J'accuse! Amazon's Rekognition reckons 1 in 5 Californian lawmakers are crims in ACLU test Amazon's Rekognition system wrongly matched one in five Californian politicians with images from a database of 25,000 wanted criminals' mugshots in tests by the American Civil Liberties Union (ACLU). If this story sounds familiar, it's because the ACLU ran a similar test last year – which matched 28 members of US Congress with …
I may be unpopular on this one, but I'm all for police wearing bodycams.
The police aren't everywhere. Everyone knows this, especially criminals, so having them wear bodycams doesn't mean blanket surveillance. In fact, it's only blanket surveillance of the police themselves.
In the event that a controversy arises regarding a police encounter, the camera footage can be reviewed (or used as evidence for either party), ensuring that actions taken by the police were fair and reasonable (and stupid stuff like unarmed civilians being shot because they approached a police car can be appraised).
It's a high-pressure job, and sometimes they're expected to make snap decisions with serious consequences. If their overview board can see the actual circumstances in which these decisions were made then it could help to sort out good cops from bad cops.
I wouldn't suggest making all bodycam footage public - that's going too far, as I feel that's invading the police officers' privacy, offering random members of the public the opportunity to covertly observe locations, and opening graphic footage of violence and deaths to voyeurs, but it should be available under a court order.
Also, and I'm going to stick my neck out even further here, if the footage were cryptographically signed and then stored in a digital vault, a checksum could be appended to some form of blockchain (yeah - I feel dirty at that bit too) which would allow public oversight to ensure that recordings weren't tampered with without leaving a trail.
So that's my 10p on that one. If you feel that I'm being a lunatic, I'd love to hear a counterpoint.
Yes, I agree that the video shouldn't be subject to mining. It should be stored intact, with checksums. That's all.
If an incident occurred with the officer or their partner, or in a location that they've visited within a time period (to be determined, but hours to days, not years), that video can be reviewed by a person.
It does need to be regulated, yes. Maybe I'm a hopeless optimist on that one!
It does need to be regulated, yes
Having the video (with an integrity mechanism) go to a third party rather than the police themselves might help, in theory. But it's hard to see how you'd prevent that third party from being coopted by one or more of several interested parties (the police, intelligence agencies, criminal organizations, gossip mongers, ...).
In the event that a controversy arises regarding a police encounter, the camera footage can be reviewed (or used as evidence for either party), ensuring that actions taken by the police were fair and reasonable (and stupid stuff like unarmed civilians being shot because they approached a police car can be appraised).
A few years back I asked one of our local (UK) police about her bodycam and she said the sound was so lousy it would record her but not anybody else unless they were seriously shouting, so it was useless for evidence, but the mere presence of it stopped a lot of people kicking off because they thought it could be used against them if they did anything stupid.
...her bodycam and she said the sound was so lousy it would record her but not anybody else unless they were seriously shouting...
Maybe the camera is what caused everybody else to not shout or speak in lower voice, thus calm the situation, deescalate problems and working as intended.
In addition, even without the voice the camera goal is to provide evidence on people's physical action at the scene, like whether or not the police or citizen acted violently.
"In addition, even without the voice the camera goal is to provide evidence on people's physical action at the scene, like whether or not the police or citizen acted violently."
The footage is only useful to a point. Being attached to the front of the officer, it's not always in a good position to capture what's going on. I've seen footage from an observer further back that records officers telling somebody to not resist and the person is shouting they aren't resisting when it's more than obvious they are. There are also clips where the person is being shouted at to stop resisting when they are al dente limp on the ground. The body cams don't show anything but cloth. The bad gals (being equal here) know that it's easier to capture audio so when in doubt, loudly proclaim you are doing what you are told and yell and scream you are being tortured when a cop is within arms reach to get it on the record.
Being attached to the front of the officer, it's not always in a good position to capture what's going on
And sometimes at a critical moment, leading to a frustrating and controversial gap in the record. That's what happened in the Deven Guilford case, where bodycam footage showed him arguing with the officer, getting hit with the officer's stun gun, and charging the officer; but not the subsequent fatal shooting.
In that case the officer involved was not charged by the investigating DA, but I believe the wrongful-death suit is still pending. If that goes to trial it will be up to a jury, and the bodycam footage can reasonably be interpreted either way.
The problem is that current bodycam programs are set up is that the bodycam footage is held by the police themselves and no-one else. This leads to situations where the police are accused of malfeasance, the evidence just straight up disappears. This is kind of the same problem with the concept of oversight by an "Internal Affairs Bureau" in that its asking an organization to oversee itself, which causes massive levels of conflict-of-interest.
This country really needs some kind of agency or organization set up for the sole purpose of providing oversight of our law enforcement system (Perhaps something that reports up through to the state / federal legislature since the police are ostensibly a part of the Executive branch and the legislature is supposed to be a counter-weight to executive power).
