Transcript of How AI facial recognition is being used to track down criminals | BBC News

Yes, it is time, a little later than scheduled for AI decoded. Facial recognition is hard for two reasons. Teaching a computer to process a human face is difficult enough. Matching that face to someone's identity in a database requires significant computing power, and billions of photographs tied to accurate data on which those computers can train. The technology has been around since the early 1970s in its most primitive form, but so unreliable was it that other biometrics, fingerprinting, retinal scanning, came to market quicker. About five years ago, though, a company called Clearview AI claimed to have made a breakthrough. Tonight, their CEO will join us live from Oakland, California. The advances in this technology are probably the biggest breakthrough in crime detection since we introduced DNA testing. This week, Police Scotland announced they will be rolling out facial recognition cameras across the country. Chief Constable Joe Farrell says it would be an abdication of her duties not to be using it. We were given a very good demonstration of it through the summer. Retrospective facial recognition tracked down many of those taking part in the riots, even those who were masked. But is it becoming too intrusive?

The software that identifies our faces is now being developed for authentication. We use it on our phones instead of a code. British Telecom are currently trialing that same technology to improve cybersecurity so that only authorized workers have access to critical systems and data. There is a lot to consider. And with me in the studio, as ever, the font of all AI knowledge, Priya Lekarne, CEO at Century Tech. She's come back from New York this morning, so bear with her.I'm going to put you on the spot, not the length.It might be a bit slow. Nonetheless, give us a quick explainer of how this technology works. Okay.

So, firstly, what we want to do is let's say we want to spot you in an image, and then, God forbid, we want to match that to a wanted list, Christian. So we've got an image of Christian here on the camera. In the background, you'll have him, let's say, in the studio, and there's lots of noise in that studio. There's tables, there's monitors. First, we want to do what we call classification. We want to detect the fact that we've got a face. We've got a face here. The way that we do that is we use deep learning models to be able to classify the image and find the face. The way that they traditionally do that is they would look at lots of images, they would tag lots of faces in those images. They'd say, These are faces, these are not faces. Then you would build up AI deep models, like deep training models, and you would train them to learn where the faces are, and they would look at those feature sets of a face. There are some models that use what we call unsupervised learning, but mostly supervised training models, where you've had that tagging and labeling.

The model then learns what the specific features are of the face, let's say the texture, the edges of what a face looks like. You've got a model that potentially can spot in that very big studio over there, all of the faces. We then take your face specifically, and we create If we can play a clip, actually, a boundary box, a bounding box around your face. I've got a clip to show our viewers. This is Friends. You can see this. It's very short then, but you saw those boxes around the face. You create these bounding boxes around the face. Then essentially, the model knows that it has a face, but we want to detect now whether that's your face and does that match something in a database. Let's put that up again. Can we do that again? Yeah. We normalize the image. Let's say you've got an image of a face. There's the box. Yeah, there's the box. Then you normalize the image. So you want to take away lots of variables that could essentially create inaccuracies. So maybe the lighting, you might want to turn all the images into grayscale. You might want to look at the alignment of the face.

And once you've done that, and this is where it's interesting, in those images, you saw there were lots of spots. The deep learning models will extract what it thinks are the key features of the face. Once it extracts those key features of Christian's face, and this is where it's really interesting, and I love this, is we turn it all into maths. We vectorise that data. What you end up with, Christian, as Christian Fraser right now, what you're used to seeing in an image, actually is a string of numbers. It's just a string of numbers. Every image of you that's taken on cameras out there that are modeling that face and using this AI, it will have a slightly different string of numbers because your position might be different, your expression might be different, but they'll be very similar because it's your face. It's a mathematical representation of your face. Then this is where it gets really interesting, is let's say you have a wanted list. You've got a database of faces that you're looking for. Those faces have been through a similar process where they have been encoded into numbers. Then it's a matching exercise.

Is there a similarity between the string of numbers that represent the face on the wanted list?

That's where the computing power comes in and where the chips come in because as the chips improve, those calculations are done that much quicker.

Do you remember, very quickly, because I know I really want to get Juan on because he is the expert in this area, but do you remember when we did an entire episode on semiconductors? We talked about how there are two big processes that happen. You've got all the We've got training data initially. You've got all the training data, for example, all the faces and the images. Then we have what we call doing inference and AI. That's when you run the model, you have the compute power. That's what you're talking about. You're then running the model against, essentially, that facial recognition to be able to find the face. Okay.

I mean, it's extremely impressive what Clearview do. Before we talk to Juan, let me show you the video, the promo video that Clearview puts out, and you'll understand what I mean.

