Maybe there is hope for 2020: AI that 'predicts criminality' from faces with '80% accuracy, no bias' gets in the sea
- Reference: 1592982365
- News link: https://www.theregister.co.uk/2020/06/24/face_criminal_ai/
- Source link:
The [1]missive , backed this week by 1,168 researchers, students, and engineers, and addressed to the academic publisher's editorial committee, listed numerous studies that rubbished the suggestion criminality can be predicted by algorithms from something as trivial as your face.
“Such claims are based on unsound scientific premises, research, and methods, which numerous studies spanning our respective disciplines have debunked over the years,” the letter read. “Nevertheless, these discredited claims continue to resurface, often under the veneer of new and purportedly neutral statistical methods such as machine learning, the primary method of the publication in question.”
The neural net in question – developed by a PhD student and two assistant professors at Harrisburg University in Pennsylvania – [2]was said to be “80 percent [accurate] and with no racial bias."
The experts in the letter argued, however, such a feat is impossible. “Let’s be clear: there is no way to develop a system that can predict or identify 'criminality' that is not racially biased — because the category of 'criminality' itself is racially biased,” the eggheads sniffed. By that, they mean, as discussed in footnote 12 of their letter, the color of your skin influences whether or not you are arrested and labeled a criminal. For example, a Black person in a poor neighborhood could get a criminal record for drinking in public where that is forbidden, while a White person in a park in an affluent neighborhood gets away with a verbal warning.
Now, Springer Nature has decided to throw out the paper, titled A Deep Neural Network Model to Predict Criminality Using Image Processing, and will not run it, as it intended, in its research book series, Transactions on Computational Science and Computational Intelligence.
The infamous AI gaydar study was repeated – and, no, code can't tell if you're straight or not just from your face [3]MORE LIKE THIS
“The paper was submitted to a forthcoming conference for which Springer had planned to publish the proceedings in the book series Transactions on Computational Science and Computational Intelligence,” a spokesperson for the publisher told The Register on Tuesday. "After a thorough peer review process the paper was rejected and therefore will not be published by us."
Harrisburg University also removed from its website a press release heralding the study, and said its authors were working to update the paper to address ethical issues. “Academic freedom is a universally acknowledged principle that has contributed to many of the world’s most profound discoveries,” the American college [4]said in a statement last month.
“This university supports the right and responsibility of university faculty to conduct research and engage in intellectual discourse, including those ideas that can be viewed from different ethical perspectives. All research conducted at the University does not necessarily reflect the views and goals of this University.”
The controversial study was written by PhD student and New York City police veteran Jonathan Korn, and Nathaniel Ashby, an assistant professor of cognitive analytics, and Roozbeh Sadeghian, an assistant professor of data analytics. Sadeghian told El Reg he has since retracted his name from the latest version of the paper due to “some disagreements among the authors about the applications.”
Korn and Ashby did not immediately respond to The Register’s request for comments. ®
Sponsored: [5]Google Security Whitepaper
[1] https://medium.com/@CoalitionForCriticalTechnology/abolish-the-techtoprisonpipeline-9b5b14366b16
[2] https://web.archive.org/web/20200506013352/https://harrisburgu.edu/hu-facial-recognition-software-identifies-potential-criminals/
[3] https://www.theregister.com/2019/03/05/ai_gaydar/
[4] https://harrisburgu.edu/hu-facial-recognition-software-identifies-potential-criminals/
[5] https://go.theregister.com/tl/1956/-8471/google-security-whitepaper?td=wptl1956
Re: Training data?
Could it be the same as the AI that could differentiate between dogs and wolves? Turned out most of the wolf pictures had snowy backgrounds.
In this case, the mugshots have a little board with the person's name. Innocent folk are all in beachwear!
Re: Training data?
Yes, probably their shell suit collars were visible in the pics.
Reminds me of when I was a student. One day the cops came in to our mess hall and were going round apparently randomly talking to people, then they came over to me. Turns out they were doing an ID parade that evening, and wanted people who were roughly a certain age, height and colouring to take part alongside the suspect.
When we turned up, the guy's solicitor, with the suspect were allowed to pick which of us they wanted in the parade. The suspect had turned up in a nice suit, obviously trying to look as uncriminal as possible, then gets put in a parade against a bunch of scruffy students. When asked if he had any objections, he did say that he stood out because we all "look like a bunch of fucking students".
I remember getting 20 quid for couple of hours doing nothing except sitting round and standing in a line briefly, and the police station had a bar so we got a couple of rounds in between performances and had a great time.
The three different witnesses all picked the suspect out. He was definitely guilty, because when he was standing next to me and they were about to bring in the last one he whispered "this fucker will get me for sure because he's an ex-cop".
Re: Training data?
How to get images of guaranteed non criminals?
Easy, just use baby photos. What could go wrong?
Re: Training data?
...they are always stealing hearts.
Fizzog
Physiognomy has been a laughable pseudoscience since Victorian times and I doubt anything has changed that will make it anything more than that.
Perhaps it is telling that one of the originators of this study is an ex cop.