But, to your point of using the 'blockchain', the blockchain is only useful in verifying a chain of transactions involving an arbitrary block of data (Thus "Block chain"). In this case, a simple cryptographic signature would be sufficient, especially if the data included such metadata as an accurate timestamp, GPS coordinates, and possibly some data from some gyros to determine the camera angle. Include that data in each frame, then run a SHA256 against the frame itself. Then when the file gets closed out, do a hash of the hashes and/or the video data itself. That would be sufficient to prove authenticity of the video itself.
My point to the blockchain aspect is to retain a public ledger of each video, so that if the file is tampered with, everyone can see it. Members of the public could conceivably maintain their own copies of that information without having the underlying video, but if the video that turns up in an investigation or a hearing has a checksum that differs from the blockchain copy that everyone can maintain then everyone knows to cry foul.
No I don't know how that would work, because I'm not really a fan of blockchain. I'm throwing it in as a method of having a persistent record that can't be altered but can be audited by all.
Totally with you on the telemetry though. We were asked to consider evidence systems including body cameras, and comprehensive telemetry was amongst the things we put forward. And checksums upon checksums. Never did the blockchain thing though - that's just something I suggested here for accountability to the public.
I have a mate who was a cop (Aussie, but also previously in the USA) who was in an Aussie equivalent of "Internal Affairs" for a period.
All of his training (think firearms, arrest procedures, physical restraining of crooks etc) and other non "on the job" stuff he was flown to another state to be trained there. Reason being is the cops dont fancy these guys in their own department because they get them in the shit and are also quite insular from the cops doing the job of policing.
"This country really needs some kind of agency or organization set up for the sole purpose of providing oversight of our law enforcement system..."
But that just moves the goalpists. How do you ensure the oversight bureau doesn't become corrupt. Or to turn the old question, "Quis custodiet ipsos custodes?"
"Quis custodiet ipsos custodes?"
Well, that's where I suggested Blockchain. That way the evidence from bodycams becomes auditable by members of the public, without making the footage public. Forgive me because I'm way out of my depth on this one, but it seems ideal to me if every time footage were submitted to the digital vault it created a quick mess of metadata (start frame timestamp, end frame timestamp, camera serial number, officer the camera is assigned to, cryptographic hash of the actual video stream - or two, because they're not exactly big), that all gets digitally signed by the storage system, and that signed metadata is appended to a dedicated blockchain. That means that each officer's actions are recorded and the recordings are securely held out of sight, but evidence that the recording exists is available to everyone. And if a recording is tampered with then the checksums don't match up. If a second recording is submitted for that officer on that date, it's abundantly clear to the public that a duplicate has been submitted (or more importantly, a different video for the day).
Most people won't care. Most people who care won't know how to replay the blockchain. But the Venn diagram will have enough people who care, and who can verify the evidence in the vault, so that final oversight of the recordings can be held by the public themselves.
I'm slightly startled because I may have uncovered a genuine, real-world purpose for blockchains.
We can play 'who watches the watchers?' all day. No system is perfect and without the possibility of corruption, I just want something that is less imperfect.
But my point is having at least something a little less terrible than the current IAB system. My big problem with the IAB system is that it reports up to the Chief of Police and/or the Police Commissioner, both people that are politically motivated to protect the police department from scandal. A lot of IAB offices also tend to be staffed with former cops that either got injured and can no longer work as regular cops, or washed out of the police academy, or at least that is how it is around here. My other big problem with the system is that in many cities, it is the IAB that has full authority over whether to file charges for police misconduct or not. There have been multiple cases of crimes committed by cops against other cops that IAB, under pressure from the police chief, the IAB never files charges. In other cases, there have been reports of officers mishandling evidence, but the IAB investigation is ended under pressure of the District Attorney since an accusation of mishandling evidence would be grounds to re-open every case that cop was involved, which can seriously damage a prosecutor's win record f people are found to have been wrongly convicted.
What I am proposing is an agency that has no law-enforcement powers over the people, only the law enforcement agencies and their employees. Essentially, they'd be the answer to 'Who watches the watchers', as to who watches them, I would say have the organization controlled by a board of commissioners that is half made up of State/Federal legislators and half elected by the people directly.
In the event that a controversy arises regarding a police encounter, the camera footage can be reviewed
Unless, of course, it was inadvertently left switched off, accidentally caught in a slamming car door several times, fell down the stairs on the way to the evidence room, or otherwise 'unavailable'
Counter-counterpoint: So you believe that police shooting civilians should never be evaluated or assessed after the fact? It's all well and good investigating the scene afterwards, but given the number of high-profile incidents where people have been harmed by police in the USA, it would be good to be able to establish that the police officer had good reason to believe that they or others around them were subject to a credible threat at the moment they fired a gun.