In 2019, Homeland Security investigations were trying to identify an adult male who was in a child abuse video. The adult male was abusing a six-year-old girl and selling this abuse video on the dark web. The only clue was a photo of the adult male who was in the background of the abuse video for just a few frames. With no other clues, the investigator in the case turned to Clearview AI. Prior to searching on Clearview AI, we must provide a reason for the search. In this example, we will choose Felon a sex offense. Secondly, you upload a photo of the suspect from your desktop and press the Search button. As you can see, Actually, 25 results now match the uploaded photo, whereas in 2019, during Homeland Security's investigation, only one result came back from Clearview AI. This is the photo. As you can see, The suspect is in the background of the photo. Press the Locate button on the top left to zoom into the photo or use the Compare button to see them side by side. In this case, the investigator clicked the link to a public social social media post online, uncovering two key pieces of information.

The photo was tagged in Las Vegas and the name of the company the suspect appeared to work for. With those two clues, the investigators at Homeland Security traveled to Las Vegas, obtained the suspect's name from the employer, and with additional corroborating evidence, secured a search warrant for the suspect's computers. The search warrant revealed that the suspect had thousands of videos and photos of child abuse material on his computer. He pled guilty and is now doing 35 years in jail, and the six-year-old girl was rescued.

Impressive. Let's speak to the CEO of Clearview, Juan Tontat. Juan, thank you very much for being with us. How many cases do you think your technology has solved? And what was the big leap forward for you?

Christian and Priya, thanks so much for having me on. It's great to be here. And And I appreciate your interest in Clearview AI. We've done now over 2 million searches on behalf of law enforcement. That's how many searches they've used on our platform. We don't know how many crimes they've actually solved, but if you take even a conservative estimate, you would be in the hundreds of thousands, at least. Sometimes cases, you have to search multiple images and so on. But anecdotally as well, most recently, we worked with the International Center of missing and Exploited Children, and we were in Ecuador and Latin America. In three days, these law enforcement agencies, about eight of them, went through a list of the hardest cold cases they haven't solved. These are missing kids. Kids have been abused, and they find them on these Internet forums as victims. And in those three days, they made 110 identifications of missing and exploited children and over 50 of them. The impact is incredible. On the flip side as well, we know it's a very powerful technology, so we've limited the usage of our application to law enforcement and governments.

But what is it? I said that there's been a lag with this technology, whereas retinal scans and fingerprinting and jump forward. What has been the real breakthrough for facial recognition?

Yeah, I think it's neural networks, which which are part of artificial intelligence. Previous algorithms for facial recognitions would try and look at the distance between the eyes or the eyes and the eyebrows or the nose and the eyes and things like that. But that doesn't work very well if you have an image from a different angle, say a security camera. With neural networks, you're able to train, as Priya said, it's called supervised learning on a lot of different examples of photos to improve accuracy. The The way we trained that algorithm was to get a lot of publicly available images. Say you have 100 photos of George Clooney, you have 100 photos of Brad Pitt. The algorithm will learn that the black and white photo of Brad Pitt with the sunglasses on is the same one of him from 20 years ago with different hair and so on. The algorithm learns what stays the same in a face. The more data you have, the more accurate it gets. There's been some great researchers out there, thanks to machine learning and all this data that the research community has done a really good job in improving it.

We built upon a lot of those innovations. What we were able to do was bring a lot of data to train our algorithm. And so now when you look at all the top facial recognition algorithms, not just Clearview, there are others as well. There's a National Institute of Standards and Technology in the US that ranks hundreds these algorithms. We can pick a photo out of a lineup of 12 million images at a 99.85% accuracy rate, and that's across all demographics. So the technology has now become much more accurate than the human eye, and the education, it really is artificial intelligence and the amount of data that's out there that you can use to train these algorithms.

Juan, what I'm really interested in is something that you just said, because I thought that prior tests showed that there was a significant statistical difference in the performance of the model when it came to certain demographics?

Yes. Nist ranks demographics as well. But if you look at the top performing algorithms, the differential is a very, very small, and you're looking at over 99% accuracy across all the demographics you do test. But if you looked four or five years ago, that's when a lot of these algorithms did have issues with accuracy, especially on in certain demographics. But today, a lot of the top algorithms are very accurate, regardless of demographics. But there's also things like angles.So.

You could get partial faces. We talked about people with masks who were identified through the summer here during the riots. But you could have, I don't know, a third of your face, someone in a Balaclava with eyebrows. You might still pick up an image.