Brings to mind Constable Savage.
Re: Fizzog
Also sounds like Phrenology!
Re: Fizzog
It's easy for you to say that... You have the brainpan of a stagecoach tilter...
Re: Fizzog
I have a feeling there is some kind of reverse process that can go on that causes criminality where people who are subject to discriminated due to their looks find it necessary to commit criminal acts to live. Institutional bullying that this will not help with at all.
Re: Fizzog
As a test, I think photos of all the senators, congressmen and European politicians should be run through this system.
Preferably the pics should be taken first thing in the morning, before breakfast or after a long session.
I wonder what kind of theory they put in the paper to explain how faces affect criminality (or vice versa). And why the first reviewers did not ask for one.
Of course
The fact it was written by a former cop is totally unsurprising at this point. Reviving a pseudoscience that's been discredited for nearly two centuries as an excuse to beat people up is incredibly on-brand for them right now.
AC for reasons that should be obvious.
Re: Of course
How dare you accuse the police of using pseudo-science? Clearly we have to use our lie-detector to find out who posted this.
There are perfectly reasons to write a paper on facial recognition of criminals but what it should be about is researching the confounding factors that have given an AI the ability to pick out pictures of criminals. There was a Chinese AI project on this which was shown to have developed the ability to recognise official police photographs of criminals because they included them in the training set. That would be good research as it would support understanding of the impact of the photographer on AI analysis of any pictures.
Recognizing official photos
This was exactly what I was speculating about in my earlier post, not knowing about the AI project you mention - do you have a link?
I'm sure that the US AI trainers must have done things like eliminate the height marker along the wall behind the face image. Didn't they...?
Re: must have eliminated the height marker
I would not bet on it. A facial gaydar neural net was tried with photos with and without the the faces blanked out. It gave the same answers both times so whatever it was making decisions on (clothes? background?) it wasn't the faces. An interesting test of this [1]Mr Cudoogo detector would be to try it with and without blanked out faces.
[1] http://www.wepsite.de/constable_savage.htm
"Copper still thinks he can spot a criminal from how they look" - and other non-news of the day
Body language
"Copper still thinks he can spot a criminal from how they look" - and other non-news of the day"
Actually, with enough experience you can get pretty good at spotting someone up to no good.
Re: Body language
black person = drug dealer.
that type of experience?
Re: Body language
They wear uniforms and drive around arresting minorities on spec rather than do some detective work.
Re: Body language
I think it sort of depends.
If your wearing a police uniform and your out and about and you see somebody that sees you, then snaps their head around to their friend and says something, and the pair then quickly stuff things in their pockets, turn around and start walking in different directions I think it's probably going to ring alarm bells for the cop because it's going to look as if they've seen you and are then trying to split up to avoid you, which probably has a reason. Some of those reasons may be innocuous, but may indicate that they are doing something dodgy.
But that's the behavior and body language that's kicked off alarm bells, and not actually the appearance of the person, so shown a static image how could the AI tell that there was an issue?
Re: Body language
Spotting someone who is actively up to no good is one thing. Looking at a picture of someone's face devoid of any context and deciding whether they're likely to be a criminal is something else entirely.
its easy
They just look for pictures with the black eye mask and the bag over their shoulder that says SWAG in large capital letters.
Even I can do that without a computer.
Re: its easy
It's harder than it looks, apparenty: [1]https://www.youtube.com/watch?v=waWugm79mQo .
[1] https://www.youtube.com/watch?v=waWugm79mQo
Very dodgy subject
but on a personal note, I find some people do look a bit dodgy from their features (nothing racial here, I'm white and I'm referring to white people). I don't know if it is exactly their facial features or their expressions or use of eye contact or whatever, but I'm automatically wary about certain people - the wife also tends to be wary of exactly the same people. There was a gang of 5 violent burglars on a BBC one programme yesterday and they all had "the look".
When certain people knock on your door, before they open their mouth, you automatically suspect they are up to no good; just something about them. You are ready for their scam before they claim to be from the gas board or that you have a slate missing on your roof.
Certain TV and movie actors also have "the look" and tend to be picked for parts where they play the baddy or psycho. Some have made a long career out of it. As to what they are like in real life, who knows. This is the problem with generalisations. I'd guess that there is maybe a bias of people who have the dodgy look who actually are criminals compared to the rest of the population but probably not enough to go down the rabbit hole of trying to make a science out of it. There are of course plenty of people who don't have "the look" who are dodgy.
5 violent burglars on a BBC one programme yesterday and they all had "the [email protected]
I share the view that there are people you meet first time, passing, etc, and your internal bell just rings for no apparent reason (sometimes quiet, but persistent, background noise to be weary). On one hand, I'd love to see research into this and intuition as well, on the other hand, re. research results, I hear this quiet, but persistent, background noise that tells me the outcome won't be used for anything good.