It's easy to condemn them afterwards by saying "but that person was unarmed", but if they had video evidence to show that the suspect snatched for something inside their coat then the police officer could easily be justified in defending themselves even if it only turned out to be a bus pass.
The best way to learn from a mistake is to see how the mistake was made.
Bzzzzzt! Wrong. They are all unarmed, or armed and friendly, until determined to be otherwise.
If one can justify shooting someone by saying, "I thought he was dangerous, so I'm justified in killing him," with no more evidence than an unsubstantiated belief that someone is armed, then that standard should apply to *all citizens*. Everyone. Without exception.
I.e. if I walk past you on the street I should be able to shoot you, then and there, because of my otherwise unfounded assumption that you are armed, and therefore a danger to me.
Symmetry: It's an important principle of ethics.
"invading the police officers' privacy"? Wait... what?
No employee of any organization has a right to privacy from the organization's owners while on the clock, drawing pay, discharging their responsibilities as an employee.
That includes police officers, or any other public official, while performing their on-the-job responsibilities. The public employ them, and have the right to know what they're up to.
Once they clock out they can enjoy privacy.
I understand your point, but I disagree. These are people, not robots. The public shouldn't be party to their conversations about their families. They shouldn't be involved in their silly / lurid jokes. If our customers could hear some of the conversations we have, they'd be shocked. But these are private conversations, and even if it's the police they should remain private unless the video is required for an investigation.
Yes it is designed for that purpose. If you know you're always being recorded you'll likley act friendlier and not do dodgy stuff.
However, it doesnt work as the cameras just get turned off/damaged (woops!), when the police feel like being dicks/doing something dodgy. Which kind of defeats the purpose...
Agree with the second part, but they need to be built so that they can't be switched off. How long is a shift? Make sure the battery life will exceed that after 5 years, and make sure they switch on the moment they're out of the charging cradle.
If an officer repeatedly damages their camera, that's grounds for disciplinary proceedings (or at least billing them for the cameras). And you'll have footage of them damaging the camera.
No, it's not perfect. But I think it's generally a good idea.
"but they need to be built so that they can't be switched off."
The problem is most of an officers day isn't chasing down people and wrestling them to the ground. If the body cams are always on, there will be hours and hours of each day where they are filling out reports, eating a meal, hiding behind road signs, having naps, etc and all of that footage will still have to be archived. The Government will suck up the world's supply of hard drives to hold all of this. There has to be a way to selectively turn the camera on and off. It might even be a rolling frame that gets overwritten if nothing happens in the same way a dash cam will retain footage if it senses that the car has been in an accident.
One, what's to say something routine, such as passing someone seemingly innocuous, turns out to be critical later (just committed a robbery and was hiding g in plain sight as seen in an Adam-12 episode)? Second, why can't the footage be whittled down later by someone else to reduce bias? Sort of like the filming philosophy "We can always edit it in post"?
It's about as accurate as their AI driven Spelcheka software.
And what's all this recommended 99% accuracy before shooting someone stuff? They probably need to have a bit more of a think about this, before the Amazon Robokop drones sre sent up for an autonomous 8 hour flight, and Sergeant Dweeb suddenly remembers that he hasn't changed the 50% accuracy default setting.
Shooting someone? The late Mr Menezes springs to mind here in Blighty, and don't the 'merkins do that kind of thing rather more often than us?
As for less-drastic stopping someone and giving them a hard time based on false positives coming from fleshbags without any help from technology, it's happened to me at least twice[1] in twenty years since I returned to Blighty. And I'm white, with grey beard.
[1] Two definite and scary occasions, plus a couple more much briefer encounters in which it wasn't clear whether they had misidentified me or just stopped me randomly.
Any technology is going to fail at least sometimes. A guy was stopped at an airport and put in back rooms for hours with all sorts of questions being thrown at him that accomplished nothing. The issue they didn't want to tell him was the bomb sniffer machine identified explosives from a swab. When somebody with brains got around to asking the guy what he did for a living, they let him go. He was a mining engineer coming directly back from working in a South American mine. The machine picked up on the residue left over from blasting. The machine worked fine, but the people that pushed the buttons didn't know how to do their jobs.
99% ? That is just like accuracy of policemen in the US.
99% of the bullets fired miss the target, and only 1% hits.
If the target is hit, he is by definition guilty of something that warranted his shooting. The US should shar forbidding unjust force.
Of the 26 lawmakers identified as possible criminals, more than half were described as "people of color"
How far over half I couldn't find the info? Surely it is about right if it's close to 50, 50 or is this just designed to get peoples backs up? One or the other is is most likely going to pull slightly ahead.
One things for sure though. The tech is no where near ready to be relied upon.