Yes, we were surprised, too. We build a software. I'm a software engineer by background. When COVID When COVID-19 happened, we had a lot of issues identifying people with masks on. What we did is we added photos with masks to our training data. About 3% of the photos, we photoshopped masks onto them. It was incredible to see that now, almost all the time, even with the mask on, our algorithm works very well in searching out billions of images. The technology even surprises people who are making it.

Juan, am I right? When you talked originally about having a lot of data to be able to add to this, and it's the data and then the labeling and the supervised learning that then allows you to have these really performant models. That data was taken from the internet. I think in terms of some of the concerns, what is Clearview doing to ensure that this technology is not being misused? Have you got safeguards in place yourselves as a business? Have you got safeguards in place with the law enforcement authorities that you're dealing with across the world? We want to hear a little bit more about that, where this data has come from, where the inputs have come from, and then how those potential outputs may be used?

Yeah, so we have a wide variety of ways to make sure that this technology is used for the best and highest purpose, which is solving crime, but also not misused by law enforcement and other users. First of all, as we do restrict this to government and law enforcement agencies, and we have a vetting process ourselves before we onboard any customer. We look at their human rights background and all those things.

Yeah, I was going to ask you about that because I was thinking of the Skripals who were poisoned in Salisbury, and of course, the Russian government would be looking for them, and I was wondering whether they would have access. But you have a vetting process for that.

Yes. So we, for example, won't sell to Russia, China, Iran, any one that's adverse to the US and US allies. And in fact, we do sell to the Ukrainians. They use it very effectively since the beginning of the war. There's been over 2,000 war crimes now where they have been solved or the suspects have been identified because of clear view. That wouldn't have been identified otherwise. It's also been...

Sorry to interrupt. How safe is this database, though? Because through our series, one of the things we've been talking about is that cat and mouse of AI, AI accessing I see AI. Are you concerned at all, given this immense amount of data that you've gathered, that it is safe from the bad actors who would want access to it?

Yeah, it's a great question. We vet every customer. We talk to them, we find out and try and verify who they are, of course, before we onboard them. Secondly, we make sure that there's an administrator in charge of the facial recognition program in any particular agency. So any law enforcement officer using it before they do a search, as you saw on the demo, they have to put in a case number and a crime type, and that allows the administrators to audit on a regular basis which offices are using the technology and what reasons. So that is another safeguard that really helps these agencies make sure it's used for the right purpose, and they can take action as appropriate if there's any misuse of the technology. I think that's another control we think is an innovation we've had that sets us apart. And finally, we give training to all the people who use Clearview to make sure they know how to not just take the search results and go with it, but verify the information that comes back from Clearview. We don't allow the search results to be used as the only source of evidence. These are leads, and these investigators do follow up research to verify the identity.

So with those things, that's how we've been able to really get the best out of the technology and minimize a lot of the downsides.

But just very quickly, humans do have to review the results of that.

Absolutely. So one thing we've done in our software is we don't actually show the person using it. If we think it's 98 or 99%, that's actually not shown in our software for that reason. That way, they have to be high.

It's amazing to talk to you. Thank you very much for coming on the program.

Thanks so Thank you very much for having me.

Quentin Tapp there from Clearbeef. That's making you feel a little Orwellian. Coming up after the break, we'll get the other side of the debate. How do we protect our freedoms? We'll speak to Big Brother Watch, the British Campaign Group, who's pushing back against the advances in state surveillance. Welcome back. Now we know how impressive this AI technology can be. It's power stretches far beyond the average search engine. This is a radical reimagining of the public space. If, as looks likely, it is widely adopted by our police forces, then increasingly the freedom to wander about without being watched will disappear. Facial recognition will bind us to our digital history in ways we've not yet imagined. It will be the end of what was previously taken for granted, the right to be publicly anonymous. With us tonight is Silky Carla. She is the Director of Big Brother Watch. Before we speak to her, let's watch her campaign video.

Out of van, there are cameras on top of the van that scan everyone's faces as they walk past.

I came to London Bridge, I was pulled up at London Bridge in regards to his facial recognition.

They were trying to threaten me in regards to an arrest. He says to me, your band.

I did cry the entire way, huh? I felt so helpless. I just really wanted to find any way to prove that I was a thief. So typically on a busy day in a London area like this, they can be scanning thousands of people's faces. It's thousands of people that are effectively walking through a digital police lineup. As you walk past, it's a passport-style check that sees if you are known to the police, if you are on a database. There are lots of different reasons that people can be put onto watch lists. You can be put on a watch list to protect you from harm, whatever that means. Typically, what we see is an alert comes up, bam, the person is stopped. If they don't stop voluntarily, then they can be physically apprehended.