As to mugshots shown on media though, I'd be VERY careful. Whoever is paraded for a mugshot opportunity, they won't be in a good mood, and our own bias tells us that they must be bad, bad people. And even if they grin stupidly (because one moment they were having a good night's sleep, the next they found themselves in a police cell and what the f... just happened?!) to us they're not just bad, they're evil monster with an evil grin, etc. Media illustrate text with a shot that is put there to reinforce the message, which is manipulation, but there's hardly any news that isn't.
Re: 5 violent burglars on a BBC one programme yesterday and they all had "the [email protected]
I share the view that there are people you meet first time, passing, etc, and your internal bell just rings for no apparent reason
Confirmation bias means you don't remember the times you got it wrong.
Some decades back I worked with a guy who everybody thought was gay - man bag, thin leather tie, seriously fashionable clothes and he could accessorize(*). Nope, wife, three kids, straight as a very straight person. Looks can be very misleading.
(*) I'm trying to remember which film that joke came from.
This.
When I lived in a big city the difference between people that shopped in the posh shopping areas compared to the poor was obvious.
The AI may be differentiating between poor and rich people (hairstyle, clothing...).
Other factors may be that the AI picks out the pale pallor and bony face of drug addicts.
In the UK, differences between Carribbeans and Asians are pronounced thanks to the 1970s expulsion of Asians from Uganda, Kenya and Tanzania (they were largely self employed entrepreneurs then sent their offspring into professions because self employment sucks) while many people from the Caribbean came over to fill employed (rather than self employed) menial positions.
Then we've got data sets. Is the AI working out the differences between mugshot photos and, say, passport style photos, mood (generally neutral in passport photos, for some reason generally less than neutral in mugshots...)
The whole area seems fascinating. A complete waste of time for distinguishing criminals but lots of lessons about history, bias and the importance of good data sets.
Re: Very dodgy subject
"Certain TV and movie actors also have "the look" and tend to be picked for parts where they play the baddy or psycho."
You probably think people with that "look" are criminals because that's the way they are portrayed in movies.
Re: Very dodgy subject
"You probably think people with that "look" are criminals because that's the way they are portrayed in movies."
A friend of mine is an actor, he gets to play a lot of roles as a "villain", usually on TV, occasionally in films. He looks like a complete thug onscreen, but that seems to be a "look" he can create.
Re: Very dodgy subject
Whenever I see a nasty villain on TV I always hope it is good acting. That's what actors are paid for. There again, they often end up at odds with the tax authorities.
Re: Very dodgy subject
Anybody with a slight bit of experience can do this. I do this in reenactments. I can and do convincingly play a very mild Sargent Wilson type, or a pantomime "hanging's too good for them" type.
Re: Very dodgy subject
I'd say you're able to tell when (most people) are forcing an expression, and it doesn't look right - so there goes the alarm. I bet there are some damn good fakers that would pass mustard, but that's the arms race I guess.
Although I've also had the same when in foreign countries and I haven't learned the cues from childhood so I don't think this intuition is anything to be relied on, quite the contrary.
Also, if someone is (or appears to be) either well above or well below you in a social hierarchy, that may trigger a fight or flight response, in my opinion.
Whether those responses are justified is case by case, of course. This research is a joke.
Re: Very dodgy subject
nah its just your prejudice based on social conditioning and fear of things you dont understand
During my "goth" phase as a teen families would cross the street to avoid the scary looking person dressed in black scowling (doing nothing and wanting nothing to do with anyone), usually into crowds of alchopop fuelled tracksuit clad teens who became known as chavs causing trouble in town centre.
Similar thing happened when i shaved my head (fund raiser for charity) i went from long haired hairy biker looking bloke to skinhead thug and the reaction was telling, especially people assuming i was pro brexit and anti migrant (couldn't be further from the truth)
fact is you cant judge a book by its cover
That said if any one knocks on my door and isnt delivering something i ordered then they are cold calling and i automatically dont care and send on way dont even bother to look at the face to know i dont want it lol
Also if you want to recreate the "look" its one part stubbing toe, 2 parts trying to remember pi to 60 decimal places, with a smidge of docile cow vacant thousand mile stare added for additional intensity, its a fine line to walk so you dont look like your severely constipated
Re: Very dodgy subject
"they all had 'The Look'..."
That generalisation might work with violent crime... broken nose, cauliflower ears... erm, sorry, that's a rugby player!
How do you identify a fraudster on your doorstep? Polite, mild-mattered, nice haircut... erm, sorry, that's a Mormon/Jehovah's Witness
Seems like the authors
might benefit from a course of retrophrenology...
Hmm... didn't we use to have a hard hat icon?
New York City police veteran
Left the force. Not racist enough.
So Lombroso was wrong?
Who knew?
Wait for it...
"We found that 99.9% of people walking into the Ritz did not have criminal features, compared to 85% of people walking into a benefits office."
Programmed by people that didn't add the faces of tax evaders, embezzlers, and stock market fraudsters to the database, naturally.
Training data?
It might very well be that the AI got good at distinguishing mugshots taken for police purposes from whatever the authors used for their "non-criminal" set. And there's a thing: how do you know that random face images do *not* represent a criminal? Maybe the AI can distinguish photos of criminals that get caught...