The 120 lawmakers are 80 representatives and 40 state senators. The current legislature has 2 african-americans senators and 8 african-american representatives, 10 total. So in this phrase "Of the 26 lawmakers identified as possible criminals, more than half were described as "people of color" ", the "people of colour" part can't be referring only to african-americans. Expanding the "people of colour" to include latinos gives 30 representatives and 7 senators. So at least 14 out of these 37 were identified as criminals (minimum 38%).
On the other hand, whites and asians make up 81 of the 120, of which at most 12 were identified as criminals (maximum 15%).
It remains to be seen how much of that difference is due to racial bias in the training set / AI, and how much is simply due to the mugshot database being predominantly non-white and having an AI that is equally rubbish at identifying any type of face.
https://www.library.ca.gov/Content/pdf/crb/reports/LegDemographicsNov16.pdf
Hey, I looked at a sample of the output from a Chinese AI, neural, trained to recognize "criminal types". What it actually did was to separate photos of Mark I Rev 0 ethnic Chinese from those of people who looked like foreigners or Chinese non-majority ethnic groups. Classic GIGO.
Tell you what: Suppose there's a one in a million chance a system wrongly matching a given photo of a person with some photograph of a different person in a database. Sounds pretty good, doesn't it? Now take a database of 1,000,000 photos of 1,000,000 different people, and match your given photograph (of a person not in the database) with all 1,000,000 of them. Guess what, the likelihood of NOT being matched to anyone in the database is 36.79% ( (1-0.000001)^1,000,000 ), or, put differently, the likelihood of at least one false match is 63.21%. Of course things get a lot worse if you realise that one-in-a-million chances crop up nine times out of ten.
Facial recognition is pretty good for authentication purposes: matching an image of a person with e.g. his or her passport image. For general surveillance, there are huge risks of false positive matches
Doffs hat to the late, great Terry Pratchett
The problem is that people who understand statistics can be told this, and will see the problem. The overwhelming majority of the public (and legislators) have little understanding of statistics, and are blinded by "even if it catches one criminal that would have otherwise been left to walk free and murder someone, it is a positive for society".
The only way people will understand why it is bad is for it to be deployed, and news stories to begin to come out about the number of people detained due to false matches, along with some high profile people in government falsely detained. Unfortunately at that point the money has already been spent, so the talk will be about how to fix it.
You can reduce the number of false matches to any arbitrary level by requiring a higher confidence of match, but that reduces the number of true matches to where it was never worth the money in the first place. Except to the company that got the contract, and the politicians who got campaign contributions from them or lobbying jobs from them after they left office. And that's all that really matters to the politicians - they aren't spending their money, after all.
As a statistician, I am more that aware of the "false positive paradox" -- when the actual incident rate (in this case "criminality") is low in the general population, the probability that a person identified as as criminal is actually a criminal is low.
Wikipedia offers a good description of the paradox with examples, at https://en.wikipedia.org/wiki/Base_rate_fallacy#False_positive_paradox
Most evidence is statistical, it increases the probability of guilt or innocence until the likelihood of guilt becomes overwhelming. If you have something that is perceived as overwhelmingly accurate then it becomes very difficult to challenge when it does make a mistake. This is an immense problem with DNA evidence these days -- most people have been schooled to believe its unerringly correct, and since 'most people' make up juries its an almost impossible task to overcome this evidence regardless of the facts, probabilities or what have you, if the police lab says there's a match you're done.
So, facial recognition introduces 'reasonable doubt' into its findings. I'm really happy with that because it means that I won't be arrested, tried and convicted for a crime that I have no knowledge of just because an algorithm said so.
So, facial recognition introduces 'reasonable doubt' into its findings.
Wrong. Reasonable doubt is inferred by juries. It's not implied by anything in the evidence. It's an attribute of interpretation, not of data.
You can claim that facial recognition should lead juries to infer a degree of reasonable doubt, but there's no guarantee that any given juror will see it that way.
I'm really happy with that because it means that I won't be arrested, tried and convicted for a crime that I have no knowledge of just because an algorithm said so.
I find your abundance of faith disturbing.
I feel for people of color... I do.
But, I’m more worried about going out in the sun...
On the other hand, this might be a business opportunity. Anyone know what Michael Jackson was using?
New product: Arrest-Off
Cream/procedure that is “guaranteed to reduce your probably of arrest by 90%”
Now, I can see the usefulness of facial recognition as *corroborating evidence*, to be combined with thorough police work and witness testimony. Relying on it as the *only* means of evidence is the problem. It should be there as a way of validating the rest of your evidence, just as DNA isn't a full proof in itself.
But just like many other technologies, a tool meant to assist your work and/or improve accuracy instead becomes an excuse to be lazy or sloppy in your own work.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
