and again that you can't keep technology in the closet. Better it be used on our terms than theirs.

We're going to have to confront the data retention and government/commercial stalking/surveillance issue eventually. Getting angry over APLRs, stingrays or facial recognition is just playing whack-a-mole. I'm still gonna play whack-a-mole until we solve the big problem though.

Edit: If people disagree I'd love to hear why. This isn't Reddit.

I followed the linked article sited in this post - "Face Off: Confronting Bias in Face Recognition AI" which states:

"Fortunately, the matter of algorithmic ethnic bias, or “the coded gaze” as Buolamwini calls it, can be corrected with a lot of cooperative effort and patience while the AI learns. If you think of machine learning in terms of teaching a child, then consider that you cannot reasonably expect a child to recognize something or someone it has never or seldom seen. Similarly, in the case of algorithmic ethnic bias, the system can only be as diverse in its recognition of ethnicities as the catalogue of photos on which it has been trained."

As this is extremely concerning. I had some questions:

Are the training sets for company's selling facial recognition technology considered proprietary and therefore not verifiable as free from bias?

Is it not possible to develop a standard or criteria that a company's training set has sufficient distribution to be free of racial bias? Is this just not possible for some technical reason?

There is not a single statement in that article that is not some type of fear implied concept that fails under observation. The "FR is biased towards people of color" issue was publicized a few years ago, and the industry adjusted - most are no longer biased, and that bias was a mathematical color space issue and nothing racial. It's an empty article cut and pasting fear statements from popular media, and simply Kiaros getting their name in the news.

Technically speaking, face recognition has been a solved problem since the early 2000s despite facing few challenges nowadays including non-frontal, tiny or partial real-time detection. The most common algorithm is to output the bounding box (i.e. rectangle) coordinates for each present face, extract the facial landmarks (nose, pupils, mouth, etc.) points for the target face (the more landmarks you have, the more accurate the result is), perform some Euclidean distance calculation (sort of hash) and compare this to the existing face hashes (i.e. police database). The highest score is considered a potential match.

You can develop your own face recognition software providing a good dataset (i.e CelebA, LFW or the fed dataset if you have access to) using our embedded computer vision library[1] that we just released earlier this month or using the PixLab /facecompare HTTP endpoint (https://pixlab.io/cmd?id=facecompare).

[1]: https://github.com/symisc/sod