Silky Carl, welcome to the program. Thanks for having me. It's difficult for someone in my position who's on television every night to argue for more privacy. But I think if I were not in this job, I think if I could choose to opt out of Juan's database, I might want to do that, but it seems very tricky now.

Yeah, the extraordinary thing is that in British law and in European law, you have the right to protect your own data, including your photographs. And of course, what he's doing with Clearview AI is actually it's not just stealing billions of photographs from the internet that people haven't consented to. There's about 30 billion in his database alone. But it's also extracting biometric data from them. That's information as sensitive as what's on your passport. Then these companies are making that very sensitive data available to the the highest bidder. It's really just a question of what does the buyer want to do with it?

You say it's stealing. Is it stealing or scraping? Is it the difference?

Well, there are two words for the same thing because you have rights over your data. We fought for a long time to have... The right to privacy is a protected right. It's a fundamental human right. The thing about facial recognition, especially when you've got companies either taking it from people on the street as they walk around through CCTV or scraping it from the internet, is that it actually reverses the presumption.

But in reality, the consent happens in stages. We give the images to Instagram, to TikTok. We assume... I mean, I don't know. Do we assume they would be used for other purposes? I don't know. But at the same time, we enjoy using Google photos, which then puts all our photos into certain order and identifies people for us. We like all that. Are you saying that we have to give up all that if we We want to keep our freedoms?

No, not at all. But there's a different level of protection for biometric data because it's so sensitive. It's like DNA. In the same way that, and I do think with facial recognition, there are ways that highly regulated You can use it for the public benefit. But the problem is we don't allow companies to go around scraping busses and playgrounds and high streets for people's DNA and making mass databases. Because it's so unregulated at the moment. That's what's happening with facial recognition.

Traditionally, when we're looking at the actual process, we're relying on people recognizing people or people remembering people in human eyewitness. We're relying on this human facial recognition rather than a machine. What I'm really interested in is the public interest argument. We've got these security forces. We've got the Met police using this technology now in the UK, the Scottish police, the Welsh police. I think it was Big Brother Watch, a quote from you where you said police are failing to turn up to even 40% of violent shoplifting incidences. The Met police started the year 1,000 officers short. They're going to probably end the year 1,400 officers short. I don't want those officers trawling through faces, trying to make those matches. You want the machines to do it for them. It's a tool, it's a technology. I can understand the point about the consent of the data and the input of the data, but how do we achieve that balance where there is still the job to be done here? I know crime is reduced by 8%, but sexual assault has actually rocketed upward. How do we use the technology to benefit us while being able to mitigate against potential risks and harms that are caused by the breach of our...

Potential breach of privacy, I should say.

We have to have regulation. We have to have laws around this. It's Completely unlike, we've got laws on fingerprints, we've got laws on DNA, we've got laws on CCTV. Facial recognition is just a vacuum at the moment. It's a Wild West for companies to go in and build these databases of billions of photos. Also, I have to say for the police. Let me give an example. I've been watching the police use this for seven years now. I'll give an example of what it actually looks like on the street. I go to a high crime area like Croydon, where police are using live facial recognition. Masses of resource, officers standing around a van looking at iPads, I walked past a robbery on my way to watch the police using facial recognition. Of course, then they are getting it wrong as well. That's where we step in as Big Brother Watch because there are injustices. People wrongly stopped, questioned, harassed by police because they've been misidentified.

There's one other issue which we've not talked about, and this is that third story I had in the introduction where British Telecom is saying we can use this for our own security. Is there a danger that actually we're being snooped on by our employer?

Absolutely. We've just released a report on this actually called Bossware. And increasingly, whether it's on construction sites, the gig economy, people are being basically told that they have give over DNA fingerprint-style data. Sometimes it is literally fingerprints and increasingly facial recognition just to get their pay packet at the end of the month. And we really have to be careful about what that data is used for. Do they have controls over it? Do they It's not a choice? Because it even happens in schools now that kids are giving facial recognition to get their school lunch. Yeah.

You know what? My boss is watching me. We're out of time.

I need a grumpy face recognition. Say that when you're with your spouse or your partner, it goes, Ping, smile. We don't want bosses looking at-We all need one of those. We need one of those, don't we?

Silky, thank you for coming in. Really interesting. Get in touch if you've got thoughts on what we've discussed. That's it for this week. As I like to remind you each week, though, if you enjoyed tonight's show, you can watch it all back on the back catalog on our YouTube channel, AI decoded. Do get in touch about that. If you've got thoughts on a program, we'd like to hear that, too. We'll do it again same time next